[llvm] [Matrix] Propagate shape information through cast instructions (PR #141869)

Jon Roelofs via llvm-commits llvm-commits at lists.llvm.org
Wed May 28 17:49:00 PDT 2025


https://github.com/jroelofs updated https://github.com/llvm/llvm-project/pull/141869

>From 42a624af192133381c84745116211f320ea37dce Mon Sep 17 00:00:00 2001
From: Jon Roelofs <jonathan_roelofs at apple.com>
Date: Wed, 28 May 2025 15:32:53 -0700
Subject: [PATCH] [Matrix] Propagate shape information through cast
 instructions

---
 .../Scalar/LowerMatrixIntrinsics.cpp          |  81 ++++++-
 .../Transforms/LowerMatrixIntrinsics/unary.ll | 202 ++++++++++++++++++
 2 files changed, 281 insertions(+), 2 deletions(-)
 create mode 100644 llvm/test/Transforms/LowerMatrixIntrinsics/unary.ll

diff --git a/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp b/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
index 756a72e6d97bc..e355a096e00b7 100644
--- a/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
+++ b/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
@@ -32,8 +32,10 @@
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/DebugInfoMetadata.h"
+#include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/MatrixBuilder.h"
@@ -232,6 +234,28 @@ static bool isUniformShape(Value *V) {
   if (I->isBinaryOp())
     return true;
 
+  if (auto *Cast = dyn_cast<CastInst>(V))
+    switch (Cast->getOpcode()) {
+    case llvm::Instruction::Trunc:
+    case llvm::Instruction::ZExt:
+    case llvm::Instruction::SExt:
+    case llvm::Instruction::FPToUI:
+    case llvm::Instruction::FPToSI:
+    case llvm::Instruction::UIToFP:
+    case llvm::Instruction::SIToFP:
+    case llvm::Instruction::FPTrunc:
+    case llvm::Instruction::FPExt:
+      return true;
+    case llvm::Instruction::AddrSpaceCast:
+    case CastInst::PtrToInt:
+    case CastInst::IntToPtr:
+    case CastInst::BitCast:
+      return false;
+    case llvm::Instruction::CastOpsEnd:
+      llvm_unreachable("not an actual cast op");
+    }
+
+
   switch (I->getOpcode()) {
   case Instruction::FNeg:
     return true;
@@ -1066,9 +1090,11 @@ class LowerMatrixIntrinsics {
       Value *Op2;
       if (auto *BinOp = dyn_cast<BinaryOperator>(Inst))
         Changed |= VisitBinaryOperator(BinOp);
-      if (auto *UnOp = dyn_cast<UnaryOperator>(Inst))
+      else if (auto *UnOp = dyn_cast<UnaryOperator>(Inst))
         Changed |= VisitUnaryOperator(UnOp);
-      if (match(Inst, m_Load(m_Value(Op1))))
+      else if (auto *Cast = dyn_cast<CastInst>(Inst))
+        Changed |= VisitCastInstruction(Cast);
+      else if (match(Inst, m_Load(m_Value(Op1))))
         Changed |= VisitLoad(cast<LoadInst>(Inst), Op1, Builder);
       else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
         Changed |= VisitStore(cast<StoreInst>(Inst), Op1, Op2, Builder);
@@ -2198,6 +2224,57 @@ class LowerMatrixIntrinsics {
     return true;
   }
 
