Merge pull request #5326 from OFFTKP/vcmpss_full

Add vcmp**_full tests

FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp

@@ -620,7 +620,7 @@ Ref OpDispatchBuilder::InsertScalarFCMPOpImpl(OpSize Size, IR::OpSize OpDstSize,
     // If either of the sources are unordered, then returns true.
     Ref Src1_U = _VFCMPEQ(Size, ElementSize, Src1, Src1);
     Ref Src2_U = _VFCMPEQ(Size, ElementSize, Src2, Src2);
-    auto Ordered = _VAnd(ElementSize, ElementSize, Src1_U, Src2_U);
+    auto Ordered = _VAnd(Size, ElementSize, Src1_U, Src2_U);
 
     Ref Compare_Ordered = _VFCMPEQ(Size, ElementSize, Src1, Src2);
     Ref Result = _VOrn(Size, ElementSize, Compare_Ordered, Ordered);
@@ -2508,8 +2508,8 @@ Ref OpDispatchBuilder::VFCMPOpImpl(OpSize Size, IR::OpSize ElementSize, Ref Src1
   case VectorCompareType::ORD_S: return _VFCMPORD(Size, ElementSize, Src1, Src2);
   case VectorCompareType::NGT_UQ:
   case VectorCompareType::NGT_US: {
-    Ref Result = _VFCMPLT(ElementSize, ElementSize, Src2, Src1);
-    return _VNot(ElementSize, ElementSize, Result);
+    Ref Result = _VFCMPLT(Size, ElementSize, Src2, Src1);
+    return _VNot(Size, ElementSize, Result);
   }
   case VectorCompareType::NGE_UQ:
   case VectorCompareType::NGE_US: {
@@ -2525,7 +2525,7 @@ Ref OpDispatchBuilder::VFCMPOpImpl(OpSize Size, IR::OpSize ElementSize, Ref Src1
     // If either of the sources are unordered, then returns true.
     Ref Src1_U = _VFCMPEQ(Size, ElementSize, Src1, Src1);
     Ref Src2_U = _VFCMPEQ(Size, ElementSize, Src2, Src2);
-    auto Ordered = _VAnd(ElementSize, ElementSize, Src1_U, Src2_U);
+    auto Ordered = _VAnd(Size, ElementSize, Src1_U, Src2_U);
 
     Ref Compare_Ordered = _VFCMPEQ(Size, ElementSize, Src1, Src2);
     return _VOrn(Size, ElementSize, Compare_Ordered, Ordered);
@@ -2541,9 +2541,9 @@ Ref OpDispatchBuilder::VFCMPOpImpl(OpSize Size, IR::OpSize ElementSize, Ref Src1
     return _VAnd(Size, ElementSize, Result, Src2_U);
   }
   case VectorCompareType::FALSE_OQ:
-  case VectorCompareType::FALSE_OS: return LoadZeroVector(OpSize::i128Bit);
+  case VectorCompareType::FALSE_OS: return LoadZeroVector(Size);
   case VectorCompareType::TRUE_UQ:
-  case VectorCompareType::TRUE_US: _VectorImm(OpSize::i128Bit, OpSize::i8Bit, -1, 0);
+  case VectorCompareType::TRUE_US: return _VectorImm(Size, OpSize::i8Bit, -1, 0);
   }
   FEX_UNREACHABLE;
 }

unittests/ASM/VEX/vcmppd_full.asm

@@ -0,0 +1,369 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+    "RCX": "448"
+  }
+}
+%endif
+
+%define true 1
+%define false 0
+
+%define EQ_OQ 0
+%define LT_OS 1
+%define LE_OS 2
+%define UNORD_Q 3
+%define NEQ_UQ 4
+%define NLT_US 5
+%define NLE_US 6
+%define ORD_Q 7
+%define EQ_UQ 8
+%define NGE_US 9
+%define NGT_US 10
+%define FALSE_OQ 11
+%define NEQ_OQ 12
+%define GE_OS 13
+%define GT_OS 14
+%define TRUE_UQ 15
+%define EQ_OS 16
+%define LT_OQ 17
+%define LE_OQ 18
+%define UNORD_S 19
+%define NEQ_US 20
+%define NLT_UQ 21
+%define NLE_UQ 22
+%define ORD_S 23
+%define EQ_US 24
+%define NGE_UQ 25
+%define NGT_UQ 26
+%define FALSE_OS 27
+%define NEQ_OS 28
+%define GE_OQ 29
+%define GT_OQ 30
+%define TRUE_US 31
+
+; Arguments: src1, src2, predicate, true/false
+%macro Compare 4
+    vcmppd result, %1, %2, %3
+    ; Construct expected result
+    vmovdqa temp, [rel .false + 32 * %4]
+    ; Compare if result and expected are equal
+    vxorpd ymm6, ymm7, ymm8
+    vptest ymm6, ymm6
+    jnz .fail
+    ; Increment counter of tests passed
+    add ecx, 1
+%endmacro
+
+lea rax, [rel .data]
+xor ecx, ecx
+
+vmovdqa ymm0, [rax + 32 * 0] ; -4.0
+vmovdqa ymm1, [rax + 32 * 1] ; -0.0
+vmovdqa ymm2, [rax + 32 * 2] ; 0.0
+vmovdqa ymm3, [rax + 32 * 3] ; 4.0
+vmovdqa ymm4, [rel .nan]     ; NaN
+
+%macro CompareAll 5
+    %define xmm_neg_four %1
+    %define xmm_neg_zero %2
+    %define xmm_zero %3
+    %define xmm_four %4
+    %define xmm_nan %5
+
+    Compare xmm_four, xmm_four, EQ_OQ, true
+    Compare xmm_neg_four, xmm_four, EQ_OQ, false
+    Compare xmm_zero, xmm_neg_zero, EQ_OQ, true
+    Compare xmm_zero, xmm_zero, EQ_OQ, true
+    Compare xmm_four, xmm_nan, EQ_OQ, false
+    Compare xmm_nan, xmm_four, EQ_OQ, false
+    Compare xmm_nan, xmm_nan, EQ_OQ, false
+
+    Compare xmm_four, xmm_neg_four, LT_OS, false
+    Compare xmm_neg_four, xmm_four, LT_OS, true
+    Compare xmm_zero, xmm_neg_zero, LT_OS, false
+    Compare xmm_neg_zero, xmm_zero, LT_OS, false
+    Compare xmm_nan, xmm_four, LT_OS, false
+    Compare xmm_four, xmm_nan, LT_OS, false
+    Compare xmm_nan, xmm_nan, LT_OS, false
+
+    Compare xmm_four, xmm_neg_four, LE_OS, false
+    Compare xmm_neg_four, xmm_four, LE_OS, true
+    Compare xmm_zero, xmm_neg_zero, LE_OS, true
+    Compare xmm_neg_zero, xmm_zero, LE_OS, true
+    Compare xmm_nan, xmm_four, LE_OS, false
+    Compare xmm_four, xmm_nan, LE_OS, false
+    Compare xmm_nan, xmm_nan, LE_OS, false
+
+    Compare xmm_four, xmm_neg_four, UNORD_Q, false
+    Compare xmm_neg_four, xmm_four, UNORD_Q, false
+    Compare xmm_zero, xmm_neg_zero, UNORD_Q, false
+    Compare xmm_neg_zero, xmm_zero, UNORD_Q, false
+    Compare xmm_nan, xmm_four, UNORD_Q, true
+    Compare xmm_four, xmm_nan, UNORD_Q, true
+    Compare xmm_nan, xmm_nan, UNORD_Q, true
+
+    Compare xmm_four, xmm_four, NEQ_UQ, false
+    Compare xmm_four, xmm_neg_four, NEQ_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NEQ_UQ, false
+    Compare xmm_zero, xmm_zero, NEQ_UQ, false
+    Compare xmm_four, xmm_nan, NEQ_UQ, true
+    Compare xmm_nan, xmm_four, NEQ_UQ, true
+    Compare xmm_nan, xmm_nan, NEQ_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NLT_US, true
+    Compare xmm_neg_four, xmm_four, NLT_US, false
+    Compare xmm_zero, xmm_neg_zero, NLT_US, true
+    Compare xmm_neg_zero, xmm_zero, NLT_US, true
+    Compare xmm_nan, xmm_four, NLT_US, true
+    Compare xmm_four, xmm_nan, NLT_US, true
+    Compare xmm_nan, xmm_nan, NLT_US, true
+
+    Compare xmm_four, xmm_neg_four, NLE_US, true
+    Compare xmm_neg_four, xmm_four, NLE_US, false
+    Compare xmm_zero, xmm_neg_zero, NLE_US, false
+    Compare xmm_neg_zero, xmm_zero, NLE_US, false
+    Compare xmm_nan, xmm_four, NLE_US, true
+    Compare xmm_four, xmm_nan, NLE_US, true
+    Compare xmm_nan, xmm_nan, NLE_US, true
+
+    Compare xmm_four, xmm_neg_four, ORD_Q, true
+    Compare xmm_neg_four, xmm_four, ORD_Q, true
+    Compare xmm_zero, xmm_neg_zero, ORD_Q, true
+    Compare xmm_neg_zero, xmm_zero, ORD_Q, true
+    Compare xmm_nan, xmm_four, ORD_Q, false
+    Compare xmm_four, xmm_nan, ORD_Q, false
+    Compare xmm_nan, xmm_nan, ORD_Q, false
+
