[InstCombine] Pull vector reverse through intrinsics

[lukel97]

This is the intrinsic version of #146349, and handles fabs as well as other intrinsics.

It's largely a copy of InstCombinerImpl::foldShuffledIntrinsicOperands but a bit simpler since we don't need to find a common mask.

Creating a separate function seems to be cleaner than trying to shoehorn it into the existing one.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Luke Lau (lukel97)

Changes

This is the intrinsic version of #146349, and handles fabs as well as other intrinsics.

It's largely a copy of InstCombinerImpl::foldShuffledIntrinsicOperands but a bit simpler since we don't need to find a common mask.

Creating a separate function seems to be cleaner than trying to shoehorn it into the existing one.

---

Full diff: https://github.com/llvm/llvm-project/pull/146384.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp (+42)
• (modified) llvm/lib/Transforms/InstCombine/InstCombineInternal.h (+1)
• (modified) llvm/test/Transforms/InstCombine/vector-reverse.ll (+133)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index e33d111167c04..9d07a4531d0a3 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -1456,6 +1456,45 @@ InstCombinerImpl::foldShuffledIntrinsicOperands(IntrinsicInst *II) {
   return new ShuffleVectorInst(NewIntrinsic, Mask);
 }
 
+/// If all arguments of the intrinsic are reverses, try to pull the reverse
+/// after the intrinsic.
+Value *InstCombinerImpl::foldReversedIntrinsicOperands(IntrinsicInst *II) {
+  if (!isTriviallyVectorizable(II->getIntrinsicID()) ||
+      !II->getCalledFunction()->isSpeculatable())
+    return nullptr;
+
+  // At least 1 operand must be a reverse with 1 use because we are creating 2
+  // instructions.
+  if (none_of(II->args(), [](Value *V) {
+        return match(V, m_OneUse(m_VecReverse(m_Value())));
+      }))
+    return nullptr;
+
+  Value *X;
+  Constant *C;
+  SmallVector<Value *> NewArgs;
+  for (Use &Arg : II->args()) {
+    if (isVectorIntrinsicWithScalarOpAtArg(II->getIntrinsicID(),
+                                           Arg.getOperandNo(), nullptr))
+      NewArgs.push_back(Arg);
+    else if (match(&Arg, m_VecReverse(m_Value(X))))
+      NewArgs.push_back(X);
+    else if (Value *Splat = getSplatValue(Arg))
+      NewArgs.push_back(Builder.CreateVectorSplat(
+          cast<VectorType>(Arg->getType())->getElementCount(), Splat));
+    else if (match(&Arg, m_ImmConstant(C)))
+      NewArgs.push_back(Builder.CreateVectorReverse(C));
+    else
+      return nullptr;
+  }
+
+  // intrinsic (reverse X), (reverse Y), ... --> reverse (intrinsic X, Y, ...)
+  Instruction *FPI = isa<FPMathOperator>(II) ? II : nullptr;
+  Instruction *NewIntrinsic = Builder.CreateIntrinsic(
+      II->getType(), II->getIntrinsicID(), NewArgs, FPI);
+  return Builder.CreateVectorReverse(NewIntrinsic);
+}
+
 /// Fold the following cases and accepts bswap and bitreverse intrinsics:
 ///   bswap(logic_op(bswap(x), y)) --> logic_op(x, bswap(y))
 ///   bswap(logic_op(bswap(x), bswap(y))) --> logic_op(x, y) (ignores multiuse)
@@ -3867,6 +3906,9 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
   if (Instruction *Shuf = foldShuffledIntrinsicOperands(II))
     return Shuf;
 
+  if (Value *Reverse = foldReversedIntrinsicOperands(II))
+    return replaceInstUsesWith(*II, Reverse);
+
   // Some intrinsics (like experimental_gc_statepoint) can be used in invoke
   // context, so it is handled in visitCallBase and we should trigger it.
   return visitCallBase(*II);
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
index 8c9de862fe8f2..a8645521fe053 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -148,6 +148,7 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
   Instruction *foldItoFPtoI(CastInst &FI);
   Instruction *visitSelectInst(SelectInst &SI);
   Instruction *foldShuffledIntrinsicOperands(IntrinsicInst *II);
+  Value *foldReversedIntrinsicOperands(IntrinsicInst *II);
   Instruction *visitCallInst(CallInst &CI);
   Instruction *visitInvokeInst(InvokeInst &II);
   Instruction *visitCallBrInst(CallBrInst &CBI);
diff --git a/llvm/test/Transforms/InstCombine/vector-reverse.ll b/llvm/test/Transforms/InstCombine/vector-reverse.ll
index c9c68d2241b34..cbf6b37692cac 100644
--- a/llvm/test/Transforms/InstCombine/vector-reverse.ll
+++ b/llvm/test/Transforms/InstCombine/vector-reverse.ll
@@ -17,6 +17,18 @@ define <vscale x 4 x i32> @binop_reverse(<vscale x 4 x i32> %a, <vscale x 4 x i3
   ret <vscale x 4 x i32> %add
 }
 
