diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp index 1ae077130..faed7085e 100644 --- a/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp +++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp @@ -5462,14 +5462,20 @@ Ref OpDispatchBuilder::PCMPXSTRXSaturateExplicitLength(OpSize ElementSize, OpSiz return _Select(OpSize::i32Bit, LengthSize, CondClass::ULT, AbsLength, Constant(NumElements), AbsLength, Constant(NumElements)); } +// Equal Any is a "is character in set" operation. +// The set of characters is passed in Src1, and we then check if every +// character in Src2 is in that set. Ref OpDispatchBuilder::PCMPXSTRXEqualAny(OpSize ElementSize, Ref Src1, Ref Src2, Ref Src1ValidElements, Ref Src2ValidElements) { const uint32_t NumElements = IR::NumElements(OpSize::i128Bit, ElementSize); - // Replace the invalid Src1 elements with Src1[0] so they can't add matches of their own. + // First we sanitize Src1, and zero out elements after the end of the string, + // removing them from the matching set. Ref Src1Element0 = _VDupElement(OpSize::i128Bit, ElementSize, Src1, 0); Ref SanitizedSrc1 = _VBSL(OpSize::i128Bit, Src1ValidElements, Src1, Src1Element0); - // A Src2 element matches if it equals any Src1 element. + // Now walk through every element in Src1, broadcast it into a temporary vector, + // and compare it against every element in Src2, then OR the results + // together to get the final match vector. Ref Matches = _VCMPEQ(OpSize::i128Bit, ElementSize, Src2, Src1Element0); for (uint32_t i = 1; i < NumElements; i++) { Ref Src1Element = _VDupElement(OpSize::i128Bit, ElementSize, SanitizedSrc1, i); @@ -5477,39 +5483,52 @@ Ref OpDispatchBuilder::PCMPXSTRXEqualAny(OpSize ElementSize, Ref Src1, Ref Src2, Matches = _VOr(OpSize::i128Bit, Matches, ElementMatches); } - // An empty Src1 matches nothing, and invalid Src2 elements never match. + // Handle the case where Src1 was entirely empty, and also sanitize the results + // for any NULLs in Src2, since those don't count towards matching. Ref Src1NotEmpty = _VDupElement(OpSize::i128Bit, ElementSize, Src1ValidElements, 0); Matches = _VAnd(OpSize::i128Bit, Matches, Src1NotEmpty); return _VAnd(OpSize::i128Bit, Matches, Src2ValidElements); } + +// The ranges aggregation is a weird one. It checks if every character in Src2 is +// within a character range (ie. a-z or A-Z) similar to regex. +// Ranges are passed in Src1 as pairs of characters in adjacent lanes. Ref OpDispatchBuilder::PCMPXSTRXRanges(OpSize ElementSize, Ref Src1, Ref Src2, Ref Src1ValidElements, Ref Src2ValidElements, bool IsSigned) { const uint32_t NumElements = IR::NumElements(OpSize::i128Bit, ElementSize); - // Signed or unsigned greater than, depending on the source data format. + // Signed or unsigned greater than comparison, based on the Imm8 value const auto GreaterThan = [&](Ref Lhs, Ref Rhs) { return IsSigned ? _VCMPGT(OpSize::i128Bit, ElementSize, Lhs, Rhs) : _VUCMPGT(OpSize::i128Bit, ElementSize, Lhs, Rhs); }; - // A Src2 element matches if it is within any valid [Src1[2k], Src1[2k+1]] range. + // First we walk through the ranges from Src1, and construct + // temporary vectors to represent the lower and upper bounds + // of the comparison. Then check if both comparisons are true, + // zero out invalid lanes, and OR the result into the final result vector. Ref Result {}; for (uint32_t i = 0; i < NumElements; i += 2) { Ref LowerBound = _VDupElement(OpSize::i128Bit, ElementSize, Src1, i); Ref UpperBound = _VDupElement(OpSize::i128Bit, ElementSize, Src1, i + 1); Ref BelowLower = GreaterThan(LowerBound, Src2); Ref AboveUpper = GreaterThan(Src2, UpperBound); - Ref OutsideRange = _VOr(OpSize::i128Bit, BelowLower, AboveUpper); // A range is only valid if its upper bound is a valid Src1 element. Ref RangeValid = _VDupElement(OpSize::i128Bit, ElementSize, Src1ValidElements, i + 1); - Ref InRange = _VAndn(OpSize::i128Bit, RangeValid, OutsideRange); + + // RangeValid & ~BelowLower & ~AboveUpper + Ref InRange = _VAndn(OpSize::i128Bit, RangeValid, BelowLower); + InRange = _VAndn(OpSize::i128Bit, InRange, AboveUpper); Result = Result ? _VOr(OpSize::i128Bit, Result, InRange) : InRange; } - // Invalid Src2 elements never match. + // Invalid (ie. NULL) Src2 characters don't count towards matching. return _VAnd(OpSize::i128Bit, Result, Src2ValidElements); } +// This aggregation is the simplest, and is the most similar +// to strcmp(). It Just compares lanewise which characters are +// equal in each string. Ref OpDispatchBuilder::PCMPXSTRXEqualEach(OpSize ElementSize, Ref Src1, Ref Src2, Ref Src1ValidElements, Ref Src2ValidElements) { // Elements match when both are valid and equal, or when both are invalid. Ref Equal = _VCMPEQ(OpSize::i128Bit, ElementSize, Src1, Src2); @@ -5519,33 +5538,43 @@ Ref OpDispatchBuilder::PCMPXSTRXEqualEach(OpSize ElementSize, Ref Src1, Ref Src2 return _VOrn(OpSize::i128Bit, ValidMatches, EitherValid); } +// EqualOrdered implements