+    Compare xmm_four, xmm_neg_four, EQ_UQ, false
+    Compare xmm_four, xmm_four, EQ_UQ, true
+    Compare xmm_zero, xmm_neg_zero, EQ_UQ, true
+    Compare xmm_zero, xmm_zero, EQ_UQ, true
+    Compare xmm_four, xmm_nan, EQ_UQ, true
+    Compare xmm_nan, xmm_four, EQ_UQ, true
+    Compare xmm_nan, xmm_nan, EQ_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NGE_US, false
+    Compare xmm_neg_four, xmm_four, NGE_US, true
+    Compare xmm_zero, xmm_neg_zero, NGE_US, false
+    Compare xmm_neg_zero, xmm_zero, NGE_US, false
+    Compare xmm_nan, xmm_four, NGE_US, true
+    Compare xmm_four, xmm_nan, NGE_US, true
+    Compare xmm_nan, xmm_nan, NGE_US, true
+
+    Compare xmm_four, xmm_neg_four, NGT_US, false
+    Compare xmm_neg_four, xmm_four, NGT_US, true
+    Compare xmm_zero, xmm_neg_zero, NGT_US, true
+    Compare xmm_neg_zero, xmm_zero, NGT_US, true
+    Compare xmm_nan, xmm_four, NGT_US, true
+    Compare xmm_four, xmm_nan, NGT_US, true
+    Compare xmm_nan, xmm_nan, NGT_US, true
+
+    Compare xmm_four, xmm_neg_four, FALSE_OQ, false
+    Compare xmm_neg_four, xmm_four, FALSE_OQ, false
+    Compare xmm_zero, xmm_neg_zero, FALSE_OQ, false
+    Compare xmm_neg_zero, xmm_zero, FALSE_OQ, false
+    Compare xmm_nan, xmm_four, FALSE_OQ, false
+    Compare xmm_four, xmm_nan, FALSE_OQ, false
+    Compare xmm_nan, xmm_nan, FALSE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, NEQ_OQ, true
+    Compare xmm_four, xmm_four, NEQ_OQ, false
+    Compare xmm_zero, xmm_neg_zero, NEQ_OQ, false
+    Compare xmm_zero, xmm_zero, NEQ_OQ, false
+    Compare xmm_four, xmm_nan, NEQ_OQ, false
+    Compare xmm_nan, xmm_four, NEQ_OQ, false
+    Compare xmm_nan, xmm_nan, NEQ_OQ, false
+
+    Compare xmm_four, xmm_neg_four, GE_OS, true
+    Compare xmm_neg_four, xmm_four, GE_OS, false
+    Compare xmm_zero, xmm_neg_zero, GE_OS, true
+    Compare xmm_neg_zero, xmm_zero, GE_OS, true
+    Compare xmm_nan, xmm_four, GE_OS, false
+    Compare xmm_four, xmm_nan, GE_OS, false
+    Compare xmm_nan, xmm_nan, GE_OS, false
+
+    Compare xmm_four, xmm_neg_four, GT_OS, true
+    Compare xmm_neg_four, xmm_four, GT_OS, false
+    Compare xmm_zero, xmm_neg_zero, GT_OS, false
+    Compare xmm_neg_zero, xmm_zero, GT_OS, false
+    Compare xmm_nan, xmm_four, GT_OS, false
+    Compare xmm_four, xmm_nan, GT_OS, false
+    Compare xmm_nan, xmm_nan, GT_OS, false
+
+    Compare xmm_four, xmm_neg_four, TRUE_UQ, true
+    Compare xmm_neg_four, xmm_four, TRUE_UQ, true
+    Compare xmm_zero, xmm_neg_zero, TRUE_UQ, true
+    Compare xmm_neg_zero, xmm_zero, TRUE_UQ, true
+    Compare xmm_nan, xmm_four, TRUE_UQ, true
+    Compare xmm_four, xmm_nan, TRUE_UQ, true
+    Compare xmm_nan, xmm_nan, TRUE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, EQ_OS, false
+    Compare xmm_neg_four, xmm_four, EQ_OS, false
+    Compare xmm_zero, xmm_neg_zero, EQ_OS, true
+    Compare xmm_neg_zero, xmm_zero, EQ_OS, true
+    Compare xmm_nan, xmm_four, EQ_OS, false
+    Compare xmm_four, xmm_nan, EQ_OS, false
+    Compare xmm_nan, xmm_nan, EQ_OS, false
+
+    Compare xmm_four, xmm_neg_four, LT_OQ, false
+    Compare xmm_neg_four, xmm_four, LT_OQ, true
+    Compare xmm_zero, xmm_neg_zero, LT_OQ, false
+    Compare xmm_neg_zero, xmm_zero, LT_OQ, false
+    Compare xmm_nan, xmm_four, LT_OQ, false
+    Compare xmm_four, xmm_nan, LT_OQ, false
+    Compare xmm_nan, xmm_nan, LT_OQ, false
+
+    Compare xmm_four, xmm_neg_four, LE_OQ, false
+    Compare xmm_neg_four, xmm_four, LE_OQ, true
+    Compare xmm_zero, xmm_neg_zero, LE_OQ, true
+    Compare xmm_neg_zero, xmm_zero, LE_OQ, true
+    Compare xmm_nan, xmm_four, LE_OQ, false
+    Compare xmm_four, xmm_nan, LE_OQ, false
+    Compare xmm_nan, xmm_nan, LE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, UNORD_S, false
+    Compare xmm_neg_four, xmm_four, UNORD_S, false
+    Compare xmm_zero, xmm_neg_zero, UNORD_S, false
+    Compare xmm_neg_zero, xmm_zero, UNORD_S, false
+    Compare xmm_nan, xmm_four, UNORD_S, true
+    Compare xmm_four, xmm_nan, UNORD_S, true
+    Compare xmm_nan, xmm_nan, UNORD_S, true
+
+    Compare xmm_four, xmm_neg_four, NEQ_US, true
+    Compare xmm_neg_four, xmm_four, NEQ_US, true
+    Compare xmm_zero, xmm_neg_zero, NEQ_US, false
+    Compare xmm_neg_zero, xmm_zero, NEQ_US, false
+    Compare xmm_nan, xmm_four, NEQ_US, true
+    Compare xmm_four, xmm_nan, NEQ_US, true
+    Compare xmm_nan, xmm_nan, NEQ_US, true
+
+    Compare xmm_four, xmm_neg_four, NLT_UQ, true
+    Compare xmm_neg_four, xmm_four, NLT_UQ, false
+    Compare xmm_zero, xmm_neg_zero, NLT_UQ, true
+    Compare xmm_neg_zero, xmm_zero, NLT_UQ, true
+    Compare xmm_nan, xmm_four, NLT_UQ, true
+    Compare xmm_four, xmm_nan, NLT_UQ, true
+    Compare xmm_nan, xmm_nan, NLT_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NLE_UQ, true
+    Compare xmm_neg_four, xmm_four, NLE_UQ, false
+    Compare xmm_zero, xmm_neg_zero, NLE_UQ, false
+    Compare xmm_neg_zero, xmm_zero, NLE_UQ, false
+    Compare xmm_nan, xmm_four, NLE_UQ, true
+    Compare xmm_four, xmm_nan, NLE_UQ, true
+    Compare xmm_nan, xmm_nan, NLE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, ORD_S, true
+    Compare xmm_neg_four, xmm_four, ORD_S, true
+    Compare xmm_zero, xmm_neg_zero, ORD_S, true
+    Compare xmm_neg_zero, xmm_zero, ORD_S, true
+    Compare xmm_nan, xmm_four, ORD_S, false
+    Compare xmm_four, xmm_nan, ORD_S, false
+    Compare xmm_nan, xmm_nan, ORD_S, false
+
+    Compare xmm_four, xmm_neg_four, EQ_US, false
+    Compare xmm_four, xmm_four, EQ_US, true
+    Compare xmm_zero, xmm_neg_zero, EQ_US, true
+    Compare xmm_zero, xmm_zero, EQ_US, true
+    Compare xmm_four, xmm_nan, EQ_US, true
+    Compare xmm_nan, xmm_four, EQ_US, true
+    Compare xmm_nan, xmm_nan, EQ_US, true
+
+    Compare xmm_four, xmm_neg_four, NGE_UQ, false
+    Compare xmm_neg_four, xmm_four, NGE_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NGE_UQ, false
+    Compare xmm_neg_zero, xmm_zero, NGE_UQ, false
+    Compare xmm_nan, xmm_four, NGE_UQ, true
+    Compare xmm_four, xmm_nan, NGE_UQ, true
+    Compare xmm_nan, xmm_nan, NGE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NGT_UQ, false
+    Compare xmm_neg_four, xmm_four, NGT_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NGT_UQ, true
+    Compare xmm_neg_zero, xmm_zero, NGT_UQ, true
+    Compare xmm_nan, xmm_four, NGT_UQ, true
+    Compare xmm_four, xmm_nan, NGT_UQ, true
+    Compare xmm_nan, xmm_nan, NGT_UQ, true
+
+    Compare xmm_four, xmm_neg_four, FALSE_OS, false
+    Compare xmm_neg_four, xmm_four, FALSE_OS, false
+    Compare xmm_zero, xmm_neg_zero, FALSE_OS, false
