[llvm] [InstCombine] recognize missed i128 split optimization (PR #129363)
Muhammad Bassiouni via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 5 09:25:48 PST 2025
https://github.com/bassiounix updated https://github.com/llvm/llvm-project/pull/129363
>From eb3a1f15d0ffee624ed87d9f496c0c3ed2f8e27f Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sat, 1 Mar 2025 07:24:38 +0200
Subject: [PATCH 1/8] [InstCombine] recognize missed i128 split optimization
---
llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 175c653f17f07..cff95338650f2 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3119,6 +3119,13 @@ static Value *matchOrConcat(Instruction &Or, InstCombiner::BuilderTy &Builder) {
match(UpperSrc, m_BitReverse(m_Value(UpperBRev))))
return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);
+ Value *X;
+ if (match(LowerSrc, m_SExt(m_Value(X))) &&
+ match(UpperSrc,
+ m_SExt(m_AShr(m_Specific(X), m_SpecificInt(HalfWidth / 2 - 1))))) {
+ return Builder.CreateSExt(X, Ty);
+ }
+
return nullptr;
}
>From 5e37f325ec097084868450d92285dcf2f252291a Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 00:26:51 +0200
Subject: [PATCH 2/8] fix: update pattern constant source
---
llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index cff95338650f2..b5a2c5df7b4ce 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3122,7 +3122,9 @@ static Value *matchOrConcat(Instruction &Or, InstCombiner::BuilderTy &Builder) {
Value *X;
if (match(LowerSrc, m_SExt(m_Value(X))) &&
match(UpperSrc,
- m_SExt(m_AShr(m_Specific(X), m_SpecificInt(HalfWidth / 2 - 1))))) {
+ m_SExt(m_AShr(
+ m_Specific(X),
+ m_SpecificInt(X->getType()->getScalarSizeInBits() - 1))))) {
return Builder.CreateSExt(X, Ty);
}
>From 899c8d559a388b6034bbbf3eb109516bcfc6863b Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 01:20:03 +0200
Subject: [PATCH 3/8] add(InstCombine): tests for the ext optimization
---
.../Transforms/InstCombine/i128-ext-split.ll | 42 +++++++++++++++++++
1 file changed, 42 insertions(+)
create mode 100644 llvm/test/Transforms/InstCombine/i128-ext-split.ll
diff --git a/llvm/test/Transforms/InstCombine/i128-ext-split.ll b/llvm/test/Transforms/InstCombine/i128-ext-split.ll
new file mode 100644
index 0000000000000..05d5deb8f7eb9
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/i128-ext-split.ll
@@ -0,0 +1,42 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+; PR129363
+
+define i128 @i128_ext_split(i32 noundef %x) {
+; CHECK-LABEL: define i128 @i128_ext_split(
+; CHECK-SAME: i32 noundef [[X:%.*]]) {
+; CHECK-NEXT: [[XX:%.*]] = sext i32 [[X]] to i128
+; CHECK-NEXT: ret i128 [[XX]]
+;
+ %coerce.sroa.0.0.extract.trunc = sext i32 %x to i64
+ %ashr = ashr i32 %x, 31
+ %coerce.sroa.2.0.extract.trunc = sext i32 %ashr to i64
+ %x.sroa.2.0.insert.ext.i = zext i64 %coerce.sroa.2.0.extract.trunc to i128
+ %x.sroa.2.0.insert.shift.i = shl nuw i128 %x.sroa.2.0.insert.ext.i, 64
+ %x.sroa.0.0.insert.ext.i = zext i64 %coerce.sroa.0.0.extract.trunc to i128
+ %x.sroa.0.0.insert.insert.i = or disjoint i128 %x.sroa.2.0.insert.shift.i, %x.sroa.0.0.insert.ext.i
+ ret i128 %x.sroa.0.0.insert.insert.i
+}
+
+define void @i128_ext_split_store(i32 %x, ptr %out) {
+; CHECK-LABEL: define void @i128_ext_split_store(