a strstr()-like semantic finding all instances +// of the string Src1 in Src2. Ref OpDispatchBuilder::PCMPXSTRXEqualOrdered(OpSize ElementSize, Ref Src1, Ref Src2, Ref Src1Length, Ref Src2Length, Ref Src1ValidElements, Ref Indices) { const uint32_t NumElements = IR::NumElements(OpSize::i128Bit, ElementSize); - // Row i: does Src1[i] match at each start position? Rows past the needle length are forced to - // match, so ANDing all rows together leaves the positions where the whole needle matches. + // Broadcast needle[i] into every lane of a temp vector, then compare it to + // the haystack vector. On succesive iterations, we shift the haystack over + // by one element, and compare it against the next needle element. + // AND together the results of all comparisons, and what is left should + // be a vector where the only 'true' elements are lanes where the full + // needle was found in the haystack, or those positions after the end of the string. Ref Result {}; for (uint32_t i = 0; i < NumElements; i++) { Ref Needle = _VDupElement(OpSize::i128Bit, ElementSize, Src1, i); Ref Haystack = i == 0 ? Src2 : _VExtr(OpSize::i128Bit, ElementSize, Needle, Src2, i); - Ref Row = _VCMPEQ(OpSize::i128Bit, ElementSize, Haystack, Needle); + Ref ElementsEqual = _VCMPEQ(OpSize::i128Bit, ElementSize, Haystack, Needle); - // Row | ~RowValid: the row itself for valid needle elements, all ones otherwise. - Ref RowValid = _VDupElement(OpSize::i128Bit, ElementSize, Src1ValidElements, i); - Row = _VOrn(OpSize::i128Bit, Row, RowValid); + // TODO: I think a more optimal version of this is possible, perhaps using + // MATCH from SVE2? + Ref ElementsValid = _VDupElement(OpSize::i128Bit, ElementSize, Src1ValidElements, i); + Ref ElementsEqualValid = _VOrn(OpSize::i128Bit, ElementsEqual, ElementsValid); - Result = Result ? _VAnd(OpSize::i128Bit, Result, Row) : Row; + // Ternary to handle the first row case + Result = Result ? _VAnd(OpSize::i128Bit, Result, ElementsEqualValid) : ElementsEqualValid; } - // Clear positions that run into invalid Src2 elements, unless Src2 is fully valid or Src1 is empty. + // Clear positions in the result after the end of the string. Ref FirstClearedPosition = Add(OpSize::i32Bit, _Sub(OpSize::i32Bit, Src2Length, Src1Length), 1); FirstClearedPosition = _Select(OpSize::i32Bit, OpSize::i32Bit, CondClass::ULT, Src2Length, Constant(NumElements), FirstClearedPosition, Constant(NumElements)); FirstClearedPosition = _Select(OpSize::i32Bit, OpSize::i32Bit, CondClass::EQ, Src1Length, Constant(0), Constant(NumElements), FirstClearedPosition); + Ref FirstClearedPositionVector = _VDupFromGPR(OpSize::i128Bit, ElementSize, FirstClearedPosition); - return _VAnd(OpSize::i128Bit, Result, _VCMPGT(OpSize::i128Bit, ElementSize, FirstClearedPositionVector, Indices)); + Ref KeptPositions = _VCMPGT(OpSize::i128Bit, ElementSize, FirstClearedPositionVector, Indices); + return _VAnd(OpSize::i128Bit, Result, KeptPositions); } void OpDispatchBuilder::PCMPXSTRXOpImpl(OpcodeArgs, bool IsExplicit, bool IsMask, bool IsAVX) { diff --git a/unittests/InstructionCountCI/SSE42_Strings.json b/unittests/InstructionCountCI/SSE42_Strings.json index a715aa82b..d5483a1c1 100644 --- a/unittests/InstructionCountCI/SSE42_Strings.json +++ b/unittests/InstructionCountCI/SSE42_Strings.json @@ -404,64 +404,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -512,32 +512,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -589,64 +589,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -697,32 +697,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -1800,64 +1800,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -1909,32 +1909,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -1987,64 +1987,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -2096,32 +2096,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -3204,64 +3204,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -3312,32 +3312,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -3389,64 +3389,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -3497,32 +3497,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -4600,64 +4600,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -4709,32 +4709,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -4787,64 +4787,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -4896,32 +4896,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -5994,64 +5994,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -6098,32 +6098,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -6170,64 +6170,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -6274,32 +6274,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -7318,64 +7318,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -7423,32 +7423,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -7496,64 +7496,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -7601,32 +7601,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -8650,64 +8650,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -8754,32 +8754,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -8826,64 +8826,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -8930,32 +8930,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -9974,64 +9974,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -10079,32 +10079,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -10152,64 +10152,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -10257,32 +10257,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -11322,64 +11322,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -11428,32 +11428,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -11502,64 +11502,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -11608,32 +11608,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -12678,64 +12678,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -12785,32 +12785,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -12860,64 +12860,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -12967,32 +12967,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -14042,64 +14042,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -14148,32 +14148,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -14222,64 +14222,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -14328,32 +14328,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -15398,64 +15398,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -15505,32 +15505,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -15580,64 +15580,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -15687,32 +15687,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -16790,64 +16790,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -16901,32 +16901,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -16980,64 +16980,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -17091,32 +17091,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -18226,64 +18226,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -18338,32 +18338,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -18418,64 +18418,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -18530,32 +18530,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "mvn v3.16b, v3.16b", @@ -19670,64 +19670,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -19781,32 +19781,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -19860,64 +19860,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.16b, #0x10", @@ -19971,32 +19971,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "movi v4.8h, #0x8", @@ -21106,64 +21106,64 @@ "dup v6.16b, v16.b[1]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -21218,32 +21218,32 @@ "dup v6.8h, v16.h[1]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -21298,64 +21298,64 @@ "dup v6.16b, v16.b[1]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[2]", "dup v7.16b, v16.b[3]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[4]", "dup v7.16b, v16.b[5]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[6]", "dup v7.16b, v16.b[7]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[7]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[7]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[8]", "dup v7.16b, v16.b[9]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[9]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[9]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[10]", "dup v7.16b, v16.b[11]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[11]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[11]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[12]", "dup v7.16b, v16.b[13]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[13]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[13]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.16b, v16.b[14]", "dup v7.16b, v16.b[15]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -21410,32 +21410,32 @@ "dup v6.8h, v16.h[1]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[1]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[1]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[2]", "dup v7.8h, v16.h[3]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[3]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[3]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[4]", "dup v7.8h, v16.h[5]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[5]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[5]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "orr v5.16b, v5.16b, v6.16b", "dup v6.8h, v16.h[6]", "dup v7.8h, v16.h[7]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v6.16b", + "bic v3.16b, v3.16b, v7.16b", "orr v3.16b, v5.16b, v3.16b", "and v3.16b, v3.16b, v4.16b", "eor v3.16b, v3.16b, v4.16b", @@ -22547,64 +22547,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -22655,32 +22655,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -22732,64 +22732,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -22840,32 +22840,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -23947,64 +23947,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -24056,32 +24056,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -24134,64 +24134,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -24243,32 +24243,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v2.16b, v2.16b", @@ -25355,64 +25355,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -25463,32 +25463,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -25540,64 +25540,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -25648,32 +25648,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "ldr q3, [x28, #3440]", @@ -26755,64 +26755,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -26864,32 +26864,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -26942,64 +26942,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -27051,32 +27051,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v2.16b, v3.16b, v2.16b", @@ -28153,64 +28153,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -28257,32 +28257,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -28329,64 +28329,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -28433,32 +28433,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -29481,64 +29481,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -29586,32 +29586,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -29659,64 +29659,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -29764,32 +29764,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v2.16b, v3.16b, v2.16b", "mvn v16.16b, v2.16b", @@ -30817,64 +30817,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -30921,32 +30921,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -30993,64 +30993,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv b2, v16.16b", @@ -31097,32 +31097,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v16.16b, v3.16b, v2.16b", "umaxv h2, v16.8h", @@ -32145,64 +32145,64 @@ "dup v5.16b, v16.b[1]", "cmhi v4.16b, v4.16b, v17.16b", "cmhi v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmhi v5.16b, v5.16b, v17.16b", "cmhi v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -32250,32 +32250,32 @@ "dup v5.8h, v16.h[1]", "cmhi v4.8h, v4.8h, v17.8h", "cmhi v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmhi v5.8h, v5.8h, v17.8h", "cmhi v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -32323,64 +32323,64 @@ "dup v5.16b, v16.b[1]", "cmgt v4.16b, v4.16b, v17.16b", "cmgt v5.16b, v17.16b, v5.16b", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.16b, v3.b[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.16b, v3.b[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[2]", "dup v6.16b, v16.b[3]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[4]", "dup v6.16b, v16.b[5]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[6]", "dup v6.16b, v16.b[7]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[7]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[7]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[8]", "dup v6.16b, v16.b[9]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[9]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[9]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[10]", "dup v6.16b, v16.b[11]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[11]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[11]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[12]", "dup v6.16b, v16.b[13]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.16b, v3.b[13]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.16b, v3.b[13]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.16b, v16.b[14]", "dup v6.16b, v16.b[15]", "cmgt v5.16b, v5.16b, v17.16b", "cmgt v6.16b, v17.16b, v6.16b", - "orr v5.16b, v5.16b, v6.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -32428,32 +32428,32 @@ "dup v5.8h, v16.h[1]", "cmgt v4.8h, v4.8h, v17.8h", "cmgt v5.8h, v17.8h, v5.8h", - "orr v4.16b, v4.16b, v5.16b", - "dup v5.8h, v3.h[1]", - "bic v4.16b, v5.16b, v4.16b", + "dup v6.8h, v3.h[1]", + "bic v4.16b, v6.16b, v4.16b", + "bic v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[2]", "dup v6.8h, v16.h[3]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[3]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[3]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[4]", "dup v6.8h, v16.h[5]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", - "dup v6.8h, v3.h[5]", - "bic v5.16b, v6.16b, v5.16b", + "dup v7.8h, v3.h[5]", + "bic v5.16b, v7.16b, v5.16b", + "bic v5.16b, v5.16b, v6.16b", "orr v4.16b, v4.16b, v5.16b", "dup v5.8h, v16.h[6]", "dup v6.8h, v16.h[7]", "cmgt v5.8h, v5.8h, v17.8h", "cmgt v6.8h, v17.8h, v6.8h", - "orr v5.16b, v5.16b, v6.