+    Compare xmm_neg_zero, xmm_zero, FALSE_OS, false
+    Compare xmm_nan, xmm_four, FALSE_OS, false
+    Compare xmm_four, xmm_nan, FALSE_OS, false
+    Compare xmm_nan, xmm_nan, FALSE_OS, false
+
+    Compare xmm_four, xmm_neg_four, NEQ_OS, true
+    Compare xmm_four, xmm_four, NEQ_OS, false
+    Compare xmm_zero, xmm_neg_zero, NEQ_OS, false
+    Compare xmm_zero, xmm_zero, NEQ_OS, false
+    Compare xmm_four, xmm_nan, NEQ_OS, false
+    Compare xmm_nan, xmm_four, NEQ_OS, false
+    Compare xmm_nan, xmm_nan, NEQ_OS, false
+
+    Compare xmm_four, xmm_neg_four, GE_OQ, true
+    Compare xmm_neg_four, xmm_four, GE_OQ, false
+    Compare xmm_zero, xmm_neg_zero, GE_OQ, true
+    Compare xmm_neg_zero, xmm_zero, GE_OQ, true
+    Compare xmm_nan, xmm_four, GE_OQ, false
+    Compare xmm_four, xmm_nan, GE_OQ, false
+    Compare xmm_nan, xmm_nan, GE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, GT_OQ, true
+    Compare xmm_neg_four, xmm_four, GT_OQ, false
+    Compare xmm_zero, xmm_neg_zero, GT_OQ, false
+    Compare xmm_neg_zero, xmm_zero, GT_OQ, false
+    Compare xmm_nan, xmm_four, GT_OQ, false
+    Compare xmm_four, xmm_nan, GT_OQ, false
+    Compare xmm_nan, xmm_nan, GT_OQ, false
+
+    Compare xmm_four, xmm_neg_four, TRUE_US, true
+    Compare xmm_neg_four, xmm_four, TRUE_US, true
+    Compare xmm_zero, xmm_neg_zero, TRUE_US, true
+    Compare xmm_neg_zero, xmm_zero, TRUE_US, true
+    Compare xmm_nan, xmm_four, TRUE_US, true
+    Compare xmm_four, xmm_nan, TRUE_US, true
+    Compare xmm_nan, xmm_nan, TRUE_US, true
+%endmacro
+
+%define temp xmm8
+%define result xmm7
+CompareAll xmm0, xmm1, xmm2, xmm3, xmm4
+%define temp ymm8
+%define result ymm7
+CompareAll ymm0, ymm1, ymm2, ymm3, ymm4
+
+; If ecx is not correct we can use the counter to determine which test failed
+.fail:
+hlt
+
+align 32
+.data:
+dq -4.0, -4.0, -4.0, -4.0
+
+dq -0.0, -0.0, -0.0, -0.0
+
+dq 0.0, 0.0, 0.0, 0.0
+
+dq 4.0, 4.0, 4.0, 4.0
+
+.nan:
+dq 0x7FF8000000000000
+dq 0x7FF8000000000000
+dq 0x7FF8000000000000
+dq 0x7FF8000000000000
+
+.false:
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+
+.true:
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF

unittests/ASM/VEX/vcmpps_full.asm

@@ -0,0 +1,373 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+    "RCX": "448"
+  }
+}
+%endif
+
+%define true 1
+%define false 0
+
+%define EQ_OQ 0
+%define LT_OS 1
+%define LE_OS 2
+%define UNORD_Q 3
+%define NEQ_UQ 4
+%define NLT_US 5
+%define NLE_US 6
+%define ORD_Q 7
+%define EQ_UQ 8
+%define NGE_US 9
+%define NGT_US 10
+%define FALSE_OQ 11
+%define NEQ_OQ 12
+%define GE_OS 13
+%define GT_OS 14
+%define TRUE_UQ 15
+%define EQ_OS 16
+%define LT_OQ 17
+%define LE_OQ 18
+%define UNORD_S 19
+%define NEQ_US 20
+%define NLT_UQ 21
+%define NLE_UQ 22
+%define ORD_S 23
+%define EQ_US 24
+%define NGE_UQ 25
+%define NGT_UQ 26
+%define FALSE_OS 27
+%define NEQ_OS 28
+%define GE_OQ 29
+%define GT_OQ 30
+%define TRUE_US 31
+
+; Arguments: src1, src2, predicate, true/false
+%macro Compare 4
+    vcmpps result, %1, %2, %3
+    ; Construct expected result
+    vmovdqa temp, [rel .false + 32 * %4]
+    ; Compare if result and expected are equal
+    vxorps ymm6, ymm7, ymm8
+    vptest ymm6, ymm6
+    jnz .fail
+    ; Increment counter of tests passed
+    add ecx, 1
+%endmacro
+
+lea rax, [rel .data]
+xor ecx, ecx
+
+vmovdqa ymm0, [rax + 32 * 0] ; -4.0
+vmovdqa ymm1, [rax + 32 * 1] ; -0.0
+vmovdqa ymm2, [rax + 32 * 2] ; 0.0
+vmovdqa ymm3, [rax + 32 * 3] ; 4.0
+vmovdqa ymm4, [rel .nan]     ; NaN
+
+%macro CompareAll 5
+    %define xmm_neg_four %1
+    %define xmm_neg_zero %2
+    %define xmm_zero %3
+    %define xmm_four %4
+    %define xmm_nan %5
+
+    Compare xmm_four, xmm_four, EQ_OQ, true
+    Compare xmm_neg_four, xmm_four, EQ_OQ, false
+    Compare xmm_zero, xmm_neg_zero, EQ_OQ, true
+    Compare xmm_zero, xmm_zero, EQ_OQ, true
+    Compare xmm_four, xmm_nan, EQ_OQ, false
+    Compare xmm_nan, xmm_four, EQ_OQ, false
+    Compare xmm_nan, xmm_nan, EQ_OQ, false
+
+    Compare xmm_four, xmm_neg_four, LT_OS, false
+    Compare xmm_neg_four, xmm_four, LT_OS, true
+    Compare xmm_zero, xmm_neg_zero, LT_OS, false
+    Compare xmm_neg_zero, xmm_zero, LT_OS, false
+    Compare xmm_nan, xmm_four, LT_OS, false
+    Compare xmm_four, xmm_nan, LT_OS, false
+    Compare xmm_nan, xmm_nan, LT_OS, false
+
+    Compare xmm_four, xmm_neg_four, LE_OS, false
+    Compare xmm_neg_four, xmm_four, LE_OS, true
+    Compare xmm_zero, xmm_neg_zero, LE_OS, true
+    Compare xmm_neg_zero, xmm_zero, LE_OS, true
+    Compare xmm_nan, xmm_four, LE_OS, false
+    Compare xmm_four, xmm_nan, LE_OS, false
+    Compare xmm_nan, xmm_nan, LE_OS, false
+
+    Compare xmm_four, xmm_neg_four, UNORD_Q, false
+    Compare xmm_neg_four, xmm_four, UNORD_Q, false
+    Compare xmm_zero, xmm_neg_zero, UNORD_Q, false
+    Compare xmm_neg_zero, xmm_zero, UNORD_Q, false
+    Compare xmm_nan, xmm_four, UNORD_Q, true
+    Compare xmm_four, xmm_nan, UNORD_Q, true
+    Compare xmm_nan, xmm_nan, UNORD_Q, true
+
+    Compare xmm_four, xmm_four, NEQ_UQ, false
+    Compare xmm_four, xmm_neg_four, NEQ_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NEQ_UQ, false
+    Compare xmm_zero, xmm_zero, NEQ_UQ, false
+    Compare xmm_four, xmm_nan, NEQ_UQ, true
+    Compare xmm_nan, xmm_four, NEQ_UQ, true
+    Compare xmm_nan, xmm_nan, NEQ_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NLT_US, true
+    Compare xmm_neg_four, xmm_four, NLT_US, false
+    Compare xmm_zero, xmm_neg_zero, NLT_US, true
+    Compare xmm_neg_zero, xmm_zero, NLT_US, true
+    Compare xmm_nan, xmm_four, NLT_US, true
+    Compare xmm_four, xmm_nan, NLT_US, true
+    Compare xmm_nan, xmm_nan, NLT_US, true
+
+    Compare xmm_four, xmm_neg_four, NLE_US, true
+    Compare xmm_neg_four, xmm_four, NLE_US, false
+    Compare xmm_zero, xmm_neg_zero, NLE_US, false
+    Compare xmm_neg_zero, xmm_zero, NLE_US, false
+    Compare xmm_nan, xmm_four, NLE_US, true
+    Compare xmm_four, xmm_nan, NLE_US, true
+    Compare xmm_nan, xmm_nan, NLE_US, true
+
+    Compare xmm_four, xmm_neg_four, ORD_Q, true
+    Compare xmm_neg_four, xmm_four, ORD_Q, true
+    Compare xmm_zero, xmm_neg_zero, ORD_Q, true
+    Compare xmm_neg_zero, xmm_zero, ORD_Q, true
+    Compare xmm_nan, xmm_four, ORD_Q, false
+    Compare xmm_four, xmm_nan, ORD_Q, false
+    Compare xmm_nan, xmm_nan, ORD_Q, false
+
+    Compare xmm_four, xmm_neg_four, EQ_UQ, false
+    Compare xmm_four, xmm_four, EQ_UQ, true
+    Compare xmm_zero, xmm_neg_zero, EQ_UQ, true
+    Compare xmm_zero, xmm_zero, EQ_UQ, true
