[llvm] 469793e - [InstCombine] `vector_reduce_mul(?ext(<n x i1>))` --> `zext(vector_reduce_and(<n x i1>))` (PR51259)
Roman Lebedev via llvm-commits
llvm-commits at lists.llvm.org
Mon Aug 2 12:03:30 PDT 2021
Author: Roman Lebedev
Date: 2021-08-02T21:57:51+03:00
New Revision: 469793efa7b518e3b5a5bd68ed5e39a9f9f4121f
URL: https://github.com/llvm/llvm-project/commit/469793efa7b518e3b5a5bd68ed5e39a9f9f4121f
DIFF: https://github.com/llvm/llvm-project/commit/469793efa7b518e3b5a5bd68ed5e39a9f9f4121f.diff
LOG: [InstCombine] `vector_reduce_mul(?ext(<n x i1>))` --> `zext(vector_reduce_and(<n x i1>))` (PR51259)
Alive2 agrees:
https://alive2.llvm.org/ce/z/PDansB (self)
https://alive2.llvm.org/ce/z/55D-Xc (zext)
https://alive2.llvm.org/ce/z/LxG3-r (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-mul-sext-zext-i1.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 1ada23b979a0..5636b5d1529a 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -2045,7 +2045,28 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
}
LLVM_FALLTHROUGH;
}
- case Intrinsic::vector_reduce_mul:
+ case Intrinsic::vector_reduce_mul: {
+ if (IID == Intrinsic::vector_reduce_mul) {
+ // Multiplicative reduction over the vector with (potentially-extended)
+ // i1 element type is actually a (potentially zero-extended)
+ // logical `and` reduction over the original non-extended value:
+ // vector_reduce_mul(?ext(<n x i1>))
+ // -->
+ // zext(vector_reduce_and(<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);
+ if (Res->getType() != II->getType())
+ Res = Builder.CreateZExt(Res, II->getType());
+ return replaceInstUsesWith(CI, Res);
+ }
+ }
+ }
+ LLVM_FALLTHROUGH;
+ }
case Intrinsic::vector_reduce_umax:
case Intrinsic::vector_reduce_umin:
case Intrinsic::vector_reduce_smax:
diff --git a/llvm/test/Transforms/InstCombine/reduction-mul-sext-zext-i1.ll b/llvm/test/Transforms/InstCombine/reduction-mul-sext-zext-i1.ll
index 82b4f66c7905..1d5e05347e5b 100644
--- a/llvm/test/Transforms/InstCombine/reduction-mul-sext-zext-i1.ll
+++ b/llvm/test/Transforms/InstCombine/reduction-mul-sext-zext-i1.ll
@@ -3,8 +3,9 @@
define i1 @reduce_mul_self(<8 x i1> %x) {
; CHECK-LABEL: @reduce_mul_self(
-; CHECK-NEXT: [[RES:%.*]] = call i1 @llvm.vector.reduce.mul.v8i1(<8 x i1> [[X:%.*]])
-; CHECK-NEXT: ret i1 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i8 [[TMP1]], -1
+; CHECK-NEXT: ret i1 [[TMP2]]
;
%res = call i1 @llvm.vector.reduce.mul.v8i32(<8 x i1> %x)
ret i1 %res
@@ -12,9 +13,10 @@ define i1 @reduce_mul_self(<8 x i1> %x) {
define i32 @reduce_mul_sext(<4 x i1> %x) {
; CHECK-LABEL: @reduce_mul_sext(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <4 x i1> [[X:%.*]] to <4 x i32>
-; CHECK-NEXT: [[RES:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[SEXT]])
-; CHECK-NEXT: ret i32 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <4 x i1> [[X:%.*]] to i4
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i4 [[TMP1]], -1
+; CHECK-NEXT: [[TMP3:%.*]] = zext 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.mul.v4i32(<4 x i32> %sext)
@@ -23,9 +25,10 @@ define i32 @reduce_mul_sext(<4 x i1> %x) {
define i64 @reduce_mul_zext(<8 x i1> %x) {
; CHECK-LABEL: @reduce_mul_zext(
-; CHECK-NEXT: [[ZEXT:%.*]] = zext <8 x i1> [[X:%.*]] to <8 x i64>
-; CHECK-NEXT: [[RES:%.*]] = call i64 @llvm.vector.reduce.mul.v8i64(<8 x i64> [[ZEXT]])
-; CHECK-NEXT: ret i64 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i8 [[TMP1]], -1
+; 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.mul.v8i64(<8 x i64> %zext)
@@ -34,9 +37,10 @@ define i64 @reduce_mul_zext(<8 x i1> %x) {
define i16 @reduce_mul_sext_same(<16 x i1> %x) {
; CHECK-LABEL: @reduce_mul_sext_same(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <16 x i1> [[X:%.*]] to <16 x i16>
-; CHECK-NEXT: [[RES:%.*]] = call i16 @llvm.vector.reduce.mul.v16i16(<16 x i16> [[SEXT]])
-; CHECK-NEXT: ret i16 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <16 x i1> [[X:%.*]] to i16
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i16 [[TMP1]], -1
+; CHECK-NEXT: [[TMP3:%.*]] = zext 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.mul.v16i16(<16 x i16> %sext)
@@ -45,9 +49,10 @@ define i16 @reduce_mul_sext_same(<16 x i1> %x) {
define i8 @reduce_mul_zext_long(<128 x i1> %x) {
; CHECK-LABEL: @reduce_mul_zext_long(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <128 x i1> [[X:%.*]] to <128 x i8>
-; CHECK-NEXT: [[RES:%.*]] = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> [[SEXT]])
-; CHECK-NEXT: ret i8 [[RES]]
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast <128 x i1> [[X:%.*]] to i128
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i128 [[TMP1]], -1
+; CHECK-NEXT: [[TMP3:%.*]] = zext 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.mul.v128i8(<128 x i8> %sext)
@@ -57,11 +62,13 @@ define i8 @reduce_mul_zext_long(<128 x i1> %x) {
@glob = external global i8, align 1
define i8 @reduce_mul_zext_long_external_use(<128 x i1> %x) {
; CHECK-LABEL: @reduce_mul_zext_long_external_use(
-; CHECK-NEXT: [[SEXT:%.*]] = sext <128 x i1> [[X:%.*]] to <128 x i8>
-; CHECK-NEXT: [[RES:%.*]] = call i8 @llvm.vector.reduce.mul.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 eq i128 [[TMP1]], -1
+; CHECK-NEXT: [[TMP3:%.*]] = zext 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.mul.v128i8(<128 x i8> %sext)
@@ -73,11 +80,13 @@ define i8 @reduce_mul_zext_long_external_use(<128 x i1> %x) {
@glob1 = external global i64, align 8
define i64 @reduce_mul_zext_external_use(<8 x i1> %x) {
; CHECK-LABEL: @reduce_mul_zext_external_use(
-; CHECK-NEXT: [[ZEXT:%.*]] = zext <8 x i1> [[X:%.*]] to <8 x i64>
-; CHECK-NEXT: [[RES:%.*]] = call i64 @llvm.vector.reduce.mul.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 eq i8 [[TMP1]], -1
+; 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.mul.v8i64(<8 x i64> %zext)
More information about the llvm-commits
mailing list