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v5.16b", + "bic v3.16b, v3.16b, v6.16b", "orr v3.16b, v4.16b, v3.16b", "and v3.16b, v3.16b, v2.16b", "eor v16.16b, v3.16b, v2.16b", @@ -33493,64 +33493,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -33598,32 +33598,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -33671,64 +33671,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -33776,32 +33776,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -34837,64 +34837,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -34943,32 +34943,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -35017,64 +35017,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -35123,32 +35123,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -36189,64 +36189,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -36294,32 +36294,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -36367,64 +36367,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -36472,32 +36472,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bsl v3.16b, v2.16b, v4.16b", @@ -37533,64 +37533,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -37639,32 +37639,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -37713,64 +37713,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -37819,32 +37819,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -38918,64 +38918,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -39029,32 +39029,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -39108,64 +39108,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -39219,32 +39219,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -40354,64 +40354,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -40466,32 +40466,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -40546,64 +40546,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -40658,32 +40658,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "mvn v3.16b, v3.16b", @@ -41798,64 +41798,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -41909,32 +41909,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -41988,64 +41988,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -42099,32 +42099,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "bit v4.16b, v2.16b, v3.16b", @@ -43234,64 +43234,64 @@ "dup v7.16b, v16.b[1]", "cmhi v6.16b, v6.16b, v17.16b", "cmhi v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmhi v7.16b, v7.16b, v17.16b", "cmhi v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -43346,32 +43346,32 @@ "dup v7.8h, v16.h[1]", "cmhi v6.8h, v6.8h, v17.8h", "cmhi v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmhi v7.8h, v7.8h, v17.8h", "cmhi v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -43426,64 +43426,64 @@ "dup v7.16b, v16.b[1]", "cmgt v6.16b, v6.16b, v17.16b", "cmgt v7.16b, v17.16b, v7.16b", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.16b, v3.b[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.16b, v3.b[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[2]", "dup v8.16b, v16.b[3]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[4]", "dup v8.16b, v16.b[5]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[6]", "dup v8.16b, v16.b[7]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[7]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[7]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[8]", "dup v8.16b, v16.b[9]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[9]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[9]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[10]", "dup v8.16b, v16.b[11]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[11]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[11]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[12]", "dup v8.16b, v16.b[13]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.16b, v3.b[13]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.16b, v3.b[13]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.16b, v16.b[14]", "dup v8.16b, v16.b[15]", "cmgt v7.16b, v7.16b, v17.16b", "cmgt v8.16b, v17.16b, v8.16b", - "orr v7.16b, v7.16b, v8.16b", "dup v3.16b, v3.b[15]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b", @@ -43538,32 +43538,32 @@ "dup v7.8h, v16.h[1]", "cmgt v6.8h, v6.8h, v17.8h", "cmgt v7.8h, v17.8h, v7.8h", - "orr v6.16b, v6.16b, v7.16b", - "dup v7.8h, v3.h[1]", - "bic v6.16b, v7.16b, v6.16b", + "dup v8.8h, v3.h[1]", + "bic v6.16b, v8.16b, v6.16b", + "bic v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[2]", "dup v8.8h, v16.h[3]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[3]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[3]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[4]", "dup v8.8h, v16.h[5]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", - "dup v8.8h, v3.h[5]", - "bic v7.16b, v8.16b, v7.16b", + "dup v9.8h, v3.h[5]", + "bic v7.16b, v9.16b, v7.16b", + "bic v7.16b, v7.16b, v8.16b", "orr v6.16b, v6.16b, v7.16b", "dup v7.8h, v16.h[6]", "dup v8.8h, v16.h[7]", "cmgt v7.8h, v7.8h, v17.8h", "cmgt v8.8h, v17.8h, v8.8h", - "orr v7.16b, v7.16b, v8.16b", "dup v3.8h, v3.h[7]", "bic v3.16b, v3.16b, v7.16b", + "bic v3.16b, v3.16b, v8.16b", "orr v3.16b, v6.16b, v3.16b", "and v3.16b, v3.16b, v5.16b", "eor v3.16b, v3.16b, v5.16b",