+define <vscale x 4 x i32> @binop_intrinsic_reverse(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; CHECK-LABEL: @binop_intrinsic_reverse(
+; CHECK-NEXT:    [[ADD:%.*]] = call <vscale x 4 x i32> @llvm.smax.nxv4i32(<vscale x 4 x i32> [[A_REV:%.*]], <vscale x 4 x i32> [[B_REV:%.*]])
+; CHECK-NEXT:    [[SMAX:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[ADD]])
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[SMAX]]
+;
+  %a.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %a)
+  %b.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %b)
+  %smax = call <vscale x 4 x i32> @llvm.smax(<vscale x 4 x i32> %a.rev, <vscale x 4 x i32> %b.rev)
+  ret <vscale x 4 x i32> %smax
+}
+
 ; %a.rev has multiple uses
 define <vscale x 4 x i32> @binop_reverse_1(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
 ; CHECK-LABEL: @binop_reverse_1(
@@ -33,6 +45,22 @@ define <vscale x 4 x i32> @binop_reverse_1(<vscale x 4 x i32> %a, <vscale x 4 x
   ret <vscale x 4 x i32> %add
 }
 
+; %a.rev has multiple uses
+define <vscale x 4 x i32> @binop_intrinsic_reverse_1(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; CHECK-LABEL: @binop_intrinsic_reverse_1(
+; CHECK-NEXT:    [[B_REV:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[B:%.*]])
+; CHECK-NEXT:    call void @use_nxv4i32(<vscale x 4 x i32> [[B_REV]])
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 4 x i32> @llvm.smax.nxv4i32(<vscale x 4 x i32> [[B]], <vscale x 4 x i32> [[B1:%.*]])
+; CHECK-NEXT:    [[SMAX:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP1]])
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[SMAX]]
+;
+  %a.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %a)
+  %b.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %b)
+  call void @use_nxv4i32(<vscale x 4 x i32>  %a.rev)
+  %smax = call <vscale x 4 x i32> @llvm.smax(<vscale x 4 x i32> %a.rev, <vscale x 4 x i32> %b.rev)
+  ret <vscale x 4 x i32> %smax
+}
+
 ; %b.rev has multiple uses
 define <vscale x 4 x i32> @binop_reverse_2(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
 ; CHECK-LABEL: @binop_reverse_2(
@@ -67,6 +95,24 @@ define <vscale x 4 x i32> @binop_reverse_3(<vscale x 4 x i32> %a, <vscale x 4 x
   ret <vscale x 4 x i32> %add
 }
 
+; %a.rev and %b.rev have multiple uses
+define <vscale x 4 x i32> @binop_intrinsic_reverse_3(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; CHECK-LABEL: @binop_intrinsic_reverse_3(
+; CHECK-NEXT:    [[A_REV:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[A:%.*]])
+; CHECK-NEXT:    [[B_REV:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[B:%.*]])
+; CHECK-NEXT:    call void @use_nxv4i32(<vscale x 4 x i32> [[A_REV]])
+; CHECK-NEXT:    call void @use_nxv4i32(<vscale x 4 x i32> [[B_REV]])
+; CHECK-NEXT:    [[SMAX:%.*]] = call <vscale x 4 x i32> @llvm.smax.nxv4i32(<vscale x 4 x i32> [[A_REV]], <vscale x 4 x i32> [[B_REV]])
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[SMAX]]
+;
+  %a.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %a)
+  %b.rev = tail call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> %b)
+  call void @use_nxv4i32(<vscale x 4 x i32> %a.rev)
+  call void @use_nxv4i32(<vscale x 4 x i32> %b.rev)
+  %smax = call <vscale x 4 x i32> @llvm.smax(<vscale x 4 x i32> %a.rev, <vscale x 4 x i32> %b.rev)
+  ret <vscale x 4 x i32> %smax
+}
+
 ; %a.rev used as both operands
 define <vscale x 4 x i32> @binop_reverse_4(<vscale x 4 x i32> %a) {
 ; CHECK-LABEL: @binop_reverse_4(
@@ -184,6 +230,17 @@ define <vscale x 4 x float> @unop_reverse_1(<vscale x 4 x float> %a) {
   ret <vscale x 4 x float> %neg
 }
 
