[llvm] [InstCombine] Fold chained GEP with constant base into single GEP (PR #170439)
Jianjian Guan via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 11 22:22:05 PST 2025
https://github.com/jacquesguan updated https://github.com/llvm/llvm-project/pull/170439
>From 58671e1c0dab3659b3a9583333d9f10efb7348b8 Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Mon, 1 Dec 2025 16:47:28 +0800
Subject: [PATCH 1/7] [InstCombine] Fold chained GEP with constant base into
single GEP
---
.../InstCombine/InstructionCombining.cpp | 27 +++++++++++
.../gep-fold-chained-const-select.ll | 46 +++++++++++++++++++
2 files changed, 73 insertions(+)
create mode 100644 llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index c6de57cb34c69..ed2df767d8b32 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2803,6 +2803,33 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
if (Src->getResultElementType() != GEP.getSourceElementType())
return nullptr;
+ // Fold chained GEP with constant base into single GEP:
+ // gep i8, (gep i8, %base, C1), (select Cond, C2, C3)
+ // -> gep i8, %base, (select Cond, C1+C2, C1+C3)
+ if (GEP.getNumIndices() == 1 && Src->getNumIndices() == 1) {
+ Value *SrcIdx = *Src->idx_begin();
+ Value *GEPIdx = *GEP.idx_begin();
+ const APInt *ConstOffset, *TrueVal, *FalseVal;
+ Value *Cond;
+ if (match(SrcIdx, m_APInt(ConstOffset)) &&
+ match(GEPIdx,
+ m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal)))) {
+ if (TrueVal->getBitWidth() != ConstOffset->getBitWidth() ||
+ FalseVal->getBitWidth() != ConstOffset->getBitWidth())
+ return nullptr;
+ APInt NewTrueVal = *ConstOffset + *TrueVal;
+ APInt NewFalseVal = *ConstOffset + *FalseVal;
+ Value *NewSelect = Builder.CreateSelect(Cond, Builder.getInt(NewTrueVal),
+ Builder.getInt(NewFalseVal));
+ GEPNoWrapFlags Flags =
+ getMergedGEPNoWrapFlags(*Src, *cast<GEPOperator>(&GEP));
+ return replaceInstUsesWith(GEP,
+ Builder.CreateGEP(GEP.getSourceElementType(),
+ Src->getPointerOperand(),
+ NewSelect, "", Flags));
+ }
+ }
+
// Find out whether the last index in the source GEP is a sequential idx.
bool EndsWithSequential = false;
for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
new file mode 100644
index 0000000000000..0591986ae4509
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -0,0 +1,46 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -passes=instcombine | FileCheck %s
+
+define ptr @fold(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @fold(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[V0:%.*]], i64 [[V2]]
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr i8, ptr %v0, i64 %v2
+ ret ptr %v3
+}
+
+define ptr @no_fold_0(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @no_fold_0(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 8148
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i16, ptr [[V0]], i64 [[V2]]
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr i16, ptr %v0, i64 %v2
+ ret ptr %v3
+}
+
+define ptr @no_fold_1(i32 %arg0, ptr %arg1, i64 %arg2) {
+; CHECK-LABEL: @no_fold_1(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 [[ARG2:%.*]]
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[V0]], i64 [[V2]]
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 %arg2
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr i8, ptr %v0, i64 %v2
+ ret ptr %v3
+}
>From 552114af6e2ecf936fe8a108dfda34e81b1041b4 Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Thu, 4 Dec 2025 14:44:56 +0800
Subject: [PATCH 2/7] Add more cases
---
.../gep-fold-chained-const-select.ll | 55 ++++++++++++++++---
1 file changed, 46 insertions(+), 9 deletions(-)
