2 files changed, 25 insertions, 1 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 08254baa61d..c79db482e81 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -202,6 +202,7 @@ public:
   GCNPassConfig(TargetMachine *TM, PassManagerBase &PM)
     : AMDGPUPassConfig(TM, PM) { }
   bool addPreISel() override;
+  void addMachineSSAOptimization() override;
   bool addInstSelector() override;
 #ifdef LLVM_BUILD_GLOBAL_ISEL
   bool addIRTranslator() override;
@@ -323,11 +324,24 @@ bool GCNPassConfig::addPreISel() {
   return false;
 }
 
+void GCNPassConfig::addMachineSSAOptimization() {
+  TargetPassConfig::addMachineSSAOptimization();
+
+  // We want to fold operands after PeepholeOptimizer has run (or as part of
+  // it), because it will eliminate extra copies making it easier to fold the
+  // real source operand. We want to eliminate dead instructions after, so that
+  // we see fewer uses of the copies. We then need to clean up the dead
+  // instructions leftover after the operands are folded as well.
+  //
+  // XXX - Can we get away without running DeadMachineInstructionElim again?
+  addPass(&SIFoldOperandsID);
+  addPass(&DeadMachineInstructionElimID);
+}
+
 bool GCNPassConfig::addInstSelector() {
   AMDGPUPassConfig::addInstSelector();
   addPass(createSILowerI1CopiesPass());
   addPass(&SIFixSGPRCopiesID);
-  addPass(createSIFoldOperandsPass());
   return false;
 }
 
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 01a805439a0..ba5b8a229b7 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1058,6 +1058,8 @@ static void removeModOperands(MachineInstr &MI) {
   MI.RemoveOperand(Src0ModIdx);
 }
 
+// TODO: Maybe this should be removed this and custom fold everything in
+// SIFoldOperands?
 bool SIInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
                                 unsigned Reg, MachineRegisterInfo *MRI) const {
   if (!MRI->hasOneNonDBGUse(Reg))
@@ -1073,6 +1075,14 @@ bool SIInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
       return false;
     }
 
+    const MachineOperand &ImmOp = DefMI->getOperand(1);
+
+    // If this is a free constant, there's no reason to do this.
+    // TODO: We could fold this here instead of letting SIFoldOperands do it
+    // later.
+    if (isInlineConstant(ImmOp, 4))
+      return false;
+
     MachineOperand *Src0 = getNamedOperand(*UseMI, AMDGPU::OpName::src0);
     MachineOperand *Src1 = getNamedOperand(*UseMI, AMDGPU::OpName::src1);
     MachineOperand *Src2 = getNamedOperand(*UseMI, AMDGPU::OpName::src2);