+define <vscale x 4 x float> @unop_intrinsic_reverse(<vscale x 4 x float> %a) {
+; CHECK-LABEL: @unop_intrinsic_reverse(
+; CHECK-NEXT:    [[NEG:%.*]] = call <vscale x 4 x float> @llvm.fabs.nxv4f32(<vscale x 4 x float> [[A_REV:%.*]])
+; CHECK-NEXT:    [[ABS:%.*]] = call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> [[NEG]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[ABS]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %abs = call <vscale x 4 x float> @llvm.fabs(<vscale x 4 x float> %a.rev)
+  ret <vscale x 4 x float> %abs
+}
+
 define <vscale x 4 x i1> @icmp_reverse(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
 ; CHECK-LABEL: @icmp_reverse(
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq <vscale x 4 x i32> [[A:%.*]], [[B:%.*]]
@@ -629,6 +686,18 @@ define <vscale x 4 x float> @reverse_binop_reverse(<vscale x 4 x float> %a, <vsc
   ret <vscale x 4 x float> %add.rev
 }
 
+define <vscale x 4 x float> @reverse_binop_intrinsic_reverse(<vscale x 4 x float> %a, <vscale x 4 x float> %b) {
+; CHECK-LABEL: @reverse_binop_intrinsic_reverse(
+; CHECK-NEXT:    [[ADD:%.*]] = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> [[A_REV:%.*]], <vscale x 4 x float> [[B_REV:%.*]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[ADD]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %b.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %b)
+  %maxnum = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> %a.rev, <vscale x 4 x float> %b.rev)
+  %maxnum.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %maxnum)
+  ret <vscale x 4 x float> %maxnum.rev
+}
+
 define <vscale x 4 x float> @reverse_binop_reverse_splat_RHS(<vscale x 4 x float> %a, float %b) {
 ; CHECK-LABEL: @reverse_binop_reverse_splat_RHS(
 ; CHECK-NEXT:    [[B_INSERT:%.*]] = insertelement <vscale x 4 x float> poison, float [[B:%.*]], i64 0
@@ -659,6 +728,49 @@ define <vscale x 4 x float> @reverse_binop_reverse_splat_LHS(<vscale x 4 x float
   ret <vscale x 4 x float> %div.rev
 }
 
+define <vscale x 4 x float> @reverse_binop_reverse_intrinsic_splat_RHS(<vscale x 4 x float> %a, float %b) {
+; CHECK-LABEL: @reverse_binop_reverse_intrinsic_splat_RHS(
+; CHECK-NEXT:    [[B_INSERT:%.*]] = insertelement <vscale x 4 x float> poison, float [[B:%.*]], i64 0
+; CHECK-NEXT:    [[B_SPLAT:%.*]] = shufflevector <vscale x 4 x float> [[B_INSERT]], <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[MAXNUM:%.*]] = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> [[A_REV:%.*]], <vscale x 4 x float> [[B_SPLAT]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[MAXNUM]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %b.insert = insertelement <vscale x 4 x float> poison, float %b, i32 0
+  %b.splat = shufflevector <vscale x 4 x float> %b.insert, <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer
+  %maxnum = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> %a.rev, <vscale x 4 x float> %b.splat)
+  %maxnum.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %maxnum)
+  ret <vscale x 4 x float> %maxnum.rev
+}
+
+define <vscale x 4 x float> @reverse_binop_reverse_intrinsic_splat_LHS(<vscale x 4 x float> %a, float %b) {
+; CHECK-LABEL: @reverse_binop_reverse_intrinsic_splat_LHS(
+; CHECK-NEXT:    [[B_INSERT:%.*]] = insertelement <vscale x 4 x float> poison, float [[B:%.*]], i64 0
+; CHECK-NEXT:    [[B_SPLAT:%.*]] = shufflevector <vscale x 4 x float> [[B_INSERT]], <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[MAXNUM:%.*]] = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> [[B_SPLAT]], <vscale x 4 x float> [[A_REV:%.*]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[MAXNUM]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %b.insert = insertelement <vscale x 4 x float> poison, float %b, i32 0
+  %b.splat = shufflevector <vscale x 4 x float> %b.insert, <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer
+  %maxnum = call <vscale x 4 x float> @llvm.maxnum.nxv4f32(<vscale x 4 x float> %b.splat, <vscale x 4 x float> %a.rev)
+  %maxnum.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %maxnum)
+  ret <vscale x 4 x float> %maxnum.rev
+}
+
+define <4 x float> @reverse_binop_reverse_intrinsic_constant_RHS(<4 x float> %a) {
+; CHECK-LABEL: @reverse_binop_reverse_intrinsic_constant_RHS(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[A:%.*]], <4 x float> <float 3.000000e+00, float 2.000000e+00, float 1.000000e+00, float 0.000000e+00>)
+; CHECK-NEXT:    [[MAXNUM:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT:    [[MAXNUM_REV:%.*]] = tail call <4 x float> @llvm.vector.reverse.v4f32(<4 x float> [[MAXNUM]])
+; CHECK-NEXT:    ret <4 x float> [[MAXNUM_REV]]
+;
+  %a.rev = tail call <4 x float> @llvm.vector.reverse(<4 x float> %a)
+  %maxnum = call <4 x float> @llvm.maxnum.v4f32(<4 x float> <float 0.0, float 1.0, float 2.0, float 3.0>, <4 x float> %a.rev)
+  %maxnum.rev = tail call <4 x float> @llvm.vector.reverse(<4 x float> %maxnum)
+  ret <4 x float> %maxnum.rev
+}
+
 define <vscale x 4 x i1> @reverse_fcmp_reverse(<vscale x 4 x float> %a, <vscale x 4 x float> %b) {
 ; CHECK-LABEL: @reverse_fcmp_reverse(
 ; CHECK-NEXT:    [[CMP1:%.*]] = fcmp fast olt <vscale x 4 x float> [[A:%.*]], [[B:%.*]]
@@ -695,6 +807,27 @@ define <vscale x 4 x float> @reverse_unop_reverse(<vscale x 4 x float> %a) {
   ret <vscale x 4 x float> %neg.rev
 }
 
