[llvm] [InstCombine] Fold `(add X, (sext/zext (icmp eq X, C)))` (PR #93840)
via llvm-commits
llvm-commits at lists.llvm.org
Sat Jun 1 10:40:52 PDT 2024
https://github.com/goldsteinn updated https://github.com/llvm/llvm-project/pull/93840
>From e51e6430733a2f55fd358d3de2388bfc8c263182 Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n at gmail.com>
Date: Thu, 30 May 2024 10:01:29 -0500
Subject: [PATCH 1/2] [InstCombine] Add tests for folding `(add X, (sext/zext
(icmp eq X, C)))`; NFC
---
.../InstCombine/fold-ext-eq-c-with-op.ll | 61 +++++++++++++++++++
1 file changed, 61 insertions(+)
create mode 100644 llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
diff --git a/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll b/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
new file mode 100644
index 0000000000000..848b0d189497f
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
@@ -0,0 +1,61 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+declare void @use.i8(i8)
+define i8 @fold_add_zext_eq_0(i8 %x) {
+; CHECK-LABEL: @fold_add_zext_eq_0(
+; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq i8 [[X:%.*]], 0
+; CHECK-NEXT: [[X_EQ_EXT:%.*]] = zext i1 [[X_EQ]] to i8
+; CHECK-NEXT: [[R:%.*]] = add i8 [[X_EQ_EXT]], [[X]]
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x_eq = icmp eq i8 %x, 0
+ %x_eq_ext = zext i1 %x_eq to i8
+ %r = add i8 %x, %x_eq_ext
+ ret i8 %r
+}
+
+define <2 x i8> @fold_add_sext_eq_4_6(<2 x i6> %xx) {
+; CHECK-LABEL: @fold_add_sext_eq_4_6(
+; CHECK-NEXT: [[X:%.*]] = zext <2 x i6> [[XX:%.*]] to <2 x i8>
+; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq <2 x i8> [[X]], <i8 4, i8 -127>
+; CHECK-NEXT: [[X_EQ_EXT:%.*]] = sext <2 x i1> [[X_EQ]] to <2 x i8>
+; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i8> [[X_EQ_EXT]], [[X]]
+; CHECK-NEXT: ret <2 x i8> [[R]]
+;
+ %x = zext <2 x i6> %xx to <2 x i8>
+ %x_eq = icmp eq <2 x i8> %x, <i8 4, i8 129>
+ %x_eq_ext = sext <2 x i1> %x_eq to <2 x i8>
+ %r = add <2 x i8> %x_eq_ext, %x
+ ret <2 x i8> %r
+}
+
+define i8 @fold_add_zext_eq_0_fail_multiuse_exp(i8 %x) {
+; CHECK-LABEL: @fold_add_zext_eq_0_fail_multiuse_exp(
+; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq i8 [[X:%.*]], 0
+; CHECK-NEXT: [[X_EQ_EXT:%.*]] = zext i1 [[X_EQ]] to i8
+; CHECK-NEXT: [[R:%.*]] = add i8 [[X_EQ_EXT]], [[X]]
+; CHECK-NEXT: call void @use.i8(i8 [[X_EQ_EXT]])
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x_eq = icmp eq i8 %x, 0
+ %x_eq_ext = zext i1 %x_eq to i8
+ %r = add i8 %x, %x_eq_ext
+ call void @use.i8(i8 %x_eq_ext)
+ ret i8 %r
+}
+
+define i8 @fold_add_sext_eq_4_fail_wrong_cond(i8 %x, i8 %y) {
+; CHECK-LABEL: @fold_add_sext_eq_4_fail_wrong_cond(
+; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq i8 [[Y:%.*]], 4
+; CHECK-NEXT: [[X_EQ_EXT:%.*]] = sext i1 [[X_EQ]] to i8
+; CHECK-NEXT: [[R:%.*]] = add i8 [[X_EQ_EXT]], [[X:%.*]]
+; CHECK-NEXT: call void @use.i8(i8 [[X_EQ_EXT]])
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x_eq = icmp eq i8 %y, 4
+ %x_eq_ext = sext i1 %x_eq to i8
+ %r = add i8 %x, %x_eq_ext
+ call void @use.i8(i8 %x_eq_ext)
+ ret i8 %r
+}
>From bc2ae3874d67ebca78fea0554fdcc020d60f5b30 Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n at gmail.com>
Date: Thu, 30 May 2024 10:01:33 -0500
Subject: [PATCH 2/2] [InstCombine] Fold `(add X, (sext/zext (icmp eq X, C)))`
We can convert this to a select based on the `(icmp eq X, C)`, then
constant fold the addition the true arm begin `(add C, (sext/zext 1))`
and the false arm being `(add X, 0)` e.g
- `(select (icmp eq X, C), (add C, (sext/zext 1)), (add X, 0))`.
