[RISCV] Recommit "Expand vp.stride.load to splat of a scalar load."

[yetingk]

This is a recommit of #98140. The old commit should be rebased on #98205 which changes the feature of hardware zero stride optimization.

It's a similar patch as a214c52 for vp.stride.load. Some targets prefer pattern (vmv.v.x (load)) instead of vlse with zero stride.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Yeting Kuo (yetingk)

Changes

This is a recommit of #98140. The old commit should be rebased on #98205 which changes the feature of hardware zero stride optimization.

It's a similar patch as a214c52 for vp.stride.load. Some targets prefer pattern (vmv.v.x (load)) instead of vlse with zero stride.

---

Full diff: https://github.com/llvm/llvm-project/pull/98579.diff

3 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp (+50)
• (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll (+30-4)
• (modified) llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll (+24)

diff --git a/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp b/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
index 6e0f429c34b2f..10e0496f16d4f 100644
--- a/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
+++ b/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
@@ -18,9 +18,11 @@
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/Dominators.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstVisitor.h"
 #include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsRISCV.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
@@ -35,6 +37,7 @@ namespace {
 class RISCVCodeGenPrepare : public FunctionPass,
                             public InstVisitor<RISCVCodeGenPrepare, bool> {
   const DataLayout *DL;
+  const DominatorTree *DT;
   const RISCVSubtarget *ST;
 
 public:
@@ -48,12 +51,14 @@ class RISCVCodeGenPrepare : public FunctionPass,
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
+    AU.addRequired<DominatorTreeWrapperPass>();
     AU.addRequired<TargetPassConfig>();
   }
 
   bool visitInstruction(Instruction &I) { return false; }
   bool visitAnd(BinaryOperator &BO);
   bool visitIntrinsicInst(IntrinsicInst &I);
+  bool expandVPStrideLoad(IntrinsicInst &I);
 };
 
 } // end anonymous namespace
@@ -128,6 +133,9 @@ bool RISCVCodeGenPrepare::visitAnd(BinaryOperator &BO) {
 // Which eliminates the scalar -> vector -> scalar crossing during instruction
 // selection.
 bool RISCVCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) {
+  if (expandVPStrideLoad(I))
+    return true;
+
   if (I.getIntrinsicID() != Intrinsic::vector_reduce_fadd)
     return false;
 
@@ -155,6 +163,47 @@ bool RISCVCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) {
   return true;
 }
 
+bool RISCVCodeGenPrepare::expandVPStrideLoad(IntrinsicInst &II) {
+  if (ST->hasOptimizedZeroStrideLoad())
+    return false;
+
+  Value *BasePtr, *VL;
+  using namespace PatternMatch;
+  if (!match(&II, m_Intrinsic<Intrinsic::experimental_vp_strided_load>(
+                      m_Value(BasePtr), m_Zero(), m_AllOnes(), m_Value(VL))))
+    return false;
+
+  if (!isKnownNonZero(VL, {*DL, DT, nullptr, &II}))
+    return false;
+
+  auto *VTy = cast<VectorType>(II.getType());
+
+  IRBuilder<> Builder(&II);
+
+  // Extend VL from i32 to XLen if needed.
+  if (ST->is64Bit())
+    VL = Builder.CreateZExt(VL, Builder.getInt64Ty());
+
+  Type *STy = VTy->getElementType();
+  Value *Val = Builder.CreateLoad(STy, BasePtr);
+  const auto &TLI = *ST->getTargetLowering();
+  Value *Res;
+
+  // TODO: Also support fixed/illegal vector types to splat with evl = vl.
+  if (isa<ScalableVectorType>(VTy) && TLI.isTypeLegal(EVT::getEVT(VTy))) {
+    unsigned VMVOp = STy->isFloatingPointTy() ? Intrinsic::riscv_vfmv_v_f
+                                              : Intrinsic::riscv_vmv_v_x;
+    Res = Builder.CreateIntrinsic(VMVOp, {VTy, VL->getType()},
+                                  {PoisonValue::get(VTy), Val, VL});
+  } else {
+    Res = Builder.CreateVectorSplat(VTy->getElementCount(), Val);
+  }
+
+  II.replaceAllUsesWith(Res);
+  II.eraseFromParent();
+  return true;
+}
+
 bool RISCVCodeGenPrepare::runOnFunction(Function &F) {
   if (skipFunction(F))
     return false;
@@ -164,6 +213,7 @@ bool RISCVCodeGenPrepare::runOnFunction(Function &F) {
   ST = &TM.getSubtarget<RISCVSubtarget>(F);
 