+    Compare xmm_four, xmm_nan, EQ_UQ, true
+    Compare xmm_nan, xmm_four, EQ_UQ, true
+    Compare xmm_nan, xmm_nan, EQ_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NGE_US, false
+    Compare xmm_neg_four, xmm_four, NGE_US, true
+    Compare xmm_zero, xmm_neg_zero, NGE_US, false
+    Compare xmm_neg_zero, xmm_zero, NGE_US, false
+    Compare xmm_nan, xmm_four, NGE_US, true
+    Compare xmm_four, xmm_nan, NGE_US, true
+    Compare xmm_nan, xmm_nan, NGE_US, true
+
+    Compare xmm_four, xmm_neg_four, NGT_US, false
+    Compare xmm_neg_four, xmm_four, NGT_US, true
+    Compare xmm_zero, xmm_neg_zero, NGT_US, true
+    Compare xmm_neg_zero, xmm_zero, NGT_US, true
+    Compare xmm_nan, xmm_four, NGT_US, true
+    Compare xmm_four, xmm_nan, NGT_US, true
+    Compare xmm_nan, xmm_nan, NGT_US, true
+
+    Compare xmm_four, xmm_neg_four, FALSE_OQ, false
+    Compare xmm_neg_four, xmm_four, FALSE_OQ, false
+    Compare xmm_zero, xmm_neg_zero, FALSE_OQ, false
+    Compare xmm_neg_zero, xmm_zero, FALSE_OQ, false
+    Compare xmm_nan, xmm_four, FALSE_OQ, false
+    Compare xmm_four, xmm_nan, FALSE_OQ, false
+    Compare xmm_nan, xmm_nan, FALSE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, NEQ_OQ, true
+    Compare xmm_four, xmm_four, NEQ_OQ, false
+    Compare xmm_zero, xmm_neg_zero, NEQ_OQ, false
+    Compare xmm_zero, xmm_zero, NEQ_OQ, false
+    Compare xmm_four, xmm_nan, NEQ_OQ, false
+    Compare xmm_nan, xmm_four, NEQ_OQ, false
+    Compare xmm_nan, xmm_nan, NEQ_OQ, false
+
+    Compare xmm_four, xmm_neg_four, GE_OS, true
+    Compare xmm_neg_four, xmm_four, GE_OS, false
+    Compare xmm_zero, xmm_neg_zero, GE_OS, true
+    Compare xmm_neg_zero, xmm_zero, GE_OS, true
+    Compare xmm_nan, xmm_four, GE_OS, false
+    Compare xmm_four, xmm_nan, GE_OS, false
+    Compare xmm_nan, xmm_nan, GE_OS, false
+
+    Compare xmm_four, xmm_neg_four, GT_OS, true
+    Compare xmm_neg_four, xmm_four, GT_OS, false
+    Compare xmm_zero, xmm_neg_zero, GT_OS, false
+    Compare xmm_neg_zero, xmm_zero, GT_OS, false
+    Compare xmm_nan, xmm_four, GT_OS, false
+    Compare xmm_four, xmm_nan, GT_OS, false
+    Compare xmm_nan, xmm_nan, GT_OS, false
+
+    Compare xmm_four, xmm_neg_four, TRUE_UQ, true
+    Compare xmm_neg_four, xmm_four, TRUE_UQ, true
+    Compare xmm_zero, xmm_neg_zero, TRUE_UQ, true
+    Compare xmm_neg_zero, xmm_zero, TRUE_UQ, true
+    Compare xmm_nan, xmm_four, TRUE_UQ, true
+    Compare xmm_four, xmm_nan, TRUE_UQ, true
+    Compare xmm_nan, xmm_nan, TRUE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, EQ_OS, false
+    Compare xmm_neg_four, xmm_four, EQ_OS, false
+    Compare xmm_zero, xmm_neg_zero, EQ_OS, true
+    Compare xmm_neg_zero, xmm_zero, EQ_OS, true
+    Compare xmm_nan, xmm_four, EQ_OS, false
+    Compare xmm_four, xmm_nan, EQ_OS, false
+    Compare xmm_nan, xmm_nan, EQ_OS, false
+
+    Compare xmm_four, xmm_neg_four, LT_OQ, false
+    Compare xmm_neg_four, xmm_four, LT_OQ, true
+    Compare xmm_zero, xmm_neg_zero, LT_OQ, false
+    Compare xmm_neg_zero, xmm_zero, LT_OQ, false
+    Compare xmm_nan, xmm_four, LT_OQ, false
+    Compare xmm_four, xmm_nan, LT_OQ, false
+    Compare xmm_nan, xmm_nan, LT_OQ, false
+
+    Compare xmm_four, xmm_neg_four, LE_OQ, false
+    Compare xmm_neg_four, xmm_four, LE_OQ, true
+    Compare xmm_zero, xmm_neg_zero, LE_OQ, true
+    Compare xmm_neg_zero, xmm_zero, LE_OQ, true
+    Compare xmm_nan, xmm_four, LE_OQ, false
+    Compare xmm_four, xmm_nan, LE_OQ, false
+    Compare xmm_nan, xmm_nan, LE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, UNORD_S, false
+    Compare xmm_neg_four, xmm_four, UNORD_S, false
+    Compare xmm_zero, xmm_neg_zero, UNORD_S, false
+    Compare xmm_neg_zero, xmm_zero, UNORD_S, false
+    Compare xmm_nan, xmm_four, UNORD_S, true
+    Compare xmm_four, xmm_nan, UNORD_S, true
+    Compare xmm_nan, xmm_nan, UNORD_S, true
+
+    Compare xmm_four, xmm_neg_four, NEQ_US, true
+    Compare xmm_neg_four, xmm_four, NEQ_US, true
+    Compare xmm_zero, xmm_neg_zero, NEQ_US, false
+    Compare xmm_neg_zero, xmm_zero, NEQ_US, false
+    Compare xmm_nan, xmm_four, NEQ_US, true
+    Compare xmm_four, xmm_nan, NEQ_US, true
+    Compare xmm_nan, xmm_nan, NEQ_US, true
+
+    Compare xmm_four, xmm_neg_four, NLT_UQ, true
+    Compare xmm_neg_four, xmm_four, NLT_UQ, false
+    Compare xmm_zero, xmm_neg_zero, NLT_UQ, true
+    Compare xmm_neg_zero, xmm_zero, NLT_UQ, true
+    Compare xmm_nan, xmm_four, NLT_UQ, true
+    Compare xmm_four, xmm_nan, NLT_UQ, true
+    Compare xmm_nan, xmm_nan, NLT_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NLE_UQ, true
+    Compare xmm_neg_four, xmm_four, NLE_UQ, false
+    Compare xmm_zero, xmm_neg_zero, NLE_UQ, false
+    Compare xmm_neg_zero, xmm_zero, NLE_UQ, false
+    Compare xmm_nan, xmm_four, NLE_UQ, true
+    Compare xmm_four, xmm_nan, NLE_UQ, true
+    Compare xmm_nan, xmm_nan, NLE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, ORD_S, true
+    Compare xmm_neg_four, xmm_four, ORD_S, true
+    Compare xmm_zero, xmm_neg_zero, ORD_S, true
+    Compare xmm_neg_zero, xmm_zero, ORD_S, true
+    Compare xmm_nan, xmm_four, ORD_S, false
+    Compare xmm_four, xmm_nan, ORD_S, false
+    Compare xmm_nan, xmm_nan, ORD_S, false
+
+    Compare xmm_four, xmm_neg_four, EQ_US, false
+    Compare xmm_four, xmm_four, EQ_US, true
+    Compare xmm_zero, xmm_neg_zero, EQ_US, true
+    Compare xmm_zero, xmm_zero, EQ_US, true
+    Compare xmm_four, xmm_nan, EQ_US, true
+    Compare xmm_nan, xmm_four, EQ_US, true
+    Compare xmm_nan, xmm_nan, EQ_US, true
+
+    Compare xmm_four, xmm_neg_four, NGE_UQ, false
+    Compare xmm_neg_four, xmm_four, NGE_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NGE_UQ, false
+    Compare xmm_neg_zero, xmm_zero, NGE_UQ, false
+    Compare xmm_nan, xmm_four, NGE_UQ, true
+    Compare xmm_four, xmm_nan, NGE_UQ, true
+    Compare xmm_nan, xmm_nan, NGE_UQ, true
+
+    Compare xmm_four, xmm_neg_four, NGT_UQ, false
+    Compare xmm_neg_four, xmm_four, NGT_UQ, true
+    Compare xmm_zero, xmm_neg_zero, NGT_UQ, true
+    Compare xmm_neg_zero, xmm_zero, NGT_UQ, true
+    Compare xmm_nan, xmm_four, NGT_UQ, true
+    Compare xmm_four, xmm_nan, NGT_UQ, true
+    Compare xmm_nan, xmm_nan, NGT_UQ, true
+
+    Compare xmm_four, xmm_neg_four, FALSE_OS, false
+    Compare xmm_neg_four, xmm_four, FALSE_OS, false
+    Compare xmm_zero, xmm_neg_zero, FALSE_OS, false
+    Compare xmm_neg_zero, xmm_zero, FALSE_OS, false
+    Compare xmm_nan, xmm_four, FALSE_OS, false
+    Compare xmm_four, xmm_nan, FALSE_OS, false
+    Compare xmm_nan, xmm_nan, FALSE_OS, false
