[AMDGPU] Preliminary patch for divergence driven instruction selection. Fold immediate SMRD offset.

[alex-t]

Summary

This patch handles specific processing of the buffer read instruction. It allows to fold immediate offset in case the COPY still retains between the move and read.

[rampitec]

Please attach full context diff.

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

4094 ↗

(On Diff #163760)

This isn't needed, copy only supports registers

[alex-t]

Unnecessary "isReg()" check removed.
Full context unified diff.

[rampitec]

LGTM

llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp

case AMDGPU::S_BUFFER_LOAD_DWORDX16_SGPR: {
         };
 
         // Find a literal offset in one of source operands.
         for (int i = 0; i < 2; i++) {
           const MachineOperand *Src =
             getNamedOperand(*Add, SrcNames[i]);
 
           if (Src->isReg()) {
-            auto Mov = MRI.getUniqueVRegDef(Src->getReg());
-            if (Mov && Mov->getOpcode() == AMDGPU::S_MOV_B32)
+            MachineInstr *Def = MRI.getUniqueVRegDef(Src->getReg());
+            if (Def) {
+              if (Def->isMoveImmediate())
+                Src = &Def->getOperand(1);
+              else if (Def->isCopy()) {
+                auto Mov = MRI.getUniqueVRegDef(Def->getOperand(1).getReg());
+                if (Mov && Mov->isMoveImmediate()) {
                   Src = &Mov->getOperand(1);
                 }
+              }
+            }
+          }
 
           if (Src) {
             if (Src->isImm())
               Offset = Src->getImm();
             else if (Src->isCImm())
               Offset = Src->getCImm()->getZExtValue();
           }
 

llvm/trunk/test/CodeGen/AMDGPU/smrd-fold-offset.mir

+# RUN: llc -march=amdgcn -run-pass si-fix-sgpr-copies -o - %s | FileCheck -check-prefix=GCN %s
+
+# GCN: BUFFER_LOAD_DWORD_OFFEN %{{[0-9]+}}, killed %{{[0-9]+}}, 0, 4095
+---
+name:            smrd_vgpr_offset_imm
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0
+
+    %4:vgpr_32 = COPY $vgpr0
+    %3:sgpr_32 = COPY $sgpr3
+    %2:sgpr_32 = COPY $sgpr2
+    %1:sgpr_32 = COPY $sgpr1
+    %0:sgpr_32 = COPY $sgpr0
+    %5:sgpr_128 = REG_SEQUENCE %0, %subreg.sub0, %1, %subreg.sub1, %2, %subreg.sub2, %3, %subreg.sub3
+    %6:sreg_32_xm0 = S_MOV_B32 4095
+    %8:vgpr_32 = COPY %6
+    %7:vgpr_32 = V_ADD_I32_e32 %4, killed %8, implicit-def dead $vcc, implicit $exec
+    %10:sreg_32 = COPY %7
+    %9:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR killed %5, killed %10, 0
+    $vgpr0 = COPY %9
+    SI_RETURN_TO_EPILOG $vgpr0
+...