[llvm] b9b7162 - [InstCombine] `vector_reduce_umax(?ext(<n x i1>))` --> `?ext(vector_reduce_or(<n x i1>))` (PR51259)
Roman Lebedev via llvm-commits
llvm-commits at lists.llvm.org
Mon Aug 2 13:02:46 PDT 2021
Author: Roman Lebedev
Date: 2021-08-02T23:02:23+03:00
New Revision: b9b7162b8bdc97ba5b4219fe3447fadc48b4a107
URL: https://github.com/llvm/llvm-project/commit/b9b7162b8bdc97ba5b4219fe3447fadc48b4a107
DIFF: https://github.com/llvm/llvm-project/commit/b9b7162b8bdc97ba5b4219fe3447fadc48b4a107.diff
LOG: [InstCombine] `vector_reduce_umax(?ext(<n x i1>))` --> `?ext(vector_reduce_or(<n x i1>))` (PR51259)
Alive2 agrees:
https://alive2.llvm.org/ce/z/NbBaeT (self)
https://alive2.llvm.org/ce/z/iEaig4 (zext)
https://alive2.llvm.org/ce/z/meGb3y (sext)
We already handle `vector_reduce_and(<n x i1>)`,
so let's just combine into the already-handled pattern
and let the existing fold do the rest.
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
llvm/test/Transforms/InstCombine/reduction-umax-sext-zext-i1.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 3b39000e2b38..1136918b73fa 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -2067,20 +2067,24 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
}
LLVM_FALLTHROUGH;
}
- case Intrinsic::vector_reduce_umin: {
- if (IID == Intrinsic::vector_reduce_umin) {
- // UMin reduction over the vector with (potentially-extended)
+ case Intrinsic::vector_reduce_umin:
+ case Intrinsic::vector_reduce_umax: {
+ if (IID == Intrinsic::vector_reduce_umin ||
+ IID == Intrinsic::vector_reduce_umax) {
+ // UMin/UMax reduction over the vector with (potentially-extended)
// i1 element type is actually a (potentially-extended)
- // logical `and` reduction over the original non-extended value:
- // vector_reduce_umin(?ext(<n x i1>))
+ // logical `and`/`or` reduction over the original non-extended value:
+ // vector_reduce_u{min,max}(?ext(<n x i1>))
// -->
- // ?ext(vector_reduce_and(<n x i1>))
+ // ?ext(vector_reduce_{and,or}(<n x i1>))
Value *Arg = II->getArgOperand(0);
Value *Vect;
if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
if (FTy->getElementType() == Builder.getInt1Ty()) {
- Value *Res = Builder.CreateAndReduce(Vect);
+ Value *Res = IID == Intrinsic::vector_reduce_umin
+ ? Builder.CreateAndReduce(Vect)
+ : Builder.CreateOrReduce(Vect);
if (Arg != Vect)
Res = Builder.CreateCast(cast<CastInst>(Arg)->getOpcode(), Res,
II->getType());
@@ -2090,7 +2094,6 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
}
LLVM_FALLTHROUGH;
}
- case Intrinsic::vector_reduce_umax:
case Intrinsic::vector_reduce_smax:
case Intrinsic::vector_reduce_smin:
case Intrinsic::vector_reduce_fmax:
diff --git a/llvm/test/Transforms/InstCombine/reduction-umax-sext-zext-i1.ll b/llvm/test/Transforms/InstCombine/reduction-umax-sext-zext-i1.ll
index 07c210a57cc7..daafab864042 100644
--- a/llvm/test/Transforms/InstCombine/reduction-umax-sext-zext-i1.ll
+++ b/llvm/test/Transforms/InstCombine/reduction-umax-sext-zext-i1.ll
@@ -3,8 +3,9 @@
define i1 @reduce_umax_self(<8 x i1> %x) {
; CHECK-LABEL: @reduce_umax_self(
-; CHECK-NEXT: [[RES:%.*]] = call i1 @llvm.vector.reduce.umax.v8i1(<8 x i1> [[X:%.*]])
-; CHECK-NEXT: ret i1 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
+; CHECK-NEXT: ret i1 [[TMP2]]
;
%res = call i1 @llvm.vector.reduce.umax.v8i32(<8 x i1> %x)
ret i1 %res
@@ -12,9 +13,10 @@ define i1 @reduce_umax_self(<8 x i1> %x) {
define i32 @reduce_umax_sext(<4 x i1> %x) {
; CHECK-LABEL: @reduce_umax_sext(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <4 x i1> [[X:%.*]] to <4 x i32>
-; CHECK-NEXT: [[RES:%.*]] = call i32 @llvm.vector.reduce.umax.v4i32(<4 x i32> [[SEXT]])
-; CHECK-NEXT: ret i32 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <4 x i1> [[X:%.*]] to i4
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i4 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = sext i1 [[TMP2]] to i32
+; CHECK-NEXT: ret i32 [[TMP3]]
;
%sext = sext <4 x i1> %x to <4 x i32>