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
index 0591986ae4509..ad955673744e6 100644
--- a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -1,22 +1,37 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -S -passes=instcombine | FileCheck %s
-define ptr @fold(i32 %arg0, ptr %arg1) {
-; CHECK-LABEL: @fold(
+define ptr @src_origin(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_origin(
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
-; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 63252, i64 29452
-; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[V0:%.*]], i64 [[V2]]
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
; CHECK-NEXT: ret ptr [[V3]]
;
- %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v0 = getelementptr i8, ptr %arg1, i64 8148
%v1 = icmp sgt i32 %arg0, 3
%v2 = select i1 %v1, i64 55104, i64 21304
%v3 = getelementptr i8, ptr %v0, i64 %v2
ret ptr %v3
}
-define ptr @no_fold_0(i32 %arg0, ptr %arg1) {
-; CHECK-LABEL: @no_fold_0(
+define ptr @src_nuw(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_nuw(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr nuw i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr nuw i8, ptr %v0, i64 %v2
+ ret ptr %v3
+}
+
+; Fail 1: Different GEP type
+define ptr @src_fail_different_type(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_fail_different_type(
; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 8148
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
@@ -30,8 +45,9 @@ define ptr @no_fold_0(i32 %arg0, ptr %arg1) {
ret ptr %v3
}
-define ptr @no_fold_1(i32 %arg0, ptr %arg1, i64 %arg2) {
-; CHECK-LABEL: @no_fold_1(
+; Fail 2: No constant idx
+define ptr @src_fail_no_constant_idx(i32 %arg0, ptr %arg1, i64 %arg2) {
+; CHECK-LABEL: @src_fail_no_constant_idx(
; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 [[ARG2:%.*]]
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
@@ -44,3 +60,24 @@ define ptr @no_fold_1(i32 %arg0, ptr %arg1, i64 %arg2) {
%v3 = getelementptr i8, ptr %v0, i64 %v2
ret ptr %v3
}
+
+; Fail 2: Multiple use
+define ptr @src_fail_select_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
+; CHECK-LABEL: @src_fail_select_multiple_use(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: [[V4:%.*]] = getelementptr i8, ptr [[ARG2:%.*]], i64 [[V2]]
+; CHECK-NEXT: store ptr [[V3]], ptr [[V4]], align 8
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr i8, ptr %v0, i64 %v2
+ %v4 = getelementptr i8, ptr %arg2, i64 %v2
+ store ptr %v3, ptr %v4, align 8
+ ret ptr %v3
+}
+
>From 1b4d39ddf658e3a9b1e55e6a1d536b1a7da6b20f Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Tue, 9 Dec 2025 16:33:32 +0800
Subject: [PATCH 3/7] Address comment
---
.../InstCombine/InstructionCombining.cpp | 16 +++-
.../gep-fold-chained-const-select.ll | 85 ++++++++++++++++++-
2 files changed, 95 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index ed2df767d8b32..e3d7009fc60d4 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2811,12 +2811,22 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
Value *GEPIdx = *GEP.idx_begin();
const APInt *ConstOffset, *TrueVal, *FalseVal;
Value *Cond;
- if (match(SrcIdx, m_APInt(ConstOffset)) &&
- match(GEPIdx,
- m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal)))) {
+
+ if ((match(SrcIdx, m_APInt(ConstOffset)) &&
+ match(GEPIdx,
+ m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal)))) ||
+ (match(GEPIdx, m_APInt(ConstOffset)) &&
+ match(SrcIdx,
+ m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal))))) {
+ // Make sure the select has only one use.
