Merge pull request #3144 from Sonicadvance1/optimize_redundant_store_load

RCLSE: Optimize redundant store->load operations

FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp

@@ -2544,12 +2544,18 @@ DEF_OP(VUShrSWide) {
   else if (HostSupportsSVE128) {
     const auto Mask = PRED_TMP_16B.Merging();
 
-    auto ShiftRegister = ShiftScalar.Z();
+    auto ShiftRegister = ShiftScalar;
     if (OpSize > 8) {
       // SVE wide shifts don't need to duplicate the low bits unless the OpSize is 16-bytes
       // Slightly more optimal for 8-byte opsize.
       dup(ARMEmitter::SubRegSize::i64Bit, VTMP1.Z(), ShiftScalar.Z(), 0);
-      ShiftRegister = VTMP1.Z();
+      ShiftRegister = VTMP1;
+    }
+
+    if (Dst == ShiftRegister) {
+      // If destination aliases the shift vector then we need to move it temporarily.
+      mov(VTMP1.Z(), ShiftRegister.Z());
+      ShiftRegister = VTMP1;
     }
 
     if (Dst != Vector) {
@@ -2557,10 +2563,10 @@ DEF_OP(VUShrSWide) {
       movprfx(Dst.Z(), Vector.Z());
     }
     if (ElementSize == 8) {
-      lsr(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      lsr(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
     else {
-      lsr_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      lsr_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
   } else {
     // uqshl + ushr of 57-bits leaves 7-bits remaining.
@@ -2611,12 +2617,18 @@ DEF_OP(VSShrSWide) {
   else if (HostSupportsSVE128) {
     const auto Mask = PRED_TMP_16B.Merging();
 
-    auto ShiftRegister = ShiftScalar.Z();
+    auto ShiftRegister = ShiftScalar;
     if (OpSize > 8) {
       // SVE wide shifts don't need to duplicate the low bits unless the OpSize is 16-bytes
       // Slightly more optimal for 8-byte opsize.
       dup(ARMEmitter::SubRegSize::i64Bit, VTMP1.Z(), ShiftScalar.Z(), 0);
-      ShiftRegister = VTMP1.Z();
+      ShiftRegister = VTMP1;
+    }
+
+    if (Dst == ShiftRegister) {
+      // If destination aliases the shift vector then we need to move it temporarily.
+      mov(VTMP1.Z(), ShiftRegister.Z());
+      ShiftRegister = VTMP1;
     }
 
     if (Dst != Vector) {
@@ -2624,10 +2636,10 @@ DEF_OP(VSShrSWide) {
       movprfx(Dst.Z(), Vector.Z());
     }
     if (ElementSize == 8) {
-      asr(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      asr(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
     else {
-      asr_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      asr_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
   } else {
     // uqshl + ushr of 57-bits leaves 7-bits remaining.
@@ -2678,12 +2690,18 @@ DEF_OP(VUShlSWide) {
   else if (HostSupportsSVE128) {
     const auto Mask = PRED_TMP_16B.Merging();
 
-    auto ShiftRegister = ShiftScalar.Z();
+    auto ShiftRegister = ShiftScalar;
     if (OpSize > 8) {
       // SVE wide shifts don't need to duplicate the low bits unless the OpSize is 16-bytes
       // Slightly more optimal for 8-byte opsize.
       dup(ARMEmitter::SubRegSize::i64Bit, VTMP1.Z(), ShiftScalar.Z(), 0);
-      ShiftRegister = VTMP1.Z();
+      ShiftRegister = VTMP1;
+    }
+
+    if (Dst == ShiftRegister) {
+      // If destination aliases the shift vector then we need to move it temporarily.
+      mov(VTMP1.Z(), ShiftRegister.Z());
+      ShiftRegister = VTMP1;
     }
 
     if (Dst != Vector) {
@@ -2691,10 +2709,10 @@ DEF_OP(VUShlSWide) {
       movprfx(Dst.Z(), Vector.Z());
     }
     if (ElementSize == 8) {
-      lsl(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      lsl(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
     else {
-      lsl_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister);
+      lsl_wide(SubRegSize, Dst.Z(), Mask, Dst.Z(), ShiftRegister.Z());
     }
   } else {
     // uqshl + ushr of 57-bits leaves 7-bits remaining.

FEXCore/Source/Interface/IR/Passes/DeadContextStoreElimination.cpp

@@ -506,17 +506,17 @@ bool RCLSE::ClassifyContextLoad(FEXCore::IR::IREmitter *IREmit, ContextInfo *Loc
   ContextMemberInfo PreviousMemberInfoCopy = *Info;
   RecordAccess(Info, Class, Offset, Size, LastAccessType::READ, CodeNode);
 
-  if (IsReadAccess(PreviousMemberInfoCopy.Accessed) &&
-      IsReadAccess(Info->Accessed) &&
-      PreviousMemberInfoCopy.AccessRegClass == Info->AccessRegClass &&
+  if (PreviousMemberInfoCopy.AccessRegClass == Info->AccessRegClass &&
       PreviousMemberInfoCopy.AccessOffset == Info->AccessOffset &&
       PreviousMemberInfoCopy.AccessSize == Size) {
-    // Optimize the case of redundant reads of the same exact value.
+    // This optimizes two cases:
+    // - Previous access was a load, and we have a redundant load of the same value.
+    // - Previous access was a store, and we are redundantly loading immediately after the store. Eliminating the store.
     IREmit->ReplaceAllUsesWithRange(CodeNode, PreviousMemberInfoCopy.ValueNode, IREmit->GetIterator(IREmit->WrapNode(CodeNode)), BlockEnd);
     RecordAccess(Info, Class, Offset, Size, LastAccessType::READ, PreviousMemberInfoCopy.ValueNode);
     return true;
   }
-  // TODO: Optimize the case of Store->Load.
+  // TODO: Optimize the case of partial loads.
   return false;
 }
 

unittests/InstructionCountCI/x87.json

@@ -9039,7 +9039,7 @@
       ]
     },
     "fucompp": {
-      "ExpectedInstructionCount": 77,
+      "ExpectedInstructionCount": 76,
       "Optimal": "No",
       "Comment": [
         "0xda 11b 0xe9 /5"
@@ -9115,7 +9115,6 @@
         "strb w22, [x28, #1010]",
         "add w20, w20, #0x1 (1)",
         "and w20, w20, #0x7",
-        "ldrb w22, [x28, #1010]",
         "lsl w21, w21, w20",
         "bic w21, w22, w21",
         "strb w21, [x28, #1010]",
@@ -19655,7 +19654,7 @@
       ]
     },
     "fcompp": {
-      "ExpectedInstructionCount": 77,
+      "ExpectedInstructionCount": 76,
       "Optimal": "No",
       "Comment": [
         "0xde 11b 0xd9 /3"
@@ -19731,7 +19730,6 @@
         "strb w22, [x28, #1010]",
         "add w20, w20, #0x1 (1)",
         "and w20, w20, #0x7",
-        "ldrb w22, [x28, #1010]",
         "lsl w21, w21, w20",
         "bic w21, w22, w21",
         "strb w21, [x28, #1010]",