+define <vscale x 4 x float> @reverse_unop_intrinsic_reverse(<vscale x 4 x float> %a) {
+; CHECK-LABEL: @reverse_unop_intrinsic_reverse(
+; CHECK-NEXT:    [[ABS:%.*]] = call <vscale x 4 x float> @llvm.fabs.nxv4f32(<vscale x 4 x float> [[A_REV:%.*]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[ABS]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %abs = call <vscale x 4 x float> @llvm.fabs(<vscale x 4 x float> %a.rev)
+  %abs.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %abs)
+  ret <vscale x 4 x float> %abs.rev
+}
+
+define <vscale x 4 x float> @reverse_unop_intrinsic_reverse_scalar_arg(<vscale x 4 x float> %a, i32 %power) {
+; CHECK-LABEL: @reverse_unop_intrinsic_reverse_scalar_arg(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 4 x float> @llvm.powi.nxv4f32.i32(<vscale x 4 x float> [[A:%.*]], i32 [[POWER:%.*]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[TMP1]]
+;
+  %a.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %a)
+  %powi = call <vscale x 4 x float> @llvm.powi.nxv4f32(<vscale x 4 x float> %a.rev, i32 %power)
+  %powi.rev = tail call <vscale x 4 x float> @llvm.vector.reverse.nxv4f32(<vscale x 4 x float> %powi)
+  ret <vscale x 4 x float> %powi.rev
+}
 
 declare void @use_nxv4i1(<vscale x 4 x i1>)
 declare void @use_nxv4i32(<vscale x 4 x i32>)

[preames]

Do we need to speculate it? The reverse isn't changing which lanes are active.

[preames]

Hm, since we're inserting a new splat, do we need the existing one to be one-use? You do need to insert a new one instead of reusing because of possible undef/poison lanes.

[lukel97]

I looked at the existing binop combine to see what it was doing and it actually doesn't create new splats, it instead checks isSplatValue which doesn't allow undef lanes:

  if (match(LHS, m_VecReverse(m_Value(V1)))) {
    // Op(rev(V1), rev(V2)) -> rev(Op(V1, V2))
    if (match(RHS, m_VecReverse(m_Value(V2))) &&
        (LHS->hasOneUse() || RHS->hasOneUse() ||
         (LHS == RHS && LHS->hasNUses(2))))
      return createBinOpReverse(V1, V2);

    // Op(rev(V1), RHSSplat)) -> rev(Op(V1, RHSSplat))
    if (LHS->hasOneUse() && isSplatValue(RHS))
      return createBinOpReverse(V1, RHS);
  }
  // Op(LHSSplat, rev(V2)) -> rev(Op(LHSSplat, V2))
  else if (isSplatValue(LHS) && match(RHS, m_OneUse(m_VecReverse(m_Value(V2)))))
    return createBinOpReverse(LHS, V2);

So I've done the same here in e0acdf3 to keep it consistent, disallowing poison lanes. I don't think supporting poison lanes is important given that reverse intrinsics are really only used for scalable vectors.

[preames]

LGTM