%res = call i32 @llvm.vector.reduce.umax.v4i32(<4 x i32> %sext)
@@ -23,9 +25,10 @@ define i32 @reduce_umax_sext(<4 x i1> %x) {
define i64 @reduce_umax_zext(<8 x i1> %x) {
; CHECK-LABEL: @reduce_umax_zext(
-; CHECK-NEXT: [[ZEXT:%.*]] = zext <8 x i1> [[X:%.*]] to <8 x i64>
-; CHECK-NEXT: [[RES:%.*]] = call i64 @llvm.vector.reduce.umax.v8i64(<8 x i64> [[ZEXT]])
-; CHECK-NEXT: ret i64 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = zext i1 [[TMP2]] to i64
+; CHECK-NEXT: ret i64 [[TMP3]]
;
%zext = zext <8 x i1> %x to <8 x i64>
%res = call i64 @llvm.vector.reduce.umax.v8i64(<8 x i64> %zext)
@@ -34,9 +37,10 @@ define i64 @reduce_umax_zext(<8 x i1> %x) {
define i16 @reduce_umax_sext_same(<16 x i1> %x) {
; CHECK-LABEL: @reduce_umax_sext_same(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <16 x i1> [[X:%.*]] to <16 x i16>
-; CHECK-NEXT: [[RES:%.*]] = call i16 @llvm.vector.reduce.umax.v16i16(<16 x i16> [[SEXT]])
-; CHECK-NEXT: ret i16 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <16 x i1> [[X:%.*]] to i16
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i16 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = sext i1 [[TMP2]] to i16
+; CHECK-NEXT: ret i16 [[TMP3]]
;
%sext = sext <16 x i1> %x to <16 x i16>
%res = call i16 @llvm.vector.reduce.umax.v16i16(<16 x i16> %sext)
@@ -45,9 +49,10 @@ define i16 @reduce_umax_sext_same(<16 x i1> %x) {
define i8 @reduce_umax_zext_long(<128 x i1> %x) {
; CHECK-LABEL: @reduce_umax_zext_long(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <128 x i1> [[X:%.*]] to <128 x i8>
-; CHECK-NEXT: [[RES:%.*]] = call i8 @llvm.vector.reduce.umax.v128i8(<128 x i8> [[SEXT]])
-; CHECK-NEXT: ret i8 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <128 x i1> [[X:%.*]] to i128
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i128 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = sext i1 [[TMP2]] to i8
+; CHECK-NEXT: ret i8 [[TMP3]]
;
%sext = sext <128 x i1> %x to <128 x i8>
%res = call i8 @llvm.vector.reduce.umax.v128i8(<128 x i8> %sext)
@@ -57,11 +62,13 @@ define i8 @reduce_umax_zext_long(<128 x i1> %x) {
@glob = external global i8, align 1
define i8 @reduce_umax_zext_long_external_use(<128 x i1> %x) {
; CHECK-LABEL: @reduce_umax_zext_long_external_use(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <128 x i1> [[X:%.*]] to <128 x i8>
-; CHECK-NEXT: [[RES:%.*]] = call i8 @llvm.vector.reduce.umax.v128i8(<128 x i8> [[SEXT]])
-; CHECK-NEXT: [[EXT:%.*]] = extractelement <128 x i8> [[SEXT]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <128 x i1> [[X:%.*]] to i128
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i128 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = sext i1 [[TMP2]] to i8
+; CHECK-NEXT: [[TMP4:%.*]] = extractelement <128 x i1> [[X]], i32 0
+; CHECK-NEXT: [[EXT:%.*]] = sext i1 [[TMP4]] to i8
; CHECK-NEXT: store i8 [[EXT]], i8* @glob, align 1
-; CHECK-NEXT: ret i8 [[RES]]
+; CHECK-NEXT: ret i8 [[TMP3]]
;
%sext = sext <128 x i1> %x to <128 x i8>
%res = call i8 @llvm.vector.reduce.umax.v128i8(<128 x i8> %sext)
@@ -73,11 +80,13 @@ define i8 @reduce_umax_zext_long_external_use(<128 x i1> %x) {
@glob1 = external global i64, align 8
define i64 @reduce_umax_zext_external_use(<8 x i1> %x) {
; CHECK-LABEL: @reduce_umax_zext_external_use(
-; CHECK-NEXT: [[ZEXT:%.*]] = zext <8 x i1> [[X:%.*]] to <8 x i64>
-; CHECK-NEXT: [[RES:%.*]] = call i64 @llvm.vector.reduce.umax.v8i64(<8 x i64> [[ZEXT]])
-; CHECK-NEXT: [[EXT:%.*]] = extractelement <8 x i64> [[ZEXT]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
+; CHECK-NEXT: [[TMP3:%.*]] = zext i1 [[TMP2]] to i64
+; CHECK-NEXT: [[TMP4:%.*]] = extractelement <8 x i1> [[X]], i32 0
+; CHECK-NEXT: [[EXT:%.*]] = zext i1 [[TMP4]] to i64
; CHECK-NEXT: store i64 [[EXT]], i64* @glob1, align 8
-; CHECK-NEXT: ret i64 [[RES]]
+; CHECK-NEXT: ret i64 [[TMP3]]
;
%zext = zext <8 x i1> %x to <8 x i64>
%res = call i64 @llvm.vector.reduce.umax.v8i64(<8 x i64> %zext)
More information about the llvm-commits
mailing list