+  /// Lower cast instructions, if shape information is available.
+  bool VisitCastInstruction(CastInst *Inst) {
+    switch (Inst->getOpcode()) {
+    case llvm::Instruction::Trunc:
+    case llvm::Instruction::ZExt:
+    case llvm::Instruction::SExt:
+    case llvm::Instruction::FPToUI:
+    case llvm::Instruction::FPToSI:
+    case llvm::Instruction::UIToFP:
+    case llvm::Instruction::SIToFP:
+    case llvm::Instruction::FPTrunc:
+    case llvm::Instruction::FPExt:
+      break;
+    case llvm::Instruction::AddrSpaceCast:
+    case CastInst::PtrToInt:
+    case CastInst::IntToPtr:
+    case CastInst::BitCast:
+      return false;
+    case llvm::Instruction::CastOpsEnd:
+      llvm_unreachable("not an actual cast op");
+    }
+
+    auto I = ShapeMap.find(Inst);
+    if (I == ShapeMap.end())
+      return false;
+
+    Value *Op = Inst->getOperand(0);
+
+    IRBuilder<> Builder(Inst);
+    ShapeInfo &Shape = I->second;
+
+    MatrixTy Result;
+    MatrixTy M = getMatrix(Op, Shape, Builder);
+
+    Builder.setFastMathFlags(getFastMathFlags(Inst));
+
+    for (unsigned I = 0; I < Shape.getNumVectors(); ++I) {
+      auto *OrigTy = cast<VectorType>(Inst->getType());
+      auto *NewTy = VectorType::get(OrigTy->getElementType(),
+                                    ElementCount::getFixed(M.getStride()));
+      Result.addVector(
+          Builder.CreateCast(Inst->getOpcode(), M.getVector(I), NewTy));
+    }
+
+    finalizeLowering(Inst,
+                     Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
+                                             Result.getNumVectors()),
+                     Builder);
+    return true;
+  }
+
   /// Helper to linearize a matrix expression tree into a string. Currently
   /// matrix expressions are linarized by starting at an expression leaf and
   /// linearizing bottom up.
diff --git a/llvm/test/Transforms/LowerMatrixIntrinsics/unary.ll b/llvm/test/Transforms/LowerMatrixIntrinsics/unary.ll
new file mode 100644
index 0000000000000..b73c1402044bc
--- /dev/null
+++ b/llvm/test/Transforms/LowerMatrixIntrinsics/unary.ll
@@ -0,0 +1,202 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s
+
+define void @fneg_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fneg_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x float>, ptr [[IN:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr float, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg <2 x float> [[COL_LOAD]]
+; CHECK-NEXT:    [[TMP2:%.*]] = fneg <2 x float> [[COL_LOAD1]]
+; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr float, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x float> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x float>, ptr %in
+  %op = fneg <4 x float> %inv
+  %opt  = call <4 x float> @llvm.matrix.transpose(<4 x float> %op, i32 2, i32 2)
+  %optt = call <4 x float> @llvm.matrix.transpose(<4 x float> %opt, i32 2, i32 2)
+  store <4 x float> %optt, ptr %out
+  ret void
+}
+
+define void @trunc_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @trunc_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x i64>, ptr [[IN:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i64, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x i64>, ptr [[VEC_GEP]], align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc <2 x i64> [[COL_LOAD]] to <2 x i32>
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc <2 x i64> [[COL_LOAD1]] to <2 x i32>
+; CHECK-NEXT:    store <2 x i32> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr i32, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x i32> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x i64>, ptr %in
+  %op = trunc <4 x i64> %inv to <4 x i32>
+  %opt  = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %op, i32 2, i32 2)
+  %optt = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %opt, i32 2, i32 2)
+  store <4 x i32> %optt, ptr %out
+  ret void
+}
+
+define void @zext_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @zext_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x i16>, ptr [[IN:%.*]], align 8
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i16, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x i16>, ptr [[VEC_GEP]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = zext <2 x i16> [[COL_LOAD]] to <2 x i32>
+; CHECK-NEXT:    [[TMP2:%.*]] = zext <2 x i16> [[COL_LOAD1]] to <2 x i32>
+; CHECK-NEXT:    store <2 x i32> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr i32, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x i32> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x i16>, ptr %in
+  %op = zext <4 x i16> %inv to <4 x i32>
+  %opt  = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %op, i32 2, i32 2)
+  %optt = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %opt, i32 2, i32 2)
+  store <4 x i32> %optt, ptr %out
+  ret void
+}
+
+define void @sext_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @sext_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x i8>, ptr [[IN:%.*]], align 4
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i8, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x i8>, ptr [[VEC_GEP]], align 2
+; CHECK-NEXT:    [[TMP1:%.*]] = sext <2 x i8> [[COL_LOAD]] to <2 x i16>
+; CHECK-NEXT:    [[TMP2:%.*]] = sext <2 x i8> [[COL_LOAD1]] to <2 x i16>
+; CHECK-NEXT:    store <2 x i16> [[TMP1]], ptr [[OUT:%.*]], align 8
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr i16, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x i16> [[TMP2]], ptr [[VEC_GEP2]], align 4
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x i8>, ptr %in
+  %op = sext <4 x i8> %inv to <4 x i16>
+  %opt  = call <4 x i16> @llvm.matrix.transpose(<4 x i16> %op, i32 2, i32 2)