+
+    Compare xmm_four, xmm_neg_four, NEQ_OS, true
+    Compare xmm_four, xmm_four, NEQ_OS, false
+    Compare xmm_zero, xmm_neg_zero, NEQ_OS, false
+    Compare xmm_zero, xmm_zero, NEQ_OS, false
+    Compare xmm_four, xmm_nan, NEQ_OS, false
+    Compare xmm_nan, xmm_four, NEQ_OS, false
+    Compare xmm_nan, xmm_nan, NEQ_OS, false
+
+    Compare xmm_four, xmm_neg_four, GE_OQ, true
+    Compare xmm_neg_four, xmm_four, GE_OQ, false
+    Compare xmm_zero, xmm_neg_zero, GE_OQ, true
+    Compare xmm_neg_zero, xmm_zero, GE_OQ, true
+    Compare xmm_nan, xmm_four, GE_OQ, false
+    Compare xmm_four, xmm_nan, GE_OQ, false
+    Compare xmm_nan, xmm_nan, GE_OQ, false
+
+    Compare xmm_four, xmm_neg_four, GT_OQ, true
+    Compare xmm_neg_four, xmm_four, GT_OQ, false
+    Compare xmm_zero, xmm_neg_zero, GT_OQ, false
+    Compare xmm_neg_zero, xmm_zero, GT_OQ, false
+    Compare xmm_nan, xmm_four, GT_OQ, false
+    Compare xmm_four, xmm_nan, GT_OQ, false
+    Compare xmm_nan, xmm_nan, GT_OQ, false
+
+    Compare xmm_four, xmm_neg_four, TRUE_US, true
+    Compare xmm_neg_four, xmm_four, TRUE_US, true
+    Compare xmm_zero, xmm_neg_zero, TRUE_US, true
+    Compare xmm_neg_zero, xmm_zero, TRUE_US, true
+    Compare xmm_nan, xmm_four, TRUE_US, true
+    Compare xmm_four, xmm_nan, TRUE_US, true
+    Compare xmm_nan, xmm_nan, TRUE_US, true
+%endmacro
+
+%define temp xmm8
+%define result xmm7
+CompareAll xmm0, xmm1, xmm2, xmm3, xmm4
+%define temp ymm8
+%define result ymm7
+CompareAll ymm0, ymm1, ymm2, ymm3, ymm4
+
+; If ecx is not correct we can use the counter to determine which test failed
+.fail:
+hlt
+
+align 32
+.data:
+dd -4.0, -4.0, -4.0, -4.0
+dd -4.0, -4.0, -4.0, -4.0
+
+dd -0.0, -0.0, -0.0, -0.0
+dd -0.0, -0.0, -0.0, -0.0
+
+dd 0.0, 0.0, 0.0, 0.0
+dd 0.0, 0.0, 0.0, 0.0
+
+dd 4.0, 4.0, 4.0, 4.0
+dd 4.0, 4.0, 4.0, 4.0
+
+.nan:
+dq 0x7FC000007FC00000
+dq 0x7FC000007FC00000
+dq 0x7FC000007FC00000
+dq 0x7FC000007FC00000
+
+.false:
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+
+.true:
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF
+dq 0xFFFFFFFFFFFFFFFF

unittests/ASM/VEX/vcmpsd_full.asm

@@ -0,0 +1,375 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+    "RCX": "224"
+  }
+}
+%endif
+
+%define true 1
+%define false 0
+
+%define EQ_OQ 0
+%define LT_OS 1
+%define LE_OS 2
+%define UNORD_Q 3
+%define NEQ_UQ 4
+%define NLT_US 5
+%define NLE_US 6
+%define ORD_Q 7
+%define EQ_UQ 8
+%define NGE_US 9
+%define NGT_US 10
+%define FALSE_OQ 11
+%define NEQ_OQ 12
+%define GE_OS 13
+%define GT_OS 14
+%define TRUE_UQ 15
+%define EQ_OS 16
+%define LT_OQ 17
+%define LE_OQ 18
+%define UNORD_S 19
+%define NEQ_US 20
+%define NLT_UQ 21
+%define NLE_UQ 22
+%define ORD_S 23
+%define EQ_US 24
+%define NGE_UQ 25
+%define NGT_UQ 26
+%define FALSE_OS 27
+%define NEQ_OS 28
+%define GE_OQ 29
+%define GT_OQ 30
+%define TRUE_US 31
+
+; Arguments: src1, src2, predicate, true/false
+%macro Compare 4
+    vcmpsd xmm7, %1, %2, %3
+    ; Construct expected result
+    ; Move truthy/falsey value to the first element based on %4
+    ; and [127:64] from src1
+    vmovdqa xmm8, [rel .false + 32 * %4]
+    vmovsd xmm5, %1, xmm8
+    ; Compare if result and expected are equal
+    vxorpd ymm6, ymm7, ymm5
+    vptest ymm6, ymm6
+    jnz .fail
+    ; Increment counter of tests passed
+    add ecx, 1
+%endmacro
+
+lea rax, [rel .data]
+xor ecx, ecx
+
+vmovdqa ymm0, [rax + 32 * 0] ; -4.0
+vmovdqa ymm1, [rax + 32 * 1] ; -0.0
+vmovdqa ymm2, [rax + 32 * 2] ; 0.0
+vmovdqa ymm3, [rax + 32 * 3] ; 4.0
+vmovdqa ymm4, [rel .nan]     ; NaN
+
+%define xmm_neg_four xmm0
+%define xmm_neg_zero xmm1
+%define xmm_zero xmm2
+%define xmm_four xmm3
+%define xmm_nan xmm4
+
+Compare xmm_four, xmm_four, EQ_OQ, true
+Compare xmm_neg_four, xmm_four, EQ_OQ, false
+Compare xmm_zero, xmm_neg_zero, EQ_OQ, true
+Compare xmm_zero, xmm_zero, EQ_OQ, true
+Compare xmm_four, xmm_nan, EQ_OQ, false
+Compare xmm_nan, xmm_four, EQ_OQ, false
+Compare xmm_nan, xmm_nan, EQ_OQ, false
+
+Compare xmm_four, xmm_neg_four, LT_OS, false
+Compare xmm_neg_four, xmm_four, LT_OS, true
+Compare xmm_zero, xmm_neg_zero, LT_OS, false
+Compare xmm_neg_zero, xmm_zero, LT_OS, false
+Compare xmm_nan, xmm_four, LT_OS, false
+Compare xmm_four, xmm_nan, LT_OS, false
+Compare xmm_nan, xmm_nan, LT_OS, false
+
+Compare xmm_four, xmm_neg_four, LE_OS, false
+Compare xmm_neg_four, xmm_four, LE_OS, true
+Compare xmm_zero, xmm_neg_zero, LE_OS, true
+Compare xmm_neg_zero, xmm_zero, LE_OS, true
+Compare xmm_nan, xmm_four, LE_OS, false
+Compare xmm_four, xmm_nan, LE_OS, false
+Compare xmm_nan, xmm_nan, LE_OS, false
+
+Compare xmm_four, xmm_neg_four, UNORD_Q, false
+Compare xmm_neg_four, xmm_four, UNORD_Q, false
+Compare xmm_zero, xmm_neg_zero, UNORD_Q, false
+Compare xmm_neg_zero, xmm_zero, UNORD_Q, false
+Compare xmm_nan, xmm_four, UNORD_Q, true
+Compare xmm_four, xmm_nan, UNORD_Q, true
+Compare xmm_nan, xmm_nan, UNORD_Q, true
+
+Compare xmm_four, xmm_four, NEQ_UQ, false
+Compare xmm_four, xmm_neg_four, NEQ_UQ, true
+Compare xmm_zero, xmm_neg_zero, NEQ_UQ, false
+Compare xmm_zero, xmm_zero, NEQ_UQ, false
+Compare xmm_four, xmm_nan, NEQ_UQ, true
+Compare xmm_nan, xmm_four, NEQ_UQ, true
+Compare xmm_nan, xmm_nan, NEQ_UQ, true
+
+Compare xmm_four, xmm_neg_four, NLT_US, true
+Compare xmm_neg_four, xmm_four, NLT_US, false
+Compare xmm_zero, xmm_neg_zero, NLT_US, true
+Compare xmm_neg_zero, xmm_zero, NLT_US, true
+Compare xmm_nan, xmm_four, NLT_US, true
+Compare xmm_four, xmm_nan, NLT_US, true
+Compare xmm_nan, xmm_nan, NLT_US, true
+
+Compare xmm_four, xmm_neg_four, NLE_US, true
+Compare xmm_neg_four, xmm_four, NLE_US, false
+Compare xmm_zero, xmm_neg_zero, NLE_US, false
+Compare xmm_neg_zero, xmm_zero, NLE_US, false
+Compare xmm_nan, xmm_four, NLE_US, true
+Compare xmm_four, xmm_nan, NLE_US, true
+Compare xmm_nan, xmm_nan, NLE_US, true
+
+Compare xmm_four, xmm_neg_four, ORD_Q, true
+Compare xmm_neg_four, xmm_four, ORD_Q, true
+Compare xmm_zero, xmm_neg_zero, ORD_Q, true
+Compare xmm_neg_zero, xmm_zero, ORD_Q, true
+Compare xmm_nan, xmm_four, ORD_Q, false
+Compare xmm_four, xmm_nan, ORD_Q, false
+Compare xmm_nan, xmm_nan, ORD_Q, false
+
+Compare xmm_four, xmm_neg_four, EQ_UQ, false
+Compare xmm_four, xmm_four, EQ_UQ, true
+Compare xmm_zero, xmm_neg_zero, EQ_UQ, true
+Compare xmm_zero, xmm_zero, EQ_UQ, true
+Compare xmm_four, xmm_nan, EQ_UQ, true
+Compare xmm_nan, xmm_four, EQ_UQ, true
+Compare xmm_nan, xmm_nan, EQ_UQ, true
+
+Compare xmm_four, xmm_neg_four, NGE_US, false
+Compare xmm_neg_four, xmm_four, NGE_US, true
+Compare xmm_zero, xmm_neg_zero, NGE_US, false
