[AMDGPU] Add off-by-default flag to control LiveRegOpt

Change-Id: Id939bf74b48b47e5ee2b432956e476fac80e3375
(cherry picked from commit 5b1a599)

Author: jrbyrnes

llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp

@@ -180,7 +180,8 @@ bool AMDGPULateCodeGenPrepare::runOnFunction(Function &F) {
   for (auto &BB : reverse(F))
     for (Instruction &I : make_early_inc_range(reverse(BB))) {
       Changed |= !HasScalarSubwordLoads && visit(I);
-      Changed |= LRO.optimizeLiveType(&I, DeadInsts);
+      if (ST.shouldCoerceIllegalTypes())
+        Changed |= LRO.optimizeLiveType(&I, DeadInsts);
     }
 
   RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts);

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

@@ -59,6 +59,11 @@ static cl::opt<unsigned> NSAThreshold("amdgpu-nsa-threshold",
                                       cl::desc("Number of addresses from which to enable MIMG NSA."),
                                       cl::init(3), cl::Hidden);
 
+static cl::opt<bool>
+    CoerceIllegal("amdgpu-coerce-illegal-types",
+                  cl::desc("Whether or not to coerce illegal types"),
+                  cl::ReallyHidden, cl::init(false));
+
 GCNSubtarget::~GCNSubtarget() = default;
 
 GCNSubtarget &
@@ -199,6 +204,8 @@ GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
   RegBankInfo.reset(new AMDGPURegisterBankInfo(*this));
   InstSelector.reset(new AMDGPUInstructionSelector(
   *this, *static_cast<AMDGPURegisterBankInfo *>(RegBankInfo.get()), TM));
+
+  ShouldCoerceIllegalTypes = CoerceIllegal;
 }
 
 unsigned GCNSubtarget::getConstantBusLimit(unsigned Opcode) const {

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

@@ -314,7 +314,8 @@ unsigned GCNTTIImpl::getNumberOfParts(Type *Tp) {
   // queries (e.g. get*InstrCost) to decide the proper handling
   // of 8 bit vectors.
   if (FixedVectorType *VTy = dyn_cast<FixedVectorType>(Tp)) {
-    if (DL.getTypeSizeInBits(VTy->getElementType()) == 8) {
+    if (ST->shouldCoerceIllegalTypes() &&
+        DL.getTypeSizeInBits(VTy->getElementType()) == 8) {
       unsigned ElCount = VTy->getElementCount().getFixedValue();
       return PowerOf2Ceil(ElCount / 4);
     }
@@ -355,10 +356,10 @@ unsigned GCNTTIImpl::getMaximumVF(unsigned ElemWidth, unsigned Opcode) const {
   if (Opcode == Instruction::Load || Opcode == Instruction::Store)
     return 32 * 4 / ElemWidth;
 
-  return (ElemWidth == 8)                              ? 4
-         : (ElemWidth == 16)                           ? 2
-         : (ElemWidth == 32 && ST->hasPackedFP32Ops()) ? 2
-                                                       : 1;
+  return (ST->shouldCoerceIllegalTypes() && ElemWidth == 8) ? 4
+         : (ElemWidth == 16)                                ? 2
+         : (ElemWidth == 32 && ST->hasPackedFP32Ops())      ? 2
+                                                            : 1;
 }
 
 unsigned GCNTTIImpl::getLoadVectorFactor(unsigned VF, unsigned LoadSize,
@@ -1154,7 +1155,8 @@ InstructionCost GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
 
   unsigned ScalarSize = DL.getTypeSizeInBits(VT->getElementType());
   if (ST->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
-      (ScalarSize == 16 || ScalarSize == 8)) {
+      (ScalarSize == 16 ||
+       (ScalarSize == 8 && ST->shouldCoerceIllegalTypes()))) {
     // Larger vector widths may require additional instructions, but are
     // typically cheaper than scalarized versions.
     unsigned NumVectorElts = cast<FixedVectorType>(VT)->getNumElements();

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -235,6 +235,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   // Dummy feature to use for assembler in tablegen.
   bool FeatureDisable = false;
 