+ if ((isa<SelectInst>(GEPIdx) && !GEPIdx->hasOneUse()) ||
+ (isa<SelectInst>(SrcIdx) && !SrcIdx->hasOneUse()))
+ return nullptr;
+
if (TrueVal->getBitWidth() != ConstOffset->getBitWidth() ||
FalseVal->getBitWidth() != ConstOffset->getBitWidth())
return nullptr;
+
APInt NewTrueVal = *ConstOffset + *TrueVal;
APInt NewFalseVal = *ConstOffset + *FalseVal;
Value *NewSelect = Builder.CreateSelect(Cond, Builder.getInt(NewTrueVal),
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
index ad955673744e6..964b1cc1337fa 100644
--- a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -29,6 +29,34 @@ define ptr @src_nuw(i32 %arg0, ptr %arg1) {
ret ptr %v3
}
+define ptr @src_inbounds_nuw(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_inbounds_nuw(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr nuw i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr nuw i8, ptr %v0, i64 %v2
+ ret ptr %v3
+}
+
+define ptr @src_swap(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_swap(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V0:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[V2]]
+; CHECK-NEXT: ret ptr [[V0]]
+;
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v0 = getelementptr i8, ptr %arg1, i64 %v2
+ %v3 = getelementptr i8, ptr %v0, i64 8148
+ ret ptr %v3
+}
+
; Fail 1: Different GEP type
define ptr @src_fail_different_type(i32 %arg0, ptr %arg1) {
; CHECK-LABEL: @src_fail_different_type(
@@ -61,13 +89,13 @@ define ptr @src_fail_no_constant_idx(i32 %arg0, ptr %arg1, i64 %arg2) {
ret ptr %v3
}
-; Fail 2: Multiple use
+; Fail 3: Multiple use of select
define ptr @src_fail_select_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
; CHECK-LABEL: @src_fail_select_multiple_use(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 8148
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
-; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
-; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[V0]], i64 [[V2]]
; CHECK-NEXT: [[V4:%.*]] = getelementptr i8, ptr [[ARG2:%.*]], i64 [[V2]]
; CHECK-NEXT: store ptr [[V3]], ptr [[V4]], align 8
; CHECK-NEXT: ret ptr [[V3]]
@@ -81,3 +109,54 @@ define ptr @src_fail_select_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
ret ptr %v3
}
+; Fail 4: Multiple use of source GEP
+define ptr @src_fail_source_gep_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
+; CHECK-LABEL: @src_fail_source_gep_multiple_use(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: [[V4:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8248
+; CHECK-NEXT: store ptr [[V3]], ptr [[V4]], align 8
+; CHECK-NEXT: ret ptr [[V3]]
+;
+ %v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %v3 = getelementptr i8, ptr %v0, i64 %v2
+ %v4 = getelementptr i8, ptr %v0, i64 100
+ store ptr %v3, ptr %v4, align 8
+ ret ptr %v3
+}
+
+; TODO: constant vector index
+
+define <2 x ptr> @fail_vector_const_vector_select(i32 %arg0, <2 x ptr> %arg1) {
+; CHECK-LABEL: @fail_vector_const_vector_select(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr i8, <2 x ptr> [[ARG1:%.*]], <2 x i64> <i64 8148, i64 8848>
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], <2 x i64> <i64 55104, i64 56654>, <2 x i64> <i64 21304, i64 21224>
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, <2 x ptr> [[V0]], <2 x i64> [[V2]]
+; CHECK-NEXT: ret <2 x ptr> [[V3]]
+;
+ %v0 = getelementptr i8, <2 x ptr> %arg1, <2 x i64> <i64 8148, i64 8848>
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, <2 x i64> <i64 55104, i64 56654>, <2 x i64> <i64 21304, i64 21224>
+ %v3 = getelementptr i8, <2 x ptr> %v0, <2 x i64> %v2
+ ret <2 x ptr> %v3
+}
+
+
+define <2 x ptr> @fail_scalar_const_vector_select(i32 %arg0, <2 x ptr> %arg1) {
+; CHECK-LABEL: @fail_scalar_const_vector_select(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr i8, <2 x ptr> [[ARG1:%.*]], i64 8148
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], <2 x i64> <i64 55104, i64 56654>, <2 x i64> <i64 21304, i64 21224>
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, <2 x ptr> [[V0]], <2 x i64> [[V2]]
+; CHECK-NEXT: ret <2 x ptr> [[V3]]
+;
+ %v0 = getelementptr i8, <2 x ptr> %arg1, i64 8148
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, <2 x i64> <i64 55104, i64 56654>, <2 x i64> <i64 21304, i64 21224>
+ %v3 = getelementptr i8, <2 x ptr> %v0, <2 x i64> %v2
+ ret <2 x ptr> %v3
+}
>From 7aa148a7032fbcb8670ce8fac01e416c9045cbaa Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Tue, 9 Dec 2025 17:47:59 +0800
Subject: [PATCH 4/7] Add splat case
---
.../InstCombine/gep-fold-chained-const-select.ll | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
index 964b1cc1337fa..c295c81e1733e 100644
--- a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -57,6 +57,20 @@ define ptr @src_swap(i32 %arg0, ptr %arg1) {
ret ptr %v3
}
+define <2 x ptr> @src_splat(i32 %arg0, <2 x ptr> %arg1) {
+; CHECK-LABEL: @src_splat(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, <2 x ptr> [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret <2 x ptr> [[V3]]
+;
+ %v0 = getelementptr i8, <2 x ptr> %arg1, <2 x i64> splat (i64 8148)
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, <2 x i64> splat (i64 55104), <2 x i64> splat (i64 21304)
+ %v3 = getelementptr i8, <2 x ptr> %v0, <2 x i64> %v2
+ ret <2 x ptr> %v3
+}
+
; Fail 1: Different GEP type
define ptr @src_fail_different_type(i32 %arg0, ptr %arg1) {
; CHECK-LABEL: @src_fail_different_type(
>From 3e7794cecdcf93b19cd8aaca13ee664243bb9b6f Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Thu, 11 Dec 2025 17:52:09 +0800
Subject: [PATCH 5/7] Address comment
---
.../InstCombine/InstructionCombining.cpp | 19 ++++++++----
.../gep-fold-chained-const-select.ll | 29 ++++++++++++++-----
2 files changed, 35 insertions(+), 13 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index e3d7009fc60d4..85d880d2c437e 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2818,10 +2818,11 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
(match(GEPIdx, m_APInt(ConstOffset)) &&
match(SrcIdx,
m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal))))) {
+ auto *Select = isa<SelectInst>(GEPIdx) ? cast<SelectInst>(GEPIdx) : cast<SelectInst>(SrcIdx);
+
// Make sure the select has only one use.
- if ((isa<SelectInst>(GEPIdx) && !GEPIdx->hasOneUse()) ||
- (isa<SelectInst>(SrcIdx) && !SrcIdx->hasOneUse()))
- return nullptr;
+ if (!Select->hasOneUse())
+ return nullptr;
if (TrueVal->getBitWidth() != ConstOffset->getBitWidth() ||
FalseVal->getBitWidth() != ConstOffset->getBitWidth())
@@ -2829,12 +2830,18 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
APInt NewTrueVal = *ConstOffset + *TrueVal;
APInt NewFalseVal = *ConstOffset + *FalseVal;
- Value *NewSelect = Builder.CreateSelect(Cond, Builder.getInt(NewTrueVal),
- Builder.getInt(NewFalseVal));
+ Constant *NewTrue = Builder.getInt(NewTrueVal);
+ Constant *NewFalse = Builder.getInt(NewFalseVal);
+ // Consider vector splat.
+ if (auto *VT = dyn_cast<VectorType>(Select->getType())) {
+ NewTrue = ConstantVector::getSplat(VT->getElementCount(), NewTrue);
+ NewFalse = ConstantVector::getSplat(VT->getElementCount(), NewFalse);
+ }
+ Value *NewSelect = Builder.CreateSelect(Cond, NewTrue, NewFalse);
GEPNoWrapFlags Flags =
getMergedGEPNoWrapFlags(*Src, *cast<GEPOperator>(&GEP));
return replaceInstUsesWith(GEP,
- Builder.CreateGEP(GEP.getSourceElementType(),
+ Builder.CreateGEP(GEP.getResultElementType(),
Src->getPointerOperand(),
NewSelect, "", Flags));
}
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
index c295c81e1733e..f3b4758fa5cc8 100644
--- a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -60,8 +60,8 @@ define ptr @src_swap(i32 %arg0, ptr %arg1) {
define <2 x ptr> @src_splat(i32 %arg0, <2 x ptr> %arg1) {
; CHECK-LABEL: @src_splat(