+Compare xmm_neg_zero, xmm_zero, NGE_US, false
+Compare xmm_nan, xmm_four, NGE_US, true
+Compare xmm_four, xmm_nan, NGE_US, true
+Compare xmm_nan, xmm_nan, NGE_US, true
+
+Compare xmm_four, xmm_neg_four, NGT_US, false
+Compare xmm_neg_four, xmm_four, NGT_US, true
+Compare xmm_zero, xmm_neg_zero, NGT_US, true
+Compare xmm_neg_zero, xmm_zero, NGT_US, true
+Compare xmm_nan, xmm_four, NGT_US, true
+Compare xmm_four, xmm_nan, NGT_US, true
+Compare xmm_nan, xmm_nan, NGT_US, true
+
+Compare xmm_four, xmm_neg_four, FALSE_OQ, false
+Compare xmm_neg_four, xmm_four, FALSE_OQ, false
+Compare xmm_zero, xmm_neg_zero, FALSE_OQ, false
+Compare xmm_neg_zero, xmm_zero, FALSE_OQ, false
+Compare xmm_nan, xmm_four, FALSE_OQ, false
+Compare xmm_four, xmm_nan, FALSE_OQ, false
+Compare xmm_nan, xmm_nan, FALSE_OQ, false
+
+Compare xmm_four, xmm_neg_four, NEQ_OQ, true
+Compare xmm_four, xmm_four, NEQ_OQ, false
+Compare xmm_zero, xmm_neg_zero, NEQ_OQ, false
+Compare xmm_zero, xmm_zero, NEQ_OQ, false
+Compare xmm_four, xmm_nan, NEQ_OQ, false
+Compare xmm_nan, xmm_four, NEQ_OQ, false
+Compare xmm_nan, xmm_nan, NEQ_OQ, false
+
+Compare xmm_four, xmm_neg_four, GE_OS, true
+Compare xmm_neg_four, xmm_four, GE_OS, false
+Compare xmm_zero, xmm_neg_zero, GE_OS, true
+Compare xmm_neg_zero, xmm_zero, GE_OS, true
+Compare xmm_nan, xmm_four, GE_OS, false
+Compare xmm_four, xmm_nan, GE_OS, false
+Compare xmm_nan, xmm_nan, GE_OS, false
+
+Compare xmm_four, xmm_neg_four, GT_OS, true
+Compare xmm_neg_four, xmm_four, GT_OS, false
+Compare xmm_zero, xmm_neg_zero, GT_OS, false
+Compare xmm_neg_zero, xmm_zero, GT_OS, false
+Compare xmm_nan, xmm_four, GT_OS, false
+Compare xmm_four, xmm_nan, GT_OS, false
+Compare xmm_nan, xmm_nan, GT_OS, false
+
+Compare xmm_four, xmm_neg_four, TRUE_UQ, true
+Compare xmm_neg_four, xmm_four, TRUE_UQ, true
+Compare xmm_zero, xmm_neg_zero, TRUE_UQ, true
+Compare xmm_neg_zero, xmm_zero, TRUE_UQ, true
+Compare xmm_nan, xmm_four, TRUE_UQ, true
+Compare xmm_four, xmm_nan, TRUE_UQ, true
+Compare xmm_nan, xmm_nan, TRUE_UQ, true
+
+Compare xmm_four, xmm_neg_four, EQ_OS, false
+Compare xmm_neg_four, xmm_four, EQ_OS, false
+Compare xmm_zero, xmm_neg_zero, EQ_OS, true
+Compare xmm_neg_zero, xmm_zero, EQ_OS, true
+Compare xmm_nan, xmm_four, EQ_OS, false
+Compare xmm_four, xmm_nan, EQ_OS, false
+Compare xmm_nan, xmm_nan, EQ_OS, false
+
+Compare xmm_four, xmm_neg_four, LT_OQ, false
+Compare xmm_neg_four, xmm_four, LT_OQ, true
+Compare xmm_zero, xmm_neg_zero, LT_OQ, false
+Compare xmm_neg_zero, xmm_zero, LT_OQ, false
+Compare xmm_nan, xmm_four, LT_OQ, false
+Compare xmm_four, xmm_nan, LT_OQ, false
+Compare xmm_nan, xmm_nan, LT_OQ, false
+
+Compare xmm_four, xmm_neg_four, LE_OQ, false
+Compare xmm_neg_four, xmm_four, LE_OQ, true
+Compare xmm_zero, xmm_neg_zero, LE_OQ, true
+Compare xmm_neg_zero, xmm_zero, LE_OQ, true
+Compare xmm_nan, xmm_four, LE_OQ, false
+Compare xmm_four, xmm_nan, LE_OQ, false
+Compare xmm_nan, xmm_nan, LE_OQ, false
+
+Compare xmm_four, xmm_neg_four, UNORD_S, false
+Compare xmm_neg_four, xmm_four, UNORD_S, false
+Compare xmm_zero, xmm_neg_zero, UNORD_S, false
+Compare xmm_neg_zero, xmm_zero, UNORD_S, false
+Compare xmm_nan, xmm_four, UNORD_S, true
+Compare xmm_four, xmm_nan, UNORD_S, true
+Compare xmm_nan, xmm_nan, UNORD_S, true
+
+Compare xmm_four, xmm_neg_four, NEQ_US, true
+Compare xmm_neg_four, xmm_four, NEQ_US, true
+Compare xmm_zero, xmm_neg_zero, NEQ_US, false
+Compare xmm_neg_zero, xmm_zero, NEQ_US, false
+Compare xmm_nan, xmm_four, NEQ_US, true
+Compare xmm_four, xmm_nan, NEQ_US, true
+Compare xmm_nan, xmm_nan, NEQ_US, true
+
+Compare xmm_four, xmm_neg_four, NLT_UQ, true
+Compare xmm_neg_four, xmm_four, NLT_UQ, false
+Compare xmm_zero, xmm_neg_zero, NLT_UQ, true
+Compare xmm_neg_zero, xmm_zero, NLT_UQ, true
+Compare xmm_nan, xmm_four, NLT_UQ, true
+Compare xmm_four, xmm_nan, NLT_UQ, true
+Compare xmm_nan, xmm_nan, NLT_UQ, true
+
+Compare xmm_four, xmm_neg_four, NLE_UQ, true
+Compare xmm_neg_four, xmm_four, NLE_UQ, false
+Compare xmm_zero, xmm_neg_zero, NLE_UQ, false
+Compare xmm_neg_zero, xmm_zero, NLE_UQ, false
+Compare xmm_nan, xmm_four, NLE_UQ, true
+Compare xmm_four, xmm_nan, NLE_UQ, true
+Compare xmm_nan, xmm_nan, NLE_UQ, true
+
+Compare xmm_four, xmm_neg_four, ORD_S, true
+Compare xmm_neg_four, xmm_four, ORD_S, true
+Compare xmm_zero, xmm_neg_zero, ORD_S, true
+Compare xmm_neg_zero, xmm_zero, ORD_S, true
+Compare xmm_nan, xmm_four, ORD_S, false
+Compare xmm_four, xmm_nan, ORD_S, false
+Compare xmm_nan, xmm_nan, ORD_S, false
+
+Compare xmm_four, xmm_neg_four, EQ_US, false
+Compare xmm_four, xmm_four, EQ_US, true
+Compare xmm_zero, xmm_neg_zero, EQ_US, true
+Compare xmm_zero, xmm_zero, EQ_US, true
+Compare xmm_four, xmm_nan, EQ_US, true
+Compare xmm_nan, xmm_four, EQ_US, true
+Compare xmm_nan, xmm_nan, EQ_US, true
+
+Compare xmm_four, xmm_neg_four, NGE_UQ, false
+Compare xmm_neg_four, xmm_four, NGE_UQ, true
+Compare xmm_zero, xmm_neg_zero, NGE_UQ, false
+Compare xmm_neg_zero, xmm_zero, NGE_UQ, false
+Compare xmm_nan, xmm_four, NGE_UQ, true
+Compare xmm_four, xmm_nan, NGE_UQ, true
+Compare xmm_nan, xmm_nan, NGE_UQ, true
+
+Compare xmm_four, xmm_neg_four, NGT_UQ, false
+Compare xmm_neg_four, xmm_four, NGT_UQ, true
+Compare xmm_zero, xmm_neg_zero, NGT_UQ, true
+Compare xmm_neg_zero, xmm_zero, NGT_UQ, true
+Compare xmm_nan, xmm_four, NGT_UQ, true
+Compare xmm_four, xmm_nan, NGT_UQ, true
+Compare xmm_nan, xmm_nan, NGT_UQ, true
+
+Compare xmm_four, xmm_neg_four, FALSE_OS, false
+Compare xmm_neg_four, xmm_four, FALSE_OS, false
+Compare xmm_zero, xmm_neg_zero, FALSE_OS, false
+Compare xmm_neg_zero, xmm_zero, FALSE_OS, false
+Compare xmm_nan, xmm_four, FALSE_OS, false
+Compare xmm_four, xmm_nan, FALSE_OS, false
+Compare xmm_nan, xmm_nan, FALSE_OS, false
+
+Compare xmm_four, xmm_neg_four, NEQ_OS, true
+Compare xmm_four, xmm_four, NEQ_OS, false
+Compare xmm_zero, xmm_neg_zero, NEQ_OS, false
+Compare xmm_zero, xmm_zero, NEQ_OS, false
+Compare xmm_four, xmm_nan, NEQ_OS, false
+Compare xmm_nan, xmm_four, NEQ_OS, false
+Compare xmm_nan, xmm_nan, NEQ_OS, false
+
+Compare xmm_four, xmm_neg_four, GE_OQ, true
+Compare xmm_neg_four, xmm_four, GE_OQ, false
+Compare xmm_zero, xmm_neg_zero, GE_OQ, true
+Compare xmm_neg_zero, xmm_zero, GE_OQ, true
+Compare xmm_nan, xmm_four, GE_OQ, false
+Compare xmm_four, xmm_nan, GE_OQ, false
+Compare xmm_nan, xmm_nan, GE_OQ, false
+
+Compare xmm_four, xmm_neg_four, GT_OQ, true
+Compare xmm_neg_four, xmm_four, GT_OQ, false
+Compare xmm_zero, xmm_neg_zero, GT_OQ, false
+Compare xmm_neg_zero, xmm_zero, GT_OQ, false