   DL = &F.getDataLayout();
+  DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
 
   bool MadeChange = false;
   for (auto &BB : F)
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
index 5e64e9fbc1a2f..cbde3e9f84ce0 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
@@ -1,10 +1,10 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+m,+d,+zfh,+v,+zvfh \
-; RUN:   -verify-machineinstrs < %s \
-; RUN:   | FileCheck %s --check-prefixes=CHECK,CHECK-RV32
+; RUN:     -verify-machineinstrs < %s \
+; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-RV32
 ; RUN: llc -mtriple=riscv64 -mattr=+m,+d,+zfh,+v,+zvfh \
-; RUN:   -verify-machineinstrs < %s \
-; RUN:   | FileCheck %s --check-prefixes=CHECK,CHECK-RV64
+; RUN:     -verify-machineinstrs < %s | FileCheck %s \
+; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-RV64
 
 declare <2 x i8> @llvm.experimental.vp.strided.load.v2i8.p0.i8(ptr, i8, <2 x i1>, i32)
 
@@ -626,3 +626,29 @@ define <33 x double> @strided_load_v33f64(ptr %ptr, i64 %stride, <33 x i1> %mask
 }
 
 declare <33 x double> @llvm.experimental.vp.strided.load.v33f64.p0.i64(ptr, i64, <33 x i1>, i32)
+
+; TODO: Use accurate evl.
+; Test unmasked integer zero strided
+define <4 x i8> @zero_strided_unmasked_vpload_4i8_i8(ptr %ptr) {
+; CHECK-LABEL: zero_strided_unmasked_vpload_4i8_i8:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    lbu a0, 0(a0)
+; CHECK-NEXT:    vsetivli zero, 4, e8, mf4, ta, ma
+; CHECK-NEXT:    vmv.v.x v8, a0
+; CHECK-NEXT:    ret
+  %load = call <4 x i8> @llvm.experimental.vp.strided.load.4i8.p0.i8(ptr %ptr, i8 0, <4 x i1> splat (i1 true), i32 3)
+  ret <4 x i8> %load
+}
+
+; TODO: Use accurate evl.
+; Test unmasked float zero strided
+define <4 x half> @zero_strided_unmasked_vpload_4f16(ptr %ptr) {
+; CHECK-LABEL: zero_strided_unmasked_vpload_4f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    flh fa5, 0(a0)
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vfmv.v.f v8, fa5
+; CHECK-NEXT:    ret
+  %load = call <4 x half> @llvm.experimental.vp.strided.load.4f16.p0.i32(ptr %ptr, i32 0, <4 x i1> splat (i1 true), i32 3)
+  ret <4 x half> %load
+}
diff --git a/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll b/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
index 4d3bced0bcb50..33e16a3c09410 100644
--- a/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
@@ -780,3 +780,27 @@ define <vscale x 16 x double> @strided_load_nxv17f64(ptr %ptr, i64 %stride, <vsc
 declare <vscale x 17 x double> @llvm.experimental.vp.strided.load.nxv17f64.p0.i64(ptr, i64, <vscale x 17 x i1>, i32)
 declare <vscale x 1 x double> @llvm.experimental.vector.extract.nxv1f64(<vscale x 17 x double> %vec, i64 %idx)
 declare <vscale x 16 x double> @llvm.experimental.vector.extract.nxv16f64(<vscale x 17 x double> %vec, i64 %idx)
+
+; Test unmasked integer zero strided
+define <vscale x 1 x i8> @zero_strided_unmasked_vpload_nxv1i8_i8(ptr %ptr) {
+; CHECK-LABEL: zero_strided_unmasked_vpload_nxv1i8_i8:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    lbu a0, 0(a0)
+; CHECK-NEXT:    vsetivli zero, 4, e8, mf8, ta, ma
+; CHECK-NEXT:    vmv.v.x v8, a0
+; CHECK-NEXT:    ret
+  %load = call <vscale x 1 x i8> @llvm.experimental.vp.strided.load.nxv1i8.p0.i8(ptr %ptr, i8 0, <vscale x 1 x i1> splat (i1 true), i32 4)
+  ret <vscale x 1 x i8> %load
+}
+
+; Test unmasked float zero strided
+define <vscale x 1 x half> @zero_strided_unmasked_vpload_nxv1f16(ptr %ptr) {
+; CHECK-LABEL: zero_strided_unmasked_vpload_nxv1f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    flh fa5, 0(a0)
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf4, ta, ma
+; CHECK-NEXT:    vfmv.v.f v8, fa5
+; CHECK-NEXT:    ret
+  %load = call <vscale x 1 x half> @llvm.experimental.vp.strided.load.nxv1f16.p0.i32(ptr %ptr, i32 0, <vscale x 1 x i1> splat (i1 true), i32 4)
+  ret <vscale x 1 x half> %load
+}

[lukel97]

Should we still add a RUN line with optimized-zero-stride-load to ensure that it gets 'unexpanded' back into a zero strided load during instruction selection?

[yetingk]

I add the test and I also add the hasOptimizedZeroStrideLoad check back.

[lukel97]

LGTM