[AMDGPU] Do not handle _SGPR SMEM instructions in SILoadStoreOptimizer

After D147334 we never select _SGPR forms of SMEM instructions on
subtargets that also support the _SGPR_IMM form, so there is no need to
handle them here.

Differential Revision: https://reviews.llvm.org/D149139

diff --git a/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp b/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
index 5079160..a0f782c 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
+++ b/llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
@@ -331,7 +331,6 @@ static unsigned getOpcodeWidth(const MachineInstr &MI, const SIInstrInfo &TII) {
 
   switch (Opc) {
   case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
-  case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR_IMM:
   case AMDGPU::S_LOAD_DWORD_IMM:
   case AMDGPU::GLOBAL_LOAD_DWORD:
@@ -342,7 +341,6 @@ static unsigned getOpcodeWidth(const MachineInstr &MI, const SIInstrInfo &TII) {
   case AMDGPU::FLAT_STORE_DWORD:
     return 1;
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM:
   case AMDGPU::S_LOAD_DWORDX2_IMM:
   case AMDGPU::GLOBAL_LOAD_DWORDX2:
@@ -360,7 +358,6 @@ static unsigned getOpcodeWidth(const MachineInstr &MI, const SIInstrInfo &TII) {
   case AMDGPU::FLAT_STORE_DWORDX3:
     return 3;
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM:
   case AMDGPU::S_LOAD_DWORDX4_IMM:
   case AMDGPU::GLOBAL_LOAD_DWORDX4:
@@ -371,7 +368,6 @@ static unsigned getOpcodeWidth(const MachineInstr &MI, const SIInstrInfo &TII) {
   case AMDGPU::FLAT_STORE_DWORDX4:
     return 4;
   case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR_IMM:
   case AMDGPU::S_LOAD_DWORDX8_IMM:
     return 8;
@@ -450,12 +446,6 @@ static InstClassEnum getInstClass(unsigned Opc, const SIInstrInfo &TII) {
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
     return S_BUFFER_LOAD_IMM;
-  // For the purposes of this optimization SGPR variants of buffer loads
-  // are considered to be zero-offsetted SGPR_IMM loads.
-  case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM:
@@ -537,12 +527,6 @@ static unsigned getInstSubclass(unsigned Opc, const SIInstrInfo &TII) {
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
     return AMDGPU::S_BUFFER_LOAD_DWORD_IMM;
-  // For the purposes of this optimization SGPR variants of buffer loads
-  // are considered to be zero-offsetted SGPR_IMM loads.
-  case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM:
@@ -645,10 +629,6 @@ static AddressRegs getRegs(unsigned Opc, const SIInstrInfo &TII) {
   switch (Opc) {
   default:
     return Result;
-  case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR:
-  case AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR:
   case AMDGPU::S_BUFFER_LOAD_DWORD_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM:
@@ -744,7 +724,7 @@ void SILoadStoreOptimizer::CombineInfo::setMI(MachineBasicBlock::iterator MI,
     Offset = 0;
   } else {
     int OffsetIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::offset);
-    Offset = OffsetIdx == -1 ? 0 : I->getOperand(OffsetIdx).getImm();
+    Offset = I->getOperand(OffsetIdx).getImm();
   }
 
   if (InstClass == TBUFFER_LOAD || InstClass == TBUFFER_STORE)
@@ -1387,10 +1367,7 @@ MachineBasicBlock::iterator SILoadStoreOptimizer::mergeSMemLoadImmPair(
           .add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::sbase));
   if (CI.InstClass == S_BUFFER_LOAD_SGPR_IMM)
     New.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::soffset));
-  // For convenience, when SGPR_IMM buffer loads are merged into a
-  // zero-offset load, we generate its SGPR variant.
-  if (AMDGPU::hasNamedOperand(Opcode, AMDGPU::OpName::offset))
-    New.addImm(MergedOffset);
+  New.addImm(MergedOffset);
   New.addImm(CI.CPol).addMemOperand(combineKnownAdjacentMMOs(CI, Paired));
 
   std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI, Paired);
@@ -1701,14 +1678,11 @@ unsigned SILoadStoreOptimizer::getNewOpcode(const CombineInfo &CI,
     default:
       return 0;
     case 2:
-      return CI.Offset == 0 ? AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR
-                            : AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM;
+      return AMDGPU::S_BUFFER_LOAD_DWORDX2_SGPR_IMM;
     case 4:
-      return CI.Offset == 0 ? AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR
-                            : AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM;
+      return AMDGPU::S_BUFFER_LOAD_DWORDX4_SGPR_IMM;
     case 8:
-      return CI.Offset == 0 ? AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR
-                            : AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR_IMM;
+      return AMDGPU::S_BUFFER_LOAD_DWORDX8_SGPR_IMM;
     }
   case S_LOAD_IMM:
     switch (Width) {

diff --git a/llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir b/llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir
index 5fc8a0e..5ff0dbe 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir
+++ b/llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir
@@ -132,7 +132,7 @@ body:             |
 ---
 
 # CHECK-LABEL: name: merge_s_buffer_load_sgpr_imm
-# CHECK: S_BUFFER_LOAD_DWORDX4_SGPR %0, %1, 0 :: (dereferenceable invariant load (s128), align 4)
+# CHECK: S_BUFFER_LOAD_DWORDX4_SGPR_IMM %0, %1, 0, 0 :: (dereferenceable invariant load (s128), align 4)
 name: merge_s_buffer_load_sgpr_imm
 tracksRegLiveness: true
 body:             |
@@ -141,7 +141,7 @@ body:             |
 
     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
     %1:sreg_32 = COPY $sgpr4
-    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR %0:sgpr_128, %1:sreg_32, 0 :: (dereferenceable invariant load (s32))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 0, 0 :: (dereferenceable invariant load (s32))
     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 4, 0 :: (dereferenceable invariant load (s32))
     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 8, 0 :: (dereferenceable invariant load (s32))
     %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 12, 0 :: (dereferenceable invariant load (s32))