[PATCH] D23402: AMDGPU/SI: Teach SIInstrInfo::FoldImmediate() to fold immediates into copies

[Tom Stellard via llvm-commits]

tstellarAMD updated this revision to Diff 70418.
tstellarAMD added a comment.
Herald added subscribers: nhaehnle, wdng.

Rebase and drop code for folding v_mov_b32_e64.


https://reviews.llvm.org/D23402

Files:
  lib/Target/AMDGPU/SIInstrInfo.cpp
  lib/Target/AMDGPU/SOPInstructions.td

Index: lib/Target/AMDGPU/SOPInstructions.td
===================================================================
--- lib/Target/AMDGPU/SOPInstructions.td
+++ lib/Target/AMDGPU/SOPInstructions.td
@@ -25,6 +25,7 @@
   let SALU = 1;
   let SOP1 = 1;
   let SchedRW = [WriteSALU];
+  let UseNamedOperandTable = 1;
 
   string Mnemonic = opName;
   string AsmOperands = asmOps;
@@ -1100,4 +1101,4 @@
 def S_SETREG_B32_vi        : SOPK_Real_vi <0x12, S_SETREG_B32>;
 //def S_GETREG_REGRD_B32_vi  : SOPK_Real_vi <0x13, S_GETREG_REGRD_B32>; // see pseudo for comments
 def S_SETREG_IMM32_B32_vi  : SOPK_Real64<0x14, S_SETREG_IMM32_B32>,
-                             Select_vi<S_SETREG_IMM32_B32.Mnemonic>;
\ No newline at end of file
+                             Select_vi<S_SETREG_IMM32_B32.Mnemonic>;
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1184,14 +1184,39 @@
   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::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.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D23402.70418.patch
Type: text/x-patch
Size: 2457 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160906/1f825504/attachment.bin>