+Compare xmm_nan, xmm_four, GT_OQ, false
+Compare xmm_four, xmm_nan, GT_OQ, false
+Compare xmm_nan, xmm_nan, GT_OQ, false
+
+Compare xmm_four, xmm_neg_four, TRUE_US, true
+Compare xmm_neg_four, xmm_four, TRUE_US, true
+Compare xmm_zero, xmm_neg_zero, TRUE_US, true
+Compare xmm_neg_zero, xmm_zero, TRUE_US, true
+Compare xmm_nan, xmm_four, TRUE_US, true
+Compare xmm_four, xmm_nan, TRUE_US, true
+Compare xmm_nan, xmm_nan, TRUE_US, true
+
+; If ecx is not correct we can use the counter to determine which test failed
+.fail:
+hlt
+
+align 32
+.data:
+dq -4.0
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dq -0.0
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dq 0.0
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dq 4.0
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+.nan:
+dq 0x7FF8000000000000
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+.false:
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+
+.true:
+dq 0xFFFFFFFFFFFFFFFF
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000

unittests/ASM/VEX/vcmpss_full.asm

@@ -0,0 +1,379 @@
+%ifdef CONFIG
+{
+  "HostFeatures": ["AVX"],
+  "RegData": {
+    "RCX": "224"
+  }
+}
+%endif
+
+%define true 1
+%define false 0
+
+%define EQ_OQ 0
+%define LT_OS 1
+%define LE_OS 2
+%define UNORD_Q 3
+%define NEQ_UQ 4
+%define NLT_US 5
+%define NLE_US 6
+%define ORD_Q 7
+%define EQ_UQ 8
+%define NGE_US 9
+%define NGT_US 10
+%define FALSE_OQ 11
+%define NEQ_OQ 12
+%define GE_OS 13
+%define GT_OS 14
+%define TRUE_UQ 15
+%define EQ_OS 16
+%define LT_OQ 17
+%define LE_OQ 18
+%define UNORD_S 19
+%define NEQ_US 20
+%define NLT_UQ 21
+%define NLE_UQ 22
+%define ORD_S 23
+%define EQ_US 24
+%define NGE_UQ 25
+%define NGT_UQ 26
+%define FALSE_OS 27
+%define NEQ_OS 28
+%define GE_OQ 29
+%define GT_OQ 30
+%define TRUE_US 31
+
+; Arguments: src1, src2, predicate, true/false
+%macro Compare 4
+    vcmpss xmm7, %1, %2, %3
+    ; Construct expected result
+    ; Move truthy/falsey value to the first element based on %4
+    ; and [127:32] from src1
+    vmovdqa xmm8, [rel .false + 32 * %4]
+    vmovss xmm5, %1, xmm8
+    ; Compare if result and expected are equal
+    vxorps ymm6, ymm7, ymm5
+    vptest ymm6, ymm6
+    jnz .fail
+    ; Increment counter of tests passed
+    add ecx, 1
+%endmacro
+
+lea rax, [rel .data]
+xor ecx, ecx
+
+vmovdqa ymm0, [rax + 32 * 0] ; -4.0
+vmovdqa ymm1, [rax + 32 * 1] ; -0.0
+vmovdqa ymm2, [rax + 32 * 2] ; 0.0
+vmovdqa ymm3, [rax + 32 * 3] ; 4.0
+vmovdqa ymm4, [rel .nan]     ; NaN
+
+%define xmm_neg_four xmm0
+%define xmm_neg_zero xmm1
+%define xmm_zero xmm2
+%define xmm_four xmm3
+%define xmm_nan xmm4
+
+Compare xmm_four, xmm_four, EQ_OQ, true
+Compare xmm_neg_four, xmm_four, EQ_OQ, false
+Compare xmm_zero, xmm_neg_zero, EQ_OQ, true
+Compare xmm_zero, xmm_zero, EQ_OQ, true
+Compare xmm_four, xmm_nan, EQ_OQ, false
+Compare xmm_nan, xmm_four, EQ_OQ, false
+Compare xmm_nan, xmm_nan, EQ_OQ, false
+
+Compare xmm_four, xmm_neg_four, LT_OS, false
+Compare xmm_neg_four, xmm_four, LT_OS, true
+Compare xmm_zero, xmm_neg_zero, LT_OS, false
+Compare xmm_neg_zero, xmm_zero, LT_OS, false
+Compare xmm_nan, xmm_four, LT_OS, false
+Compare xmm_four, xmm_nan, LT_OS, false
+Compare xmm_nan, xmm_nan, LT_OS, false
+
+Compare xmm_four, xmm_neg_four, LE_OS, false
+Compare xmm_neg_four, xmm_four, LE_OS, true
+Compare xmm_zero, xmm_neg_zero, LE_OS, true
+Compare xmm_neg_zero, xmm_zero, LE_OS, true
+Compare xmm_nan, xmm_four, LE_OS, false
+Compare xmm_four, xmm_nan, LE_OS, false
+Compare xmm_nan, xmm_nan, LE_OS, false
+
+Compare xmm_four, xmm_neg_four, UNORD_Q, false
+Compare xmm_neg_four, xmm_four, UNORD_Q, false
+Compare xmm_zero, xmm_neg_zero, UNORD_Q, false
+Compare xmm_neg_zero, xmm_zero, UNORD_Q, false
+Compare xmm_nan, xmm_four, UNORD_Q, true
+Compare xmm_four, xmm_nan, UNORD_Q, true
+Compare xmm_nan, xmm_nan, UNORD_Q, true
+
+Compare xmm_four, xmm_four, NEQ_UQ, false
+Compare xmm_four, xmm_neg_four, NEQ_UQ, true
+Compare xmm_zero, xmm_neg_zero, NEQ_UQ, false
+Compare xmm_zero, xmm_zero, NEQ_UQ, false
+Compare xmm_four, xmm_nan, NEQ_UQ, true
+Compare xmm_nan, xmm_four, NEQ_UQ, true
+Compare xmm_nan, xmm_nan, NEQ_UQ, true
+
+Compare xmm_four, xmm_neg_four, NLT_US, true
+Compare xmm_neg_four, xmm_four, NLT_US, false
+Compare xmm_zero, xmm_neg_zero, NLT_US, true
+Compare xmm_neg_zero, xmm_zero, NLT_US, true
+Compare xmm_nan, xmm_four, NLT_US, true
+Compare xmm_four, xmm_nan, NLT_US, true
+Compare xmm_nan, xmm_nan, NLT_US, true
+
+Compare xmm_four, xmm_neg_four, NLE_US, true
+Compare xmm_neg_four, xmm_four, NLE_US, false
+Compare xmm_zero, xmm_neg_zero, NLE_US, false
+Compare xmm_neg_zero, xmm_zero, NLE_US, false
+Compare xmm_nan, xmm_four, NLE_US, true
+Compare xmm_four, xmm_nan, NLE_US, true
+Compare xmm_nan, xmm_nan, NLE_US, true
+
+Compare xmm_four, xmm_neg_four, ORD_Q, true
+Compare xmm_neg_four, xmm_four, ORD_Q, true
+Compare xmm_zero, xmm_neg_zero, ORD_Q, true
+Compare xmm_neg_zero, xmm_zero, ORD_Q, true
+Compare xmm_nan, xmm_four, ORD_Q, false
+Compare xmm_four, xmm_nan, ORD_Q, false
+Compare xmm_nan, xmm_nan, ORD_Q, false
+
+Compare xmm_four, xmm_neg_four, EQ_UQ, false
+Compare xmm_four, xmm_four, EQ_UQ, true
+Compare xmm_zero, xmm_neg_zero, EQ_UQ, true
+Compare xmm_zero, xmm_zero, EQ_UQ, true
+Compare xmm_four, xmm_nan, EQ_UQ, true
+Compare xmm_nan, xmm_four, EQ_UQ, true
+Compare xmm_nan, xmm_nan, EQ_UQ, true
+
+Compare xmm_four, xmm_neg_four, NGE_US, false
+Compare xmm_neg_four, xmm_four, NGE_US, true
+Compare xmm_zero, xmm_neg_zero, NGE_US, false
+Compare xmm_neg_zero, xmm_zero, NGE_US, false
+Compare xmm_nan, xmm_four, NGE_US, true
+Compare xmm_four, xmm_nan, NGE_US, true
+Compare xmm_nan, xmm_nan, NGE_US, true
+
+Compare xmm_four, xmm_neg_four, NGT_US, false
+Compare xmm_neg_four, xmm_four, NGT_US, true
+Compare xmm_zero, xmm_neg_zero, NGT_US, true
+Compare xmm_neg_zero, xmm_zero, NGT_US, true
+Compare xmm_nan, xmm_four, NGT_US, true
+Compare xmm_four, xmm_nan, NGT_US, true
+Compare xmm_nan, xmm_nan, NGT_US, true
+
+Compare xmm_four, xmm_neg_four, FALSE_OQ, false
+Compare xmm_neg_four, xmm_four, FALSE_OQ, false
+Compare xmm_zero, xmm_neg_zero, FALSE_OQ, false
+Compare xmm_neg_zero, xmm_zero, FALSE_OQ, false
+Compare xmm_nan, xmm_four, FALSE_OQ, false
+Compare xmm_four, xmm_nan, FALSE_OQ, false
+Compare xmm_nan, xmm_nan, FALSE_OQ, false
+
+Compare xmm_four, xmm_neg_four, NEQ_OQ, true
+Compare xmm_four, xmm_four, NEQ_OQ, false
+Compare xmm_zero, xmm_neg_zero, NEQ_OQ, false
+Compare xmm_zero, xmm_zero, NEQ_OQ, false
+Compare xmm_four, xmm_nan, NEQ_OQ, false
+Compare xmm_nan, xmm_four, NEQ_OQ, false