+  bool ShouldCoerceIllegalTypes = false;
+
   SelectionDAGTargetInfo TSInfo;
 private:
   SIInstrInfo InstrInfo;
@@ -1305,6 +1307,10 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   // of sign-extending.
   bool hasGetPCZeroExtension() const { return GFX12Insts; }
 
+  /// \returns whether or not we should coerce illegal types into vectors of
+  // legal types for values that span basic blocks.
+  bool shouldCoerceIllegalTypes() const { return ShouldCoerceIllegalTypes; }
+
   /// \returns SGPR allocation granularity supported by the subtarget.
   unsigned getSGPRAllocGranule() const {
     return AMDGPU::IsaInfo::getSGPRAllocGranule(this);

llvm/test/Analysis/CostModel/AMDGPU/shufflevector.ll

@@ -8,30 +8,29 @@ define amdgpu_kernel void @eggs(i1 %arg, ptr addrspace(1) %arg1, ptr %arg2, ptr
 ; CHECK:       ; %bb.0: ; %bb
 ; CHECK-NEXT:    s_load_dword s0, s[4:5], 0x0
 ; CHECK-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x8
-; CHECK-NEXT:    v_mov_b32_e32 v0, 0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_bitcmp0_b32 s0, 0
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB0_2
 ; CHECK-NEXT:  ; %bb.1: ; %bb10
-; CHECK-NEXT:    global_load_dwordx2 v[8:9], v0, s[8:9]
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    global_load_dwordx2 v[0:1], v0, s[8:9]
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; CHECK-NEXT:    v_bfe_u32 v6, v8, 8, 8
-; CHECK-NEXT:    v_bfe_u32 v5, v8, 16, 8
-; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 24, v8
-; CHECK-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; CHECK-NEXT:    v_bfe_u32 v2, v9, 8, 8
-; CHECK-NEXT:    v_bfe_u32 v1, v9, 16, 8
-; CHECK-NEXT:    v_lshrrev_b32_e32 v0, 24, v9
+; CHECK-NEXT:    v_lshrrev_b32_e32 v7, 8, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v5, 24, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v2, 24, v1
 ; CHECK-NEXT:    s_branch .LBB0_3
 ; CHECK-NEXT:  .LBB0_2:
-; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v3, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v4, 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v5, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v6, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v7, 0
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
 ; CHECK-NEXT:  .LBB0_3: ; %bb41
 ; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x48
 ; CHECK-NEXT:    v_mov_b32_e32 v8, s10
@@ -48,16 +47,16 @@ define amdgpu_kernel void @eggs(i1 %arg, ptr addrspace(1) %arg1, ptr %arg2, ptr
 ; CHECK-NEXT:    v_mov_b32_e32 v19, s21
 ; CHECK-NEXT:    v_mov_b32_e32 v20, s22
 ; CHECK-NEXT:    v_mov_b32_e32 v21, s23
-; CHECK-NEXT:    flat_store_byte v[8:9], v7
-; CHECK-NEXT:    flat_store_byte v[10:11], v6
-; CHECK-NEXT:    flat_store_byte v[12:13], v5
-; CHECK-NEXT:    flat_store_byte v[14:15], v4
-; CHECK-NEXT:    flat_store_byte v[16:17], v3
-; CHECK-NEXT:    flat_store_byte v[18:19], v2
-; CHECK-NEXT:    flat_store_byte v[20:21], v1
+; CHECK-NEXT:    flat_store_byte v[8:9], v0
+; CHECK-NEXT:    flat_store_byte v[10:11], v7
+; CHECK-NEXT:    flat_store_byte v[12:13], v6
+; CHECK-NEXT:    flat_store_byte v[14:15], v5
+; CHECK-NEXT:    flat_store_byte v[16:17], v1
+; CHECK-NEXT:    flat_store_byte v[18:19], v4
+; CHECK-NEXT:    flat_store_byte v[20:21], v3
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    v_pk_mov_b32 v[2:3], s[0:1], s[0:1] op_sel:[0,1]
-; CHECK-NEXT:    flat_store_byte v[2:3], v0
+; CHECK-NEXT:    v_pk_mov_b32 v[0:1], s[0:1], s[0:1] op_sel:[0,1]
+; CHECK-NEXT:    flat_store_byte v[0:1], v2
 ; CHECK-NEXT:    s_endpgm
 bb:
   br i1 %arg, label %bb10, label %bb41