[llvm] r283476 - AMDGPU: Don't fold undef uses or copies with implicit uses

[Matt Arsenault via llvm-commits]

Author: arsenm
Date: Thu Oct  6 13:12:13 2016
New Revision: 283476

URL: http://llvm.org/viewvc/llvm-project?rev=283476&view=rev
Log:
AMDGPU: Don't fold undef uses or copies with implicit uses

Modified:
    llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp

Modified: llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp?rev=283476&r1=283475&r2=283476&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp Thu Oct  6 13:12:13 2016
@@ -92,11 +92,18 @@ FunctionPass *llvm::createSIFoldOperands
   return new SIFoldOperands();
 }
 
-static bool isSafeToFold(unsigned Opcode) {
-  switch(Opcode) {
+static bool isSafeToFold(const MachineInstr &MI) {
+  switch (MI.getOpcode()) {
   case AMDGPU::V_MOV_B32_e32:
   case AMDGPU::V_MOV_B32_e64:
-  case AMDGPU::V_MOV_B64_PSEUDO:
+  case AMDGPU::V_MOV_B64_PSEUDO: {
+    // If there are additional implicit register operands, this may be used for
+    // register indexing so the source register operand isn't simply copied.
+    unsigned NumOps = MI.getDesc().getNumOperands() +
+      MI.getDesc().getNumImplicitUses();
+
+    return MI.getNumOperands() == NumOps;
+  }
   case AMDGPU::S_MOV_B32:
   case AMDGPU::S_MOV_B64:
   case AMDGPU::COPY:
@@ -203,6 +210,14 @@ static bool tryAddToFoldList(std::vector
   return true;
 }
 
+// If the use operand doesn't care about the value, this may be an operand only
+// used for register indexing, in which case it is unsafe to fold.
+static bool isUseSafeToFold(const MachineInstr &MI,
+                            const MachineOperand &UseMO) {
+  return !UseMO.isUndef();
+  //return !MI.hasRegisterImplicitUseOperand(UseMO.getReg());
+}
+
 static void foldOperand(MachineOperand &OpToFold, MachineInstr *UseMI,
                         unsigned UseOpIdx,
                         std::vector<FoldCandidate> &FoldList,
@@ -211,6 +226,9 @@ static void foldOperand(MachineOperand &
                         MachineRegisterInfo &MRI) {
   const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx);
 
+  if (!isUseSafeToFold(*UseMI, UseOp))
+    return;
+
   // FIXME: Fold operands with subregs.
   if (UseOp.isReg() && OpToFold.isReg()) {
     if (UseOp.isImplicit() || UseOp.getSubReg() != AMDGPU::NoSubRegister)
@@ -477,7 +495,7 @@ bool SIFoldOperands::runOnMachineFunctio
       Next = std::next(I);
       MachineInstr &MI = *I;
 
-      if (!isSafeToFold(MI.getOpcode()))
+      if (!isSafeToFold(MI))
         continue;
 
       unsigned OpSize = TII->getOpSize(MI, 1);