AMDGPU/SI: Don't move copies of immediates to the VALU

[• tstellarAMD]

Summary

If we write an immediate to a VGPR and then copy the VGPR to an
SGPR, we can replace the copy with a S_MOV_B32 sgpr, imm, rather than
moving the copy to the SALU.

[arsenm]

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

319–320

I think this should be unreachable since we never emit this. I'm not sure the modifiers even work, so I don't think there's any reason to ever use this

366–367

This will leave behind the implicit exec use

[• tstellarAMD]

Address review comments.

[• tstellarAMD]

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

366–367

MI is a copy here, so there are no implicit exec uses.

[arsenm]

LGTM

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

static bool hasTerminatorThatModifiesExec(const MachineBasicBlock &MBB,
   for (MachineBasicBlock::const_iterator I = MBB.getFirstTerminator(),
        E = MBB.end(); I != E; ++I) {
     if (I->modifiesRegister(AMDGPU::EXEC, &TRI))
       return true;
   }
   return false;
 }
 
+static bool isSafeToFoldImmIntoCopy(const MachineInstr *Copy,
+                                    const MachineInstr *MoveImm,
+                                    const SIInstrInfo *TII,
+                                    unsigned &SMovOp,
+                                    int64_t &Imm) {
+
+  if (!MoveImm->isMoveImmediate())
+    return false;
+
+  const MachineOperand *ImmOp =
+      TII->getNamedOperand(*MoveImm, AMDGPU::OpName::src0);
+  if (!ImmOp->isImm())
+    return false;
+
+  // FIXME: Handle copies with sub-regs.
+  if (Copy->getOperand(0).getSubReg())
+    return false;
+
+  switch (MoveImm->getOpcode()) {
+  default:
+    return false;
+  case AMDGPU::V_MOV_B32_e32:
+    SMovOp = AMDGPU::S_MOV_B32;
+    break;
+  case AMDGPU::V_MOV_B64_PSEUDO:
+    SMovOp = AMDGPU::S_MOV_B64;
+    break;
+  }
+  Imm = ImmOp->getImm();
+  return true;
+}
+
 bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
   MachineRegisterInfo &MRI = MF.getRegInfo();
   const SIRegisterInfo *TRI = ST.getRegisterInfo();
   const SIInstrInfo *TII = ST.getInstrInfo();
   MDT = &getAnalysis<MachineDominatorTree>();
 
   SmallVector<MachineInstr *, 16> Worklist;
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
         // If the destination register is a physical register there isn't really
         // much we can do to fix this.
         if (!TargetRegisterInfo::isVirtualRegister(MI.getOperand(0).getReg()))
           continue;
 
         const TargetRegisterClass *SrcRC, *DstRC;
         std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, MRI);
         if (isVGPRToSGPRCopy(SrcRC, DstRC, *TRI)) {
-          DEBUG(dbgs() << "Fixing VGPR -> SGPR copy: " << MI);
+          MachineInstr *DefMI = MRI.getVRegDef(MI.getOperand(1).getReg());
+          unsigned SMovOp;
+          int64_t Imm;
+          // If we are just copying an immediate, we can replace the copy with
+          // s_mov_b32.
+          if (isSafeToFoldImmIntoCopy(&MI, DefMI, TII, SMovOp, Imm)) {
+            MI.getOperand(1).ChangeToImmediate(Imm);
+            MI.addImplicitDefUseOperands(MF);
+            MI.setDesc(TII->get(SMovOp));
+            break;
+          }
           TII->moveToVALU(MI);
         }
 
         break;
       }
       case AMDGPU::PHI: {
         unsigned Reg = MI.getOperand(0).getReg();
         if (!TRI->isSGPRClass(MRI.getRegClass(Reg)))

test/CodeGen/AMDGPU/salu-to-valu.ll

bb2:
   %tmp3 = add nsw i32 %tmp, 1
   br label %bb4
 
 bb4:
   %tmp5 = phi i32 [ %tmp3, %bb2 ], [ %tmp, %bb1 ]
   br label %bb1
 }
 
+; GCN-LABEL: {{^}}phi_imm_in_sgprs
+; GCN: s_movk_i32 [[A:s[0-9]+]], 0x400
+; GCN: s_movk_i32 [[B:s[0-9]+]], 0x400
+; GCN: [[LOOP_LABEL:[0-9a-zA-Z_]+]]:
+; GCN: s_xor_b32 [[B]], [[B]], [[A]]
+; GCN: s_cbranch_scc{{[01]}} [[LOOP_LABEL]]
+define void @phi_imm_in_sgprs(i32 addrspace(3)* %out, i32 %cond) {
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [0, %entry], [%i.add, %loop]
+  %offset = phi i32 [1024, %entry], [%offset.xor, %loop]
+  %offset.xor = xor i32 %offset, 1024
+  %offset.i = add i32 %offset.xor, %i
+  %ptr = getelementptr i32, i32 addrspace(3)* %out, i32 %offset.i
+  store i32 0, i32 addrspace(3)* %ptr
+  %i.add = add i32 %i, 1
+  %cmp = icmp ult i32 %i.add, %cond
+  br i1 %cmp, label %loop, label %exit
+
+exit:
+  ret void
+}
+
 attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind }