[AMDGPU] Filter candidates of LiveRegOptimizer for profitable cases

It is known that for vector whose element fits in i16 will be split
and scalarized in SelectionDag's type legalizer
(see SIISelLowering::getPreferredVectorAction).

LRO attempts to undo the scalarizing of vectors across basic block
boundary and shoehorn Values in VGPRs. LRO is beneficial for operations
that natively work on illegal vector types to prevent flip-flopping
between unpacked and packed. If we know that operations on vector will be
split and scalarized, then we don't want to shoehorn them back to packed
VGPR.

Operations that we know to work natively on illegal vector types
usually come in the form of intrinsics (MFMA, DOT8), buffer store,
shuffle, insert/extract element, phi nodes to name a few.

Author: choikwa

llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp

@@ -20,6 +20,7 @@
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstVisitor.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/KnownBits.h"
 #include "llvm/Transforms/Utils/Local.h"
@@ -75,6 +76,7 @@ class LiveRegOptimizer {
   Module &Mod;
   const DataLayout &DL;
   const GCNSubtarget &ST;
+
   /// The scalar type to convert to
   Type *const ConvertToScalar;
   /// The set of visited Instructions
@@ -125,6 +127,119 @@ class LiveRegOptimizer {
     return LK.first != TargetLoweringBase::TypeLegal;
   }
 