This is essentially a specialization of the only case that sees to
actually show up from #89020
---
.../InstCombine/InstCombineAddSub.cpp | 18 ++++++++++++++++++
.../test/Transforms/InstCombine/apint-shift.ll | 17 +++++------------
.../InstCombine/fold-ext-eq-c-with-op.ll | 7 ++-----
3 files changed, 25 insertions(+), 17 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index bff09f5676680..8205b49dfbe2f 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1694,6 +1694,24 @@ Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
return BinaryOperator::CreateOr(LHS, Zext);
}
+ {
+ Value *Cond, *Ext;
+ Constant *C;
+ // (add X, (sext/zext (icmp eq X, C)))
+ // -> (select (icmp eq X, C), (add C, (sext/zext 1)), X)
+ auto CondMatcher = m_CombineAnd(
+ m_Value(Cond), m_ICmp(Pred, m_Deferred(A), m_ImmConstant(C)));
+
+ if (match(&I,
+ m_c_Add(m_Value(A),
+ m_CombineAnd(m_Value(Ext), m_ZExtOrSExt(CondMatcher)))) &&
+ Pred == ICmpInst::ICMP_EQ && Ext->hasOneUse()) {
+ Value *Add = isa<ZExtInst>(Ext) ? InstCombiner::AddOne(C)
+ : InstCombiner::SubOne(C);
+ return replaceInstUsesWith(I, Builder.CreateSelect(Cond, Add, A));
+ }
+ }
+
if (Instruction *Ashr = foldAddToAshr(I))
return Ashr;
diff --git a/llvm/test/Transforms/InstCombine/apint-shift.ll b/llvm/test/Transforms/InstCombine/apint-shift.ll
index 05c3db70ce1ca..ecf9c4e9c4e69 100644
--- a/llvm/test/Transforms/InstCombine/apint-shift.ll
+++ b/llvm/test/Transforms/InstCombine/apint-shift.ll
@@ -240,8 +240,8 @@ define i23 @test11(i23 %x) {
define i47 @test12(i47 %X) {
; CHECK-LABEL: @test12(
-; CHECK-NEXT: [[SH2:%.*]] = and i47 [[X:%.*]], -256
-; CHECK-NEXT: ret i47 [[SH2]]
+; CHECK-NEXT: [[SH1:%.*]] = and i47 [[X:%.*]], -256
+; CHECK-NEXT: ret i47 [[SH1]]
;
%sh1 = ashr i47 %X, 8
%sh2 = shl i47 %sh1, 8
@@ -250,8 +250,8 @@ define i47 @test12(i47 %X) {
define <2 x i47> @test12_splat_vec(<2 x i47> %X) {
; CHECK-LABEL: @test12_splat_vec(
-; CHECK-NEXT: [[SH2:%.*]] = and <2 x i47> [[X:%.*]], <i47 -256, i47 -256>
-; CHECK-NEXT: ret <2 x i47> [[SH2]]
+; CHECK-NEXT: [[SH1:%.*]] = and <2 x i47> [[X:%.*]], <i47 -256, i47 -256>
+; CHECK-NEXT: ret <2 x i47> [[SH1]]
;
%sh1 = ashr <2 x i47> %X, <i47 8, i47 8>
%sh2 = shl <2 x i47> %sh1, <i47 8, i47 8>
@@ -564,14 +564,7 @@ define i40 @test26(i40 %A) {
; https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=9880
define i177 @ossfuzz_9880(i177 %X) {
; CHECK-LABEL: @ossfuzz_9880(
-; CHECK-NEXT: [[A:%.*]] = alloca i177, align 8
-; CHECK-NEXT: [[L1:%.*]] = load i177, ptr [[A]], align 4
-; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i177 [[L1]], -1
-; CHECK-NEXT: [[B5_NEG:%.*]] = sext i1 [[TMP1]] to i177
-; CHECK-NEXT: [[B14:%.*]] = add i177 [[L1]], [[B5_NEG]]
-; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i177 [[B14]], -1
-; CHECK-NEXT: [[B1:%.*]] = zext i1 [[TMP2]] to i177
-; CHECK-NEXT: ret i177 [[B1]]
+; CHECK-NEXT: ret i177 0
;
%A = alloca i177
%L1 = load i177, ptr %A
diff --git a/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll b/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
index 848b0d189497f..4d02d492d2aa7 100644
--- a/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
+++ b/llvm/test/Transforms/InstCombine/fold-ext-eq-c-with-op.ll
@@ -4,9 +4,7 @@
declare void @use.i8(i8)
define i8 @fold_add_zext_eq_0(i8 %x) {
; CHECK-LABEL: @fold_add_zext_eq_0(
-; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq i8 [[X:%.*]], 0
-; CHECK-NEXT: [[X_EQ_EXT:%.*]] = zext i1 [[X_EQ]] to i8
-; CHECK-NEXT: [[R:%.*]] = add i8 [[X_EQ_EXT]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = call i8 @llvm.umax.i8(i8 [[X:%.*]], i8 1)
; CHECK-NEXT: ret i8 [[R]]
;
%x_eq = icmp eq i8 %x, 0
@@ -19,8 +17,7 @@ define <2 x i8> @fold_add_sext_eq_4_6(<2 x i6> %xx) {
; CHECK-LABEL: @fold_add_sext_eq_4_6(
; CHECK-NEXT: [[X:%.*]] = zext <2 x i6> [[XX:%.*]] to <2 x i8>
; CHECK-NEXT: [[X_EQ:%.*]] = icmp eq <2 x i8> [[X]], <i8 4, i8 -127>
-; CHECK-NEXT: [[X_EQ_EXT:%.*]] = sext <2 x i1> [[X_EQ]] to <2 x i8>
-; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i8> [[X_EQ_EXT]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = select <2 x i1> [[X_EQ]], <2 x i8> <i8 3, i8 -128>, <2 x i8> [[X]]
; CHECK-NEXT: ret <2 x i8> [[R]]
;
%x = zext <2 x i6> %xx to <2 x i8>
More information about the llvm-commits
mailing list