OpcodeDispatcher: avoid xblock liveness with ROR/ROL

Signed-off-by: Alyssa Rosenzweig <[email protected]>
FEX-Emu · Mar 29, 2024 · 7da1066 · 7da1066
1 parent 5b94d81
commit 7da1066
Show file tree

Hide file tree

Showing 3 changed files with 31 additions and 25 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -1928,7 +1928,7 @@ void OpDispatchBuilder::ROROp(OpcodeArgs) {
   if constexpr (Is1Bit) {
     GenerateFlags_RotateRightImmediate(Op, ALUOp, Dest, 1);
   } else {
-    GenerateFlags_RotateRight(Op, ALUOp, Dest, Src);
+    GenerateFlags_RotateRight(Op, Src);
   }
 }
 
@@ -2005,7 +2005,7 @@ void OpDispatchBuilder::ROLOp(OpcodeArgs) {
   if constexpr (Is1Bit) {
     GenerateFlags_RotateLeftImmediate(Op, ALUOp, Dest, 1);
   } else {
-    GenerateFlags_RotateLeft(Op, ALUOp, Dest, Src);
+    GenerateFlags_RotateLeft(Op, Src);
   }
 }
 

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -1669,7 +1669,7 @@ friend class FEXCore::IR::PassManager;
         OrderedNode *Src1;
       } OneSource;
 
-      // Logical, LSHL, LSHR, ASHR, ROR, ROL
+      // Logical, LSHL, LSHR, ASHR
       struct {
         OrderedNode *Src1;
         OrderedNode *Src2;
@@ -1688,6 +1688,12 @@ friend class FEXCore::IR::PassManager;
 
         bool UpdateCF;
       } TwoSrcImmediate;
+
+      // ROL, ROR
+      struct {
+        X86Tables::DecodedOp Op;
+        OrderedNode *Src2;
+      } Decoded;
     } Sources{};
   };
 
@@ -1785,8 +1791,8 @@ friend class FEXCore::IR::PassManager;
   void CalculateFlags_ShiftRightImmediateCommon(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift);
   void CalculateFlags_SignShiftRight(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2);
   void CalculateFlags_SignShiftRightImmediate(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift);
-  void CalculateFlags_RotateRight(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2);
-  void CalculateFlags_RotateLeft(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2);
+  void CalculateFlags_RotateRight(uint8_t SrcSize, X86Tables::DecodedOp Op, OrderedNode *Src2);
+  void CalculateFlags_RotateLeft(uint8_t SrcSize, X86Tables::DecodedOp Op, OrderedNode *Src2);
   void CalculateFlags_RotateRightImmediate(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift);
   void CalculateFlags_RotateLeftImmediate(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, uint64_t Shift);
   void CalculateFlags_BEXTR(OrderedNode *Src);
@@ -1976,34 +1982,32 @@ friend class FEXCore::IR::PassManager;
     };
   }
 
-  void GenerateFlags_RotateRight(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  void GenerateFlags_RotateRight(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Src2) {
     // Doesn't set all the flags, needs to calculate.
     CalculateDeferredFlags();
 
     CurrentDeferredFlags = DeferredFlagData {
       .Type = FlagsGenerationType::TYPE_ROR,
       .SrcSize = GetSrcSize(Op),
-      .Res = Res,
       .Sources = {
-        .TwoSource = {
-          .Src1 = Src1,
+        .Decoded = {
+          .Op = Op,
           .Src2 = Src2,
         },
       },
     };
   }
 
-  void GenerateFlags_RotateLeft(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
+  void GenerateFlags_RotateLeft(FEXCore::X86Tables::DecodedOp Op, OrderedNode *Src2) {
     // Doesn't set all the flags, needs to calculate.
     CalculateDeferredFlags();
 
     CurrentDeferredFlags = DeferredFlagData {
       .Type = FlagsGenerationType::TYPE_ROL,
       .SrcSize = GetSrcSize(Op),
-      .Res = Res,
       .Sources = {
-        .TwoSource = {
-          .Src1 = Src1,
+        .Decoded = {
+          .Op = Op,
           .Src2 = Src2,
         },
       },