+  /// Check if intrinsic natively operates on 8-bit or 16-bit
+  bool isNativeIntrinsic(Intrinsic::ID ID) {
+    switch (ID) {
+    case Intrinsic::amdgcn_dot4_f32_fp8_bf8:
+    case Intrinsic::amdgcn_dot4_f32_bf8_fp8:
+    case Intrinsic::amdgcn_dot4_f32_fp8_fp8:
+    case Intrinsic::amdgcn_dot4_f32_bf8_bf8:
+    case Intrinsic::amdgcn_mfma_i32_4x4x4i8:
+    case Intrinsic::amdgcn_mfma_i32_16x16x4i8:
+    case Intrinsic::amdgcn_mfma_i32_32x32x4i8:
+    case Intrinsic::amdgcn_mfma_i32_16x16x16i8:
+    case Intrinsic::amdgcn_mfma_i32_32x32x8i8:
+    case Intrinsic::amdgcn_mfma_i32_16x16x64_i8:
+    case Intrinsic::amdgcn_mfma_i32_32x32x32_i8:
+    case Intrinsic::amdgcn_mfma_i32_32x32x16_i8:
+    case Intrinsic::amdgcn_mfma_i32_16x16x32_i8:
+    case Intrinsic::amdgcn_mfma_f32_16x16x32_bf8_bf8:
+    case Intrinsic::amdgcn_mfma_f32_16x16x32_bf8_fp8:
+    case Intrinsic::amdgcn_mfma_f32_16x16x32_fp8_bf8:
+    case Intrinsic::amdgcn_mfma_f32_16x16x32_fp8_fp8:
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_bf8_bf8:
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_bf8_fp8:
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_fp8_bf8:
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_fp8_fp8:
+    case Intrinsic::amdgcn_smfmac_i32_16x16x64_i8:
+    case Intrinsic::amdgcn_smfmac_i32_32x32x32_i8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x64_bf8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x64_bf8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x64_fp8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x64_fp8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x32_bf8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x32_bf8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x32_fp8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x32_fp8_fp8:
+    case Intrinsic::amdgcn_smfmac_i32_16x16x128_i8:
+    case Intrinsic::amdgcn_smfmac_i32_32x32x64_i8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x64_fp8_bf8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x64_fp8_fp8:
+    case Intrinsic::amdgcn_wmma_f32_16x16x16_fp8_fp8:
+    case Intrinsic::amdgcn_wmma_f32_16x16x16_fp8_bf8:
+    case Intrinsic::amdgcn_wmma_f32_16x16x16_bf8_fp8:
+    case Intrinsic::amdgcn_wmma_f32_16x16x16_bf8_bf8:
+    case Intrinsic::amdgcn_swmmac_f32_16x16x32_fp8_fp8:
+    case Intrinsic::amdgcn_swmmac_f32_16x16x32_fp8_bf8:
+    case Intrinsic::amdgcn_swmmac_f32_16x16x32_bf8_fp8:
+    case Intrinsic::amdgcn_swmmac_f32_16x16x32_bf8_bf8:
+    case Intrinsic::amdgcn_wmma_i32_16x16x16_iu8:
+    case Intrinsic::amdgcn_wmma_i32_16x16x16_iu4:
+    case Intrinsic::amdgcn_wmma_i32_16x16x32_iu4:
+    case Intrinsic::amdgcn_swmmac_i32_16x16x32_iu8:
+    case Intrinsic::amdgcn_swmmac_i32_16x16x32_iu4:
+    case Intrinsic::amdgcn_swmmac_i32_16x16x64_iu4:
+      return true;
+    default:
+      return false;
+    }
+  }
+
+  bool isOpLegal(Instruction *I) {
+    if (const auto *Intr = dyn_cast<IntrinsicInst>(I)) {
+      Intrinsic::ID ID = Intr->getIntrinsicID();
+      if (isNativeIntrinsic(ID))
+        return true;
+    }
+    // Stores
+    if (isa<StoreInst>(I))
+      return true;
+    return false;
+  }
+
+  bool isCoercionProfitable(Instruction *II) {
+    SmallPtrSet<Instruction *, 4> CVisited;
+    SmallVector<Instruction *, 4> UserList;
+
+    // Check users for profitable conditions (across block user which can
+    // natively handle the illegal vector).
+    for (User *V : II->users())
+      if (auto *UseInst = dyn_cast<Instruction>(V))
+        UserList.push_back(UseInst);
+
+    auto IsLookThru = [](Instruction *II) {
+      if (const auto *Intr = dyn_cast<IntrinsicInst>(II))
+        return Intr->getIntrinsicID() == Intrinsic::amdgcn_perm;
+      return isa<PHINode>(II) || isa<ShuffleVectorInst>(II) ||
+             isa<InsertElementInst>(II) || isa<ExtractElementInst>(II) ||
+             isa<CastInst>(II);
+    };
+
+    while (!UserList.empty()) {
+      auto CII = UserList.pop_back_val();
+      if (!CVisited.insert(CII).second)
+        continue;
+
+      if (CII->getParent() == II->getParent() && !IsLookThru(II))
+        continue;
+
+      if (isOpLegal(CII))
+        return true;
+
+      if (IsLookThru(CII))
+        for (User *V : CII->users())
+          if (auto *UseInst = dyn_cast<Instruction>(V))
+            UserList.push_back(UseInst);
+    }
+    return false;
+  }
+
   LiveRegOptimizer(Module &Mod, const GCNSubtarget &ST)
       : Mod(Mod), DL(Mod.getDataLayout()), ST(ST),
         ConvertToScalar(Type::getInt32Ty(Mod.getContext())) {}
@@ -259,6 +374,9 @@ bool LiveRegOptimizer::optimizeLiveType(
     if (!shouldReplace(II->getType()))
       continue;
 
+    if (!isCoercionProfitable(II))
+      continue;
+
     if (PHINode *Phi = dyn_cast<PHINode>(II)) {
       PhiNodes.insert(Phi);
       // Collect all the incoming values of problematic PHI nodes.
@@ -478,7 +596,6 @@ bool AMDGPULateCodeGenPrepare::visitLoadInst(LoadInst &LI) {
 PreservedAnalyses
 AMDGPULateCodeGenPreparePass::run(Function &F, FunctionAnalysisManager &FAM) {
   const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
-
   AssumptionCache &AC = FAM.getResult<AssumptionAnalysis>(F);
   UniformityInfo &UI = FAM.getResult<UniformityInfoAnalysis>(F);