+; CHECK-SAME: i32 [[X:%.*]], ptr [[OUT:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[RES:%.*]] = sext i32 [[X]] to i128
+; CHECK-NEXT: store i128 [[RES]], ptr [[OUT]], align 16
+; CHECK-NEXT: ret void
+;
+entry:
+ %LowerSrc = sext i32 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = ashr i32 %x, 31
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+
+ %res = or disjoint i128 %hi, %lo
+ store i128 %res, ptr %out, align 16
+ ret void
+}
>From efd069d4cc4741f3327b10a7cf87a81406bfa85f Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 17:31:22 +0200
Subject: [PATCH 4/8] chore(test): rename test file
---
.../Transforms/InstCombine/{i128-ext-split.ll => iX-ext-split.ll} | 0
1 file changed, 0 insertions(+), 0 deletions(-)
rename llvm/test/Transforms/InstCombine/{i128-ext-split.ll => iX-ext-split.ll} (100%)
diff --git a/llvm/test/Transforms/InstCombine/i128-ext-split.ll b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
similarity index 100%
rename from llvm/test/Transforms/InstCombine/i128-ext-split.ll
rename to llvm/test/Transforms/InstCombine/iX-ext-split.ll
>From d8ef2ec9cc1a1a3168ab9781b62442508b223609 Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 17:42:23 +0200
Subject: [PATCH 5/8] chore(InstCombine): add more types to the test
---
.../Transforms/InstCombine/iX-ext-split.ll | 48 +++++++++++++++++++
1 file changed, 48 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/iX-ext-split.ll b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
index 05d5deb8f7eb9..67955fa30dee7 100644
--- a/llvm/test/Transforms/InstCombine/iX-ext-split.ll
+++ b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
@@ -3,6 +3,7 @@
; PR129363
+; ext split from i32 to i128
define i128 @i128_ext_split(i32 noundef %x) {
; CHECK-LABEL: define i128 @i128_ext_split(
; CHECK-SAME: i32 noundef [[X:%.*]]) {
@@ -19,6 +20,7 @@ define i128 @i128_ext_split(i32 noundef %x) {
ret i128 %x.sroa.0.0.insert.insert.i
}
+; ext split from i32 to i128
define void @i128_ext_split_store(i32 %x, ptr %out) {
; CHECK-LABEL: define void @i128_ext_split_store(
; CHECK-SAME: i32 [[X:%.*]], ptr [[OUT:%.*]]) {
@@ -40,3 +42,49 @@ entry:
store i128 %res, ptr %out, align 16
ret void
}
+
+; ext split from i16 to i64
+define void @i64_ext_split_store(i16 %x, ptr %out) {
+; CHECK-LABEL: define void @i64_ext_split_store(
+; CHECK-SAME: i16 [[X:%.*]], ptr [[OUT:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[RES:%.*]] = sext i16 [[X]] to i64
+; CHECK-NEXT: store i64 [[RES]], ptr [[OUT]], align 16
+; CHECK-NEXT: ret void
+;
+entry:
+ %LowerSrc = sext i16 %x to i32
+ %lo = zext i32 %LowerSrc to i64
+
+ %sign = ashr i16 %x, 15
+ %UpperSrc = sext i16 %sign to i32
+ %widen = zext i32 %UpperSrc to i64
+ %hi = shl nuw i64 %widen, 32
+
+ %res = or disjoint i64 %hi, %lo
+ store i64 %res, ptr %out, align 16
+ ret void
+}
+
+; ext split from i16 to i128
+define void @i128_ext_split_store_i16(i16 %x, ptr %out) {
+; CHECK-LABEL: define void @i128_ext_split_store_i16(
+; CHECK-SAME: i16 [[X:%.*]], ptr [[OUT:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[RES:%.*]] = sext i16 [[X]] to i128
+; CHECK-NEXT: store i128 [[RES]], ptr [[OUT]], align 16
+; CHECK-NEXT: ret void
+;
+entry:
+ %LowerSrc = sext i16 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = ashr i16 %x, 15
+ %UpperSrc = sext i16 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+
+ %res = or disjoint i128 %hi, %lo
+ store i128 %res, ptr %out, align 16
+ ret void
+}
>From cc1e8cce0814915506e3c6733f90384880c87047 Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 17:57:47 +0200
Subject: [PATCH 6/8] add(InstCombine): negative test for ext split
optimization
---
.../Transforms/InstCombine/iX-ext-split.ll | 106 ++++++++++++++++++
1 file changed, 106 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/iX-ext-split.ll b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
index 67955fa30dee7..4dab883065cc2 100644
--- a/llvm/test/Transforms/InstCombine/iX-ext-split.ll
+++ b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
@@ -88,3 +88,109 @@ entry:
store i128 %res, ptr %out, align 16
ret void
}
+
+; negative test - wrong constant value
+define i128 @i128_ext_split_neg1(i32 %x) {
+; CHECK-LABEL: define i128 @i128_ext_split_neg1(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LOWERSRC:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[LO:%.*]] = zext i64 [[LOWERSRC]] to i128
+; CHECK-NEXT: [[SIGN:%.*]] = ashr i32 [[X]], 31
+; CHECK-NEXT: [[UPPERSRC:%.*]] = sext i32 [[SIGN]] to i64
+; CHECK-NEXT: [[WIDEN:%.*]] = zext i64 [[UPPERSRC]] to i128
+; CHECK-NEXT: [[HI:%.*]] = shl nuw i128 [[WIDEN]], 65
+; CHECK-NEXT: [[RES:%.*]] = or disjoint i128 [[HI]], [[LO]]
+; CHECK-NEXT: ret i128 [[RES]]
+;
+entry:
+ %LowerSrc = sext i32 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = ashr i32 %x, 31
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 65
+
+ %res = or disjoint i128 %hi, %lo
+ ret i128 %res
+}
+
+; negative test - wrong shift value
+define i128 @i128_ext_split_neg2(i32 %x) {
+; CHECK-LABEL: define i128 @i128_ext_split_neg2(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LOWERSRC:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[LO:%.*]] = zext i64 [[LOWERSRC]] to i128
+; CHECK-NEXT: [[SIGN:%.*]] = ashr i32 [[X]], 3
+; CHECK-NEXT: [[UPPERSRC:%.*]] = sext i32 [[SIGN]] to i64
+; CHECK-NEXT: [[WIDEN:%.*]] = zext i64 [[UPPERSRC]] to i128
+; CHECK-NEXT: [[HI:%.*]] = shl nuw i128 [[WIDEN]], 64
+; CHECK-NEXT: [[RES:%.*]] = or disjoint i128 [[HI]], [[LO]]
+; CHECK-NEXT: ret i128 [[RES]]
+;
+entry:
+ %LowerSrc = sext i32 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = ashr i32 %x, 3
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+
+ %res = or disjoint i128 %hi, %lo
+ ret i128 %res
+}
+
+; negative test - wrong ext instruction
+define i128 @i128_ext_split_neg3(i32 %x) {
+; CHECK-LABEL: define i128 @i128_ext_split_neg3(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LO:%.*]] = zext i32 [[X]] to i128
+; CHECK-NEXT: [[SIGN:%.*]] = ashr i32 [[X]], 31
+; CHECK-NEXT: [[UPPERSRC:%.*]] = sext i32 [[SIGN]] to i64
+; CHECK-NEXT: [[WIDEN:%.*]] = zext i64 [[UPPERSRC]] to i128
+; CHECK-NEXT: [[HI:%.*]] = shl nuw i128 [[WIDEN]], 64
+; CHECK-NEXT: [[RES:%.*]] = or disjoint i128 [[HI]], [[LO]]
+; CHECK-NEXT: ret i128 [[RES]]
+;
+entry:
+ %LowerSrc = zext i32 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = ashr i32 %x, 31
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+