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp
@@ -404,9 +404,8 @@ void OpDispatchBuilder::CalculateDeferredFlags(uint32_t FlagsToCalculateMask) {
     case FlagsGenerationType::TYPE_ROR:
       CalculateFlags_RotateRight(
         CurrentDeferredFlags.SrcSize,
-        CurrentDeferredFlags.Res,
-        CurrentDeferredFlags.Sources.TwoSource.Src1,
-        CurrentDeferredFlags.Sources.TwoSource.Src2);
+        CurrentDeferredFlags.Sources.Decoded.Op,
+        CurrentDeferredFlags.Sources.Decoded.Src2);
       break;
     case FlagsGenerationType::TYPE_RORI:
       CalculateFlags_RotateRightImmediate(
@@ -418,9 +417,8 @@ void OpDispatchBuilder::CalculateDeferredFlags(uint32_t FlagsToCalculateMask) {
     case FlagsGenerationType::TYPE_ROL:
       CalculateFlags_RotateLeft(
         CurrentDeferredFlags.SrcSize,
-        CurrentDeferredFlags.Res,
-        CurrentDeferredFlags.Sources.TwoSource.Src1,
-        CurrentDeferredFlags.Sources.TwoSource.Src2);
+        CurrentDeferredFlags.Sources.Decoded.Op,
+        CurrentDeferredFlags.Sources.Decoded.Src2);
       break;
     case FlagsGenerationType::TYPE_ROLI:
       CalculateFlags_RotateLeftImmediate(
@@ -835,11 +833,14 @@ void OpDispatchBuilder::CalculateFlags_ShiftRightDoubleImmediate(uint8_t SrcSize
   }
 }
 
-void OpDispatchBuilder::CalculateFlags_RotateRight(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
-  CalculateFlags_ShiftVariable(Src2, [this, SrcSize, Res](){
+void OpDispatchBuilder::CalculateFlags_RotateRight(uint8_t SrcSize, X86Tables::DecodedOp Op, OrderedNode *Src2) {
+  CalculateFlags_ShiftVariable(Src2, [this, SrcSize, Op](){
     auto SizeBits = SrcSize * 8;
     const auto OpSize = SrcSize == 8 ? OpSize::i64Bit : OpSize::i32Bit;
 
+    // Rematerialize inside block. Only lower SrcSize bits are used.
+    OrderedNode *Res = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, {.AllowUpperGarbage = true});
+
     // Ends up faster overall if we don't have FlagM, slower if we do...
     // If Shift != 1, OF is undefined so we choose to zero here.
     if (!CTX->HostFeatures.SupportsFlagM)
@@ -855,19 +856,20 @@ void OpDispatchBuilder::CalculateFlags_RotateRight(uint8_t SrcSize, OrderedNode
   });
 }
 
-void OpDispatchBuilder::CalculateFlags_RotateLeft(uint8_t SrcSize, OrderedNode *Res, OrderedNode *Src1, OrderedNode *Src2) {
-  CalculateFlags_ShiftVariable(Src2, [this, SrcSize, Res](){
+void OpDispatchBuilder::CalculateFlags_RotateLeft(uint8_t SrcSize, X86Tables::DecodedOp Op, OrderedNode *Src2) {
+  CalculateFlags_ShiftVariable(Src2, [this, SrcSize, Op](){
     const auto OpSize = SrcSize == 8 ? OpSize::i64Bit : OpSize::i32Bit;
     auto SizeBits = SrcSize * 8;
 
+    // Rematerialize inside block. Only lower SrcSize bits are used.
+    OrderedNode *Res = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, {.AllowUpperGarbage = true});
+
     // Ends up faster overall if we don't have FlagM, slower if we do...
     // If Shift != 1, OF is undefined so we choose to zero here.
     if (!CTX->HostFeatures.SupportsFlagM)
       ZeroCV();
 
     // Extract the last bit shifted in to CF
-    //auto Size = _Constant(GetSrcSize(Res) * 8);
-    //auto ShiftAmt = _Sub(OpSize::i64Bit, Size, Src2);
     SetRFLAG<FEXCore::X86State::RFLAG_CF_RAW_LOC>(Res, 0, true);
 
     // OF is the LSB and MSB XOR'd together.