AMDGPU/SI: Teach SIInstrInfo::FoldImmediate() to fold immediates into copies

[• tstellarAMD]

Summary

I put this code here, because I want to re-use it in a few other places.
This supersedes some of the immediate folding code we have in SIFoldOperands.
I think the peephole optimizers is probably a better place for folding
immediates into copies, since it does some register coalescing in the same time.

This will also make it easier to transition SIFoldOperands into a smarter pass,
where it looks at all uses of instruction at once to determine the optimal way to
fold operands. Right now, the pass just considers one operand at a time.

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

1225–1226

I don't think this supports modifiers. I don't think there's any reason to ever use the v_mov_b32 as vop3 so it can probably just be skipped

[arsenm]

lib/Target/AMDGPU/SIInstrFormats.td

296–297

This isn't necessary here since they aren't used for any SOP instructions here

[• tstellarAMD]

Rebase and drop code for folding v_mov_b32_e64.

[arsenm]

LGTM

lib/Target/AMDGPU/SIInstrFormats.td

 class SOP1 <dag outs, dag ins, string asm, list<dag> pattern> :
     InstSI<outs, ins, asm, pattern> {
   let mayLoad = 0;
   let mayStore = 0;
   let hasSideEffects = 0;
   let isCodeGenOnly = 0;
   let SALU = 1;
   let SOP1 = 1;
+
+  let UseNamedOperandTable = 1;
 }
 
 class SOP2 <dag outs, dag ins, string asm, list<dag> pattern> :
     InstSI <outs, ins, asm, pattern> {
 
   let mayLoad = 0;
   let mayStore = 0;
   let hasSideEffects = 0;

lib/Target/AMDGPU/SIInstrInfo.cpp

static void removeModOperands(MachineInstr &MI) {
   int Src2ModIdx = AMDGPU::getNamedOperandIdx(Opc,
                                               AMDGPU::OpName::src2_modifiers);
 
   MI.RemoveOperand(Src2ModIdx);
   MI.RemoveOperand(Src1ModIdx);
   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))
     return false;
 
   unsigned Opc = UseMI.getOpcode();
+  if (Opc == AMDGPU::COPY) {
+    bool isVGPRCopy = RI.isVGPR(*MRI, UseMI.getOperand(0).getReg());
+    switch (DefMI.getOpcode()) {
+    default:
+      return false;
+    case AMDGPU::V_MOV_B32_e64:
+      if (hasModifiersSet(DefMI, AMDGPU::OpName::src0_modifiers))
+        return false;
+    case AMDGPU::S_MOV_B64:
+      // TODO: We could fold 64-bit immediates, but this get compilicated
+      // when there are sub-registers.
+      return false;
+
+    case AMDGPU::V_MOV_B32_e32:
+    case AMDGPU::S_MOV_B32:
+      break;
+    }
+    unsigned NewOpc = isVGPRCopy ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32;
+    const MachineOperand *ImmOp = getNamedOperand(DefMI, AMDGPU::OpName::src0);
+    assert(ImmOp);
+    // FIXME: We could handle FrameIndex values here.
+    if (!ImmOp->isImm()) {
+      return false;
+    }
+    UseMI.setDesc(get(NewOpc));
+    UseMI.getOperand(1).ChangeToImmediate(ImmOp->getImm());
+    UseMI.addImplicitDefUseOperands(*UseMI.getParent()->getParent());
+    return true;
+  }
+
   if (Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64) {
     // Don't fold if we are using source modifiers. The new VOP2 instructions
     // don't have them.
     if (hasModifiersSet(UseMI, AMDGPU::OpName::src0_modifiers) ||
         hasModifiersSet(UseMI, AMDGPU::OpName::src1_modifiers) ||
         hasModifiersSet(UseMI, AMDGPU::OpName::src2_modifiers)) {
       return false;
     }