[AMDGPU] Add merging into S_BUFFER_LOAD_DWORDX8_IMM

[piotr]

Summary

Extend SILoadStoreOptimizer to merge into DWORDX8 variant of S_BUFFER_LOAD.

Merging into DWORDX2 and DWORDX4 variants is handled already.

[piotr]

Add test.

[foad]

LGTM.

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

static unsigned getOpcodeWidth(const MachineInstr &MI, const SIInstrInfo &TII) {
 
   switch (Opc) {
   case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
     return 1;
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
     return 2;
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
     return 4;
+  case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
+    return 8;
   case AMDGPU::DS_READ_B32:      LLVM_FALLTHROUGH;
   case AMDGPU::DS_READ_B32_gfx9: LLVM_FALLTHROUGH;
   case AMDGPU::DS_WRITE_B32:     LLVM_FALLTHROUGH;
   case AMDGPU::DS_WRITE_B32_gfx9:
     return 1;
   case AMDGPU::DS_READ_B64:      LLVM_FALLTHROUGH;
   case AMDGPU::DS_READ_B64_gfx9: LLVM_FALLTHROUGH;
   case AMDGPU::DS_WRITE_B64:     LLVM_FALLTHROUGH;
if (TII.isMTBUF(Opc)) {
       case AMDGPU::TBUFFER_STORE_FORMAT_X_OFFSET_exact:
         return TBUFFER_STORE;
       }
     }
     return UNKNOWN;
   case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
+  case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
     return S_BUFFER_LOAD_IMM;
   case AMDGPU::DS_READ_B32:
   case AMDGPU::DS_READ_B32_gfx9:
   case AMDGPU::DS_READ_B64:
   case AMDGPU::DS_READ_B64_gfx9:
     return DS_READ;
   case AMDGPU::DS_WRITE_B32:
   case AMDGPU::DS_WRITE_B32_gfx9:
static unsigned getInstSubclass(unsigned Opc, const SIInstrInfo &TII) {
   case AMDGPU::DS_WRITE_B32:
   case AMDGPU::DS_WRITE_B32_gfx9:
   case AMDGPU::DS_WRITE_B64:
   case AMDGPU::DS_WRITE_B64_gfx9:
     return Opc;
   case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
+  case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
     return AMDGPU::S_BUFFER_LOAD_DWORD_IMM;
   }
 }
 
 static AddressRegs getRegs(unsigned Opc, const SIInstrInfo &TII) {
   AddressRegs Result;
 
   if (TII.isMUBUF(Opc)) {
static AddressRegs getRegs(unsigned Opc, const SIInstrInfo &TII) {
   }
 
   switch (Opc) {
   default:
     return Result;
   case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
   case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
+  case AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM:
     Result.SBase = true;
     return Result;
   case AMDGPU::DS_READ_B32:
   case AMDGPU::DS_READ_B64:
   case AMDGPU::DS_READ_B32_gfx9:
   case AMDGPU::DS_READ_B64_gfx9:
   case AMDGPU::DS_WRITE_B32:
   case AMDGPU::DS_WRITE_B64:
bool SILoadStoreOptimizer::widthsFit(const GCNSubtarget &STM,
   default:
     return (Width <= 4) && (STM.hasDwordx3LoadStores() || (Width != 3));
   case S_BUFFER_LOAD_IMM:
     switch (Width) {
     default:
       return false;
     case 2:
     case 4:
+    case 8:
       return true;
     }
   }
 }
 
 const TargetRegisterClass *
 SILoadStoreOptimizer::getDataRegClass(const MachineInstr &MI) const {
   if (const auto *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst)) {
unsigned SILoadStoreOptimizer::getNewOpcode(const CombineInfo &CI,
   case S_BUFFER_LOAD_IMM:
     switch (Width) {
     default:
       return 0;
     case 2:
       return AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM;
     case 4:
       return AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM;
+    case 8:
+      return AMDGPU::S_BUFFER_LOAD_DWORDX8_IMM;
     }
   case MIMG:
-    assert("No overlaps" && (countPopulation(CI.DMask | Paired.DMask) == Width));
+    assert((countPopulation(CI.DMask | Paired.DMask) == Width) &&
+           "No overlaps");
     return AMDGPU::getMaskedMIMGOp(CI.I->getOpcode(), Width);
   }
 }
 
 std::pair<unsigned, unsigned>
-SILoadStoreOptimizer::getSubRegIdxs(const CombineInfo &CI, const CombineInfo &Paired) {
+SILoadStoreOptimizer::getSubRegIdxs(const CombineInfo &CI,
+                                    const CombineInfo &Paired) {
 
-  if (CI.Width == 0 || Paired.Width == 0 || CI.Width + Paired.Width > 4)
-    return std::make_pair(0, 0);
+  assert(CI.Width != 0 && Paired.Width != 0 && "Width cannot be zero");
 
   bool ReverseOrder;
   if (CI.InstClass == MIMG) {
-    assert((countPopulation(CI.DMask | Paired.DMask) == CI.Width + Paired.Width) &&
-           "No overlaps");
+    assert(
+        (countPopulation(CI.DMask | Paired.DMask) == CI.Width + Paired.Width) &&
     ReverseOrder = CI.DMask > Paired.DMask;
   } else
     ReverseOrder = CI.Offset > Paired.Offset;
 