+  %optt = call <4 x i16> @llvm.matrix.transpose(<4 x i16> %opt, i32 2, i32 2)
+  store <4 x i16> %optt, ptr %out
+  ret void
+}
+
+define void @fptoui_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fptoui_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x float>, ptr [[IN:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr float, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = fptoui <2 x float> [[COL_LOAD]] to <2 x i32>
+; CHECK-NEXT:    [[TMP2:%.*]] = fptoui <2 x float> [[COL_LOAD1]] to <2 x i32>
+; CHECK-NEXT:    store <2 x i32> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr i32, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x i32> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x float>, ptr %in
+  %op = fptoui <4 x float> %inv to <4 x i32>
+  %opt  = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %op, i32 2, i32 2)
+  %optt = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %opt, i32 2, i32 2)
+  store <4 x i32> %optt, ptr %out
+  ret void
+}
+
+define void @fptosi_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fptosi_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x float>, ptr [[IN:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr float, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = fptosi <2 x float> [[COL_LOAD]] to <2 x i32>
+; CHECK-NEXT:    [[TMP2:%.*]] = fptosi <2 x float> [[COL_LOAD1]] to <2 x i32>
+; CHECK-NEXT:    store <2 x i32> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr i32, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x i32> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x float>, ptr %in
+  %op = fptosi <4 x float> %inv to <4 x i32>
+  %opt  = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %op, i32 2, i32 2)
+  %optt = call <4 x i32> @llvm.matrix.transpose(<4 x i32> %opt, i32 2, i32 2)
+  store <4 x i32> %optt, ptr %out
+  ret void
+}
+
+define void @uitofp_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @uitofp_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x i64>, ptr [[IN:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i64, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x i64>, ptr [[VEC_GEP]], align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = uitofp <2 x i64> [[COL_LOAD]] to <2 x double>
+; CHECK-NEXT:    [[TMP2:%.*]] = uitofp <2 x i64> [[COL_LOAD1]] to <2 x double>
+; CHECK-NEXT:    store <2 x double> [[TMP1]], ptr [[OUT:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr double, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x double> [[TMP2]], ptr [[VEC_GEP2]], align 16
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x i64>, ptr %in
+  %op = uitofp <4 x i64> %inv to <4 x double>
+  %opt  = call <4  x double> @llvm.matrix.transpose(<4  x double> %op, i32 2, i32 2)
+  %optt = call <4  x double> @llvm.matrix.transpose(<4  x double> %opt, i32 2, i32 2)
+  store <4  x double> %optt, ptr %out
+  ret void
+}
+
+define void @sitofp_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @sitofp_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x i64>, ptr [[IN:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i64, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x i64>, ptr [[VEC_GEP]], align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = sitofp <2 x i64> [[COL_LOAD]] to <2 x double>
+; CHECK-NEXT:    [[TMP2:%.*]] = sitofp <2 x i64> [[COL_LOAD1]] to <2 x double>
+; CHECK-NEXT:    store <2 x double> [[TMP1]], ptr [[OUT:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr double, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x double> [[TMP2]], ptr [[VEC_GEP2]], align 16
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x i64>, ptr %in
+  %op = sitofp <4 x i64> %inv to <4 x double>
+  %opt  = call <4  x double> @llvm.matrix.transpose(<4  x double> %op, i32 2, i32 2)
+  %optt = call <4  x double> @llvm.matrix.transpose(<4  x double> %opt, i32 2, i32 2)
+  store <4  x double> %optt, ptr %out
+  ret void
+}
+
+define void @fptrunc_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fptrunc_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x double>, ptr [[IN:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc nnan <2 x double> [[COL_LOAD]] to <2 x float>
+; CHECK-NEXT:    [[TMP2:%.*]] = fptrunc nnan <2 x double> [[COL_LOAD1]] to <2 x float>
+; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr float, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x float> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x double>, ptr %in
+  %op = fptrunc nnan <4 x double> %inv to <4 x float>
+  %opt  = call <4 x float> @llvm.matrix.transpose(<4 x float> %op, i32 2, i32 2)
+  %optt = call <4 x float> @llvm.matrix.transpose(<4 x float> %opt, i32 2, i32 2)
+  store <4 x float> %optt, ptr %out
+  ret void
+}
+
+define void @fpext_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fpext_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x float>, ptr [[IN:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr float, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = fpext <2 x float> [[COL_LOAD]] to <2 x double>
+; CHECK-NEXT:    [[TMP2:%.*]] = fpext <2 x float> [[COL_LOAD1]] to <2 x double>
+; CHECK-NEXT:    store <2 x double> [[TMP1]], ptr [[OUT:%.*]], align 32
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr double, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x double> [[TMP2]], ptr [[VEC_GEP2]], align 16
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x float>, ptr %in
+  %op = fpext <4 x float> %inv to <4 x double>
+  %opt  = call <4 x double> @llvm.matrix.transpose(<4 x double> %op, i32 2, i32 2)
+  %optt = call <4 x double> @llvm.matrix.transpose(<4 x double> %opt, i32 2, i32 2)
+  store <4 x double> %optt, ptr %out
+  ret void
+}



More information about the llvm-commits mailing list