+Compare xmm_nan, xmm_nan, NEQ_OQ, false
+
+Compare xmm_four, xmm_neg_four, GE_OS, true
+Compare xmm_neg_four, xmm_four, GE_OS, false
+Compare xmm_zero, xmm_neg_zero, GE_OS, true
+Compare xmm_neg_zero, xmm_zero, GE_OS, true
+Compare xmm_nan, xmm_four, GE_OS, false
+Compare xmm_four, xmm_nan, GE_OS, false
+Compare xmm_nan, xmm_nan, GE_OS, false
+
+Compare xmm_four, xmm_neg_four, GT_OS, true
+Compare xmm_neg_four, xmm_four, GT_OS, false
+Compare xmm_zero, xmm_neg_zero, GT_OS, false
+Compare xmm_neg_zero, xmm_zero, GT_OS, false
+Compare xmm_nan, xmm_four, GT_OS, false
+Compare xmm_four, xmm_nan, GT_OS, false
+Compare xmm_nan, xmm_nan, GT_OS, false
+
+Compare xmm_four, xmm_neg_four, TRUE_UQ, true
+Compare xmm_neg_four, xmm_four, TRUE_UQ, true
+Compare xmm_zero, xmm_neg_zero, TRUE_UQ, true
+Compare xmm_neg_zero, xmm_zero, TRUE_UQ, true
+Compare xmm_nan, xmm_four, TRUE_UQ, true
+Compare xmm_four, xmm_nan, TRUE_UQ, true
+Compare xmm_nan, xmm_nan, TRUE_UQ, true
+
+Compare xmm_four, xmm_neg_four, EQ_OS, false
+Compare xmm_neg_four, xmm_four, EQ_OS, false
+Compare xmm_zero, xmm_neg_zero, EQ_OS, true
+Compare xmm_neg_zero, xmm_zero, EQ_OS, true
+Compare xmm_nan, xmm_four, EQ_OS, false
+Compare xmm_four, xmm_nan, EQ_OS, false
+Compare xmm_nan, xmm_nan, EQ_OS, false
+
+Compare xmm_four, xmm_neg_four, LT_OQ, false
+Compare xmm_neg_four, xmm_four, LT_OQ, true
+Compare xmm_zero, xmm_neg_zero, LT_OQ, false
+Compare xmm_neg_zero, xmm_zero, LT_OQ, false
+Compare xmm_nan, xmm_four, LT_OQ, false
+Compare xmm_four, xmm_nan, LT_OQ, false
+Compare xmm_nan, xmm_nan, LT_OQ, false
+
+Compare xmm_four, xmm_neg_four, LE_OQ, false
+Compare xmm_neg_four, xmm_four, LE_OQ, true
+Compare xmm_zero, xmm_neg_zero, LE_OQ, true
+Compare xmm_neg_zero, xmm_zero, LE_OQ, true
+Compare xmm_nan, xmm_four, LE_OQ, false
+Compare xmm_four, xmm_nan, LE_OQ, false
+Compare xmm_nan, xmm_nan, LE_OQ, false
+
+Compare xmm_four, xmm_neg_four, UNORD_S, false
+Compare xmm_neg_four, xmm_four, UNORD_S, false
+Compare xmm_zero, xmm_neg_zero, UNORD_S, false
+Compare xmm_neg_zero, xmm_zero, UNORD_S, false
+Compare xmm_nan, xmm_four, UNORD_S, true
+Compare xmm_four, xmm_nan, UNORD_S, true
+Compare xmm_nan, xmm_nan, UNORD_S, true
+
+Compare xmm_four, xmm_neg_four, NEQ_US, true
+Compare xmm_neg_four, xmm_four, NEQ_US, true
+Compare xmm_zero, xmm_neg_zero, NEQ_US, false
+Compare xmm_neg_zero, xmm_zero, NEQ_US, false
+Compare xmm_nan, xmm_four, NEQ_US, true
+Compare xmm_four, xmm_nan, NEQ_US, true
+Compare xmm_nan, xmm_nan, NEQ_US, true
+
+Compare xmm_four, xmm_neg_four, NLT_UQ, true
+Compare xmm_neg_four, xmm_four, NLT_UQ, false
+Compare xmm_zero, xmm_neg_zero, NLT_UQ, true
+Compare xmm_neg_zero, xmm_zero, NLT_UQ, true
+Compare xmm_nan, xmm_four, NLT_UQ, true
+Compare xmm_four, xmm_nan, NLT_UQ, true
+Compare xmm_nan, xmm_nan, NLT_UQ, true
+
+Compare xmm_four, xmm_neg_four, NLE_UQ, true
+Compare xmm_neg_four, xmm_four, NLE_UQ, false
+Compare xmm_zero, xmm_neg_zero, NLE_UQ, false
+Compare xmm_neg_zero, xmm_zero, NLE_UQ, false
+Compare xmm_nan, xmm_four, NLE_UQ, true
+Compare xmm_four, xmm_nan, NLE_UQ, true
+Compare xmm_nan, xmm_nan, NLE_UQ, true
+
+Compare xmm_four, xmm_neg_four, ORD_S, true
+Compare xmm_neg_four, xmm_four, ORD_S, true
+Compare xmm_zero, xmm_neg_zero, ORD_S, true
+Compare xmm_neg_zero, xmm_zero, ORD_S, true
+Compare xmm_nan, xmm_four, ORD_S, false
+Compare xmm_four, xmm_nan, ORD_S, false
+Compare xmm_nan, xmm_nan, ORD_S, false
+
+Compare xmm_four, xmm_neg_four, EQ_US, false
+Compare xmm_four, xmm_four, EQ_US, true
+Compare xmm_zero, xmm_neg_zero, EQ_US, true
+Compare xmm_zero, xmm_zero, EQ_US, true
+Compare xmm_four, xmm_nan, EQ_US, true
+Compare xmm_nan, xmm_four, EQ_US, true
+Compare xmm_nan, xmm_nan, EQ_US, true
+
+Compare xmm_four, xmm_neg_four, NGE_UQ, false
+Compare xmm_neg_four, xmm_four, NGE_UQ, true
+Compare xmm_zero, xmm_neg_zero, NGE_UQ, false
+Compare xmm_neg_zero, xmm_zero, NGE_UQ, false
+Compare xmm_nan, xmm_four, NGE_UQ, true
+Compare xmm_four, xmm_nan, NGE_UQ, true
+Compare xmm_nan, xmm_nan, NGE_UQ, true
+
+Compare xmm_four, xmm_neg_four, NGT_UQ, false
+Compare xmm_neg_four, xmm_four, NGT_UQ, true
+Compare xmm_zero, xmm_neg_zero, NGT_UQ, true
+Compare xmm_neg_zero, xmm_zero, NGT_UQ, true
+Compare xmm_nan, xmm_four, NGT_UQ, true
+Compare xmm_four, xmm_nan, NGT_UQ, true
+Compare xmm_nan, xmm_nan, NGT_UQ, true
+
+Compare xmm_four, xmm_neg_four, FALSE_OS, false
+Compare xmm_neg_four, xmm_four, FALSE_OS, false
+Compare xmm_zero, xmm_neg_zero, FALSE_OS, false
+Compare xmm_neg_zero, xmm_zero, FALSE_OS, false
+Compare xmm_nan, xmm_four, FALSE_OS, false
+Compare xmm_four, xmm_nan, FALSE_OS, false
+Compare xmm_nan, xmm_nan, FALSE_OS, false
+
+Compare xmm_four, xmm_neg_four, NEQ_OS, true
+Compare xmm_four, xmm_four, NEQ_OS, false
+Compare xmm_zero, xmm_neg_zero, NEQ_OS, false
+Compare xmm_zero, xmm_zero, NEQ_OS, false
+Compare xmm_four, xmm_nan, NEQ_OS, false
+Compare xmm_nan, xmm_four, NEQ_OS, false
+Compare xmm_nan, xmm_nan, NEQ_OS, false
+
+Compare xmm_four, xmm_neg_four, GE_OQ, true
+Compare xmm_neg_four, xmm_four, GE_OQ, false
+Compare xmm_zero, xmm_neg_zero, GE_OQ, true
+Compare xmm_neg_zero, xmm_zero, GE_OQ, true
+Compare xmm_nan, xmm_four, GE_OQ, false
+Compare xmm_four, xmm_nan, GE_OQ, false
+Compare xmm_nan, xmm_nan, GE_OQ, false
+
+Compare xmm_four, xmm_neg_four, GT_OQ, true
+Compare xmm_neg_four, xmm_four, GT_OQ, false
+Compare xmm_zero, xmm_neg_zero, GT_OQ, false
+Compare xmm_neg_zero, xmm_zero, GT_OQ, false
+Compare xmm_nan, xmm_four, GT_OQ, false
+Compare xmm_four, xmm_nan, GT_OQ, false
+Compare xmm_nan, xmm_nan, GT_OQ, false
+
+Compare xmm_four, xmm_neg_four, TRUE_US, true
+Compare xmm_neg_four, xmm_four, TRUE_US, true
+Compare xmm_zero, xmm_neg_zero, TRUE_US, true
+Compare xmm_neg_zero, xmm_zero, TRUE_US, true
+Compare xmm_nan, xmm_four, TRUE_US, true
+Compare xmm_four, xmm_nan, TRUE_US, true
+Compare xmm_nan, xmm_nan, TRUE_US, true
+
+; If ecx is not correct we can use the counter to determine which test failed
+.fail:
+hlt
+
+align 32
+.data:
+dd -4.0
+dd 0xCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dd -0.0
+dd 0xCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dd 0.0
+dd 0xCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+dd 4.0
+dd 0xCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+.nan:
+dq 0xCCCCCCCC7FC00000
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+dq 0xCCCCCCCCCCCCCCCC
+
+.false:
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000
+
+.true:
+dq 0x00000000FFFFFFFF
+dq 0x0000000000000000
+dq 0x0000000000000000
+dq 0x0000000000000000