+  unsigned Idx0;
+  unsigned Idx1;
+
+  if (CI.Width + Paired.Width > 4) {
+    assert(CI.Width == 4 && Paired.Width == 4);
+
+    if (ReverseOrder) {
+      Idx1 = AMDGPU::sub0_sub1_sub2_sub3;
+      Idx0 = AMDGPU::sub4_sub5_sub6_sub7;
+    } else {
+      Idx0 = AMDGPU::sub0_sub1_sub2_sub3;
+      Idx1 = AMDGPU::sub4_sub5_sub6_sub7;
+    }
+  } else {
     static const unsigned Idxs[4][4] = {
         {AMDGPU::sub0, AMDGPU::sub0_sub1, AMDGPU::sub0_sub1_sub2, AMDGPU::sub0_sub1_sub2_sub3},
         {AMDGPU::sub1, AMDGPU::sub1_sub2, AMDGPU::sub1_sub2_sub3, 0},
         {AMDGPU::sub2, AMDGPU::sub2_sub3, 0, 0},
         {AMDGPU::sub3, 0, 0, 0},
     };
-  unsigned Idx0;
-  unsigned Idx1;
 
     assert(CI.Width >= 1 && CI.Width <= 3);
     assert(Paired.Width >= 1 && Paired.Width <= 3);
 
     if (ReverseOrder) {
       Idx1 = Idxs[0][Paired.Width - 1];
       Idx0 = Idxs[Paired.Width][CI.Width - 1];
     } else {
       Idx0 = Idxs[0][CI.Width - 1];
       Idx1 = Idxs[CI.Width][Paired.Width - 1];
     }
+  }
 
   return std::make_pair(Idx0, Idx1);
 }
 
 const TargetRegisterClass *
 SILoadStoreOptimizer::getTargetRegisterClass(const CombineInfo &CI,
                                              const CombineInfo &Paired) {
   if (CI.InstClass == S_BUFFER_LOAD_IMM) {
case DS_WRITE: {
           mergeWrite2Pair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       break;
     }
     case S_BUFFER_LOAD_IMM: {
       MachineBasicBlock::iterator NewMI =
           mergeSBufferLoadImmPair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
-      OptimizeListAgain |= (CI.Width + Paired.Width) < 16;
+      OptimizeListAgain |= (CI.Width + Paired.Width) < 8;
       break;
     }
     case BUFFER_LOAD: {
       MachineBasicBlock::iterator NewMI =
           mergeBufferLoadPair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;

llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir

+# RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s
+
+# CHECK-LABEL: name: merge_s_buffer_load_x2
+# CHECK: S_BUFFER_LOAD_DWORDX2_IMM %0, 0, 0 :: (dereferenceable invariant load (s64), align 4)
+
+name: merge_s_buffer_load_x2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
+
+    S_ENDPGM 0
+...
+---
+
+# CHECK-LABEL: name: merge_s_buffer_load_x4
+# CHECK: S_BUFFER_LOAD_DWORDX4_IMM %0, 0, 0 :: (dereferenceable invariant load (s128), align 4)
+name: merge_s_buffer_load_x4
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
+    %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
+    %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
+
+    S_ENDPGM 0
+...
+---
+
+# CHECK-LABEL: name: merge_s_buffer_load_x8
+# CHECK: S_BUFFER_LOAD_DWORDX8_IMM %0, 0, 0 :: (dereferenceable invariant load (s256), align 4)
+name: merge_s_buffer_load_x8
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
+    %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
+    %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
+    %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
+    %6:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
+    %7:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s32))
+    %8:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 28, 0 :: (dereferenceable invariant load (s32))
+
+    S_ENDPGM 0
+...
+---
+
+# CHECK-LABEL: name: merge_s_buffer_load_x8_reordered
+# CHECK: S_BUFFER_LOAD_DWORDX8_IMM %0, 0, 0 :: (dereferenceable invariant load (s256), align 4)
+name: merge_s_buffer_load_x8_reordered
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
+    %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
+    %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 28, 0 :: (dereferenceable invariant load (s32))
+    %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
+    %6:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
+    %7:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
+    %8:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s32))
+
+    S_ENDPGM 0
+...
+---
+
+# CHECK-LABEL: name: merge_s_buffer_load_x8_out_of_x2
+# CHECK: S_BUFFER_LOAD_DWORDX8_IMM %0, 0, 0 :: (dereferenceable invariant load (s256), align 8)
+name: merge_s_buffer_load_x8_out_of_x2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s64))
+    %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s64))
+    %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
+    %4:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
+
+    S_ENDPGM 0
+...
+---
+
+# CHECK-LABEL: name: merge_s_buffer_load_x8_out_of_x4
+# CHECK: S_BUFFER_LOAD_DWORDX8_IMM %0, 0, 0 :: (dereferenceable invariant load (s256), align 16)
+name: merge_s_buffer_load_x8_out_of_x4
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
+    %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
+
+    S_ENDPGM 0
+...
+---
+
+
+# CHECK-LABEL: name: merge_s_buffer_load_x8_mixed
+# CHECK: S_BUFFER_LOAD_DWORDX8_IMM %0, 0, 0 :: (dereferenceable invariant load (s256), align 16)
+name: merge_s_buffer_load_x8_mixed
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
+    %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
+    %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
+    %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
+
+    S_ENDPGM 0
+...
+---