+ %res = or disjoint i128 %hi, %lo
+ ret i128 %res
+}
+
+; negative test - wrong shift
+define i128 @i128_ext_split_neg4(i32 %x) {
+; CHECK-LABEL: define i128 @i128_ext_split_neg4(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LOWERSRC:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[LO:%.*]] = zext i64 [[LOWERSRC]] to i128
+; CHECK-NEXT: [[SIGN:%.*]] = lshr i32 [[X]], 31
+; CHECK-NEXT: [[WIDEN:%.*]] = zext nneg i32 [[SIGN]] to i128
+; CHECK-NEXT: [[HI:%.*]] = shl nuw nsw i128 [[WIDEN]], 64
+; CHECK-NEXT: [[RES:%.*]] = or disjoint i128 [[HI]], [[LO]]
+; CHECK-NEXT: ret i128 [[RES]]
+;
+entry:
+ %LowerSrc = sext i32 %x to i64
+ %lo = zext i64 %LowerSrc to i128
+
+ %sign = lshr i32 %x, 31
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+
+ %res = or disjoint i128 %hi, %lo
+ ret i128 %res
+}
>From 4e423c631fc04a1e549aaf68eb3d9f283dac68ba Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Sun, 2 Mar 2025 19:51:29 +0200
Subject: [PATCH 7/8] fix(style): change variable names to be readable
---
llvm/test/Transforms/InstCombine/iX-ext-split.ll | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/llvm/test/Transforms/InstCombine/iX-ext-split.ll b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
index 4dab883065cc2..b88b62aa04375 100644
--- a/llvm/test/Transforms/InstCombine/iX-ext-split.ll
+++ b/llvm/test/Transforms/InstCombine/iX-ext-split.ll
@@ -10,14 +10,14 @@ define i128 @i128_ext_split(i32 noundef %x) {
; CHECK-NEXT: [[XX:%.*]] = sext i32 [[X]] to i128
; CHECK-NEXT: ret i128 [[XX]]
;
- %coerce.sroa.0.0.extract.trunc = sext i32 %x to i64
- %ashr = ashr i32 %x, 31
- %coerce.sroa.2.0.extract.trunc = sext i32 %ashr to i64
- %x.sroa.2.0.insert.ext.i = zext i64 %coerce.sroa.2.0.extract.trunc to i128
- %x.sroa.2.0.insert.shift.i = shl nuw i128 %x.sroa.2.0.insert.ext.i, 64
- %x.sroa.0.0.insert.ext.i = zext i64 %coerce.sroa.0.0.extract.trunc to i128
- %x.sroa.0.0.insert.insert.i = or disjoint i128 %x.sroa.2.0.insert.shift.i, %x.sroa.0.0.insert.ext.i
- ret i128 %x.sroa.0.0.insert.insert.i
+ %LowerSrc = sext i32 %x to i64
+ %sign = ashr i32 %x, 31
+ %UpperSrc = sext i32 %sign to i64
+ %widen = zext i64 %UpperSrc to i128
+ %hi = shl nuw i128 %widen, 64
+ %lo = zext i64 %LowerSrc to i128
+ %res = or disjoint i128 %hi, %lo
+ ret i128 %res
}
; ext split from i32 to i128
>From aa838564282cfe0d55f268d35933d8d49bc4b746 Mon Sep 17 00:00:00 2001
From: Muhammad Bassiouni <60100307+bassiounix at users.noreply.github.com>
Date: Wed, 5 Mar 2025 17:25:24 +0000
Subject: [PATCH 8/8] chore(InstCombine): add comments describing the pattern
---
llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index b5a2c5df7b4ce..330358ee749e5 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3119,6 +3119,8 @@ static Value *matchOrConcat(Instruction &Or, InstCombiner::BuilderTy &Builder) {
match(UpperSrc, m_BitReverse(m_Value(UpperBRev))))
return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);
+ // iX ext split: extending or(zext(x),shl(zext(y),bw/2) pattern
+ // to consume sext/ashr: or(zext(sext(x)),shl(zext(sext(ashr(x))),bw/2)
Value *X;
if (match(LowerSrc, m_SExt(m_Value(X))) &&
match(UpperSrc,
More information about the llvm-commits
mailing list