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
-; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
-; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, <2 x ptr> [[ARG1:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], <2 x i64> splat (i64 63252), <2 x i64> splat (i64 29452)
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, <2 x ptr> [[ARG1:%.*]], <2 x i64> [[TMP1]]
; CHECK-NEXT: ret <2 x ptr> [[V3]]
;
%v0 = getelementptr i8, <2 x ptr> %arg1, <2 x i64> splat (i64 8148)
@@ -71,6 +71,20 @@ define <2 x ptr> @src_splat(i32 %arg0, <2 x ptr> %arg1) {
ret <2 x ptr> %v3
}
+define <2 x ptr> @src_splat_scalar_ptr(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_splat_scalar_ptr(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], <2 x i64> splat (i64 63252), <2 x i64> splat (i64 29452)
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], <2 x i64> [[TMP1]]
+; CHECK-NEXT: ret <2 x ptr> [[V3]]
+;
+ %v0 = getelementptr i8, ptr %arg1, <2 x i64> splat (i64 8148)
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, <2 x i64> splat (i64 55104), <2 x i64> splat (i64 21304)
+ %v3 = getelementptr i8, <2 x ptr> %v0, <2 x i64> %v2
+ ret <2 x ptr> %v3
+}
+
; Fail 1: Different GEP type
define ptr @src_fail_different_type(i32 %arg0, ptr %arg1) {
; CHECK-LABEL: @src_fail_different_type(
@@ -123,22 +137,23 @@ define ptr @src_fail_select_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
ret ptr %v3
}
+declare void @use(ptr)
+
; Fail 4: Multiple use of source GEP
define ptr @src_fail_source_gep_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
; CHECK-LABEL: @src_fail_source_gep_multiple_use(
+; CHECK-NEXT: [[V0:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 8148
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: call void @use(ptr nonnull [[V0]])
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
-; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
-; CHECK-NEXT: [[V4:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8248
-; CHECK-NEXT: store ptr [[V3]], ptr [[V4]], align 8
+; CHECK-NEXT: [[V3:%.*]] = getelementptr i8, ptr [[ARG1]], i64 [[TMP1]]
; CHECK-NEXT: ret ptr [[V3]]
;
%v0 = getelementptr inbounds nuw i8, ptr %arg1, i64 8148
%v1 = icmp sgt i32 %arg0, 3
%v2 = select i1 %v1, i64 55104, i64 21304
+ call void @use(ptr %v0)
%v3 = getelementptr i8, ptr %v0, i64 %v2
- %v4 = getelementptr i8, ptr %v0, i64 100
- store ptr %v3, ptr %v4, align 8
ret ptr %v3
}
>From 66afed322f037fdf99f83327fea50f4cabb9c1cd Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Thu, 11 Dec 2025 17:56:12 +0800
Subject: [PATCH 6/7] Format
---
llvm/lib/Transforms/InstCombine/InstructionCombining.cpp | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 85d880d2c437e..17c2618f0ffef 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2818,11 +2818,12 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
(match(GEPIdx, m_APInt(ConstOffset)) &&
match(SrcIdx,
m_Select(m_Value(Cond), m_APInt(TrueVal), m_APInt(FalseVal))))) {
- auto *Select = isa<SelectInst>(GEPIdx) ? cast<SelectInst>(GEPIdx) : cast<SelectInst>(SrcIdx);
+ auto *Select = isa<SelectInst>(GEPIdx) ? cast<SelectInst>(GEPIdx)
+ : cast<SelectInst>(SrcIdx);
// Make sure the select has only one use.
if (!Select->hasOneUse())
- return nullptr;
+ return nullptr;
if (TrueVal->getBitWidth() != ConstOffset->getBitWidth() ||
FalseVal->getBitWidth() != ConstOffset->getBitWidth())
>From 53be14b21c2d38e6fa8fe07ad5ef09a20a1c6177 Mon Sep 17 00:00:00 2001
From: Jianjian GUAN <jacquesguan at me.com>
Date: Fri, 12 Dec 2025 14:21:39 +0800
Subject: [PATCH 7/7] Address comment
---
.../InstCombine/InstructionCombining.cpp | 9 ++------
.../gep-fold-chained-const-select.ll | 22 +++++++++++++++++--
2 files changed, 22 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 17c2618f0ffef..6e08af63f0863 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2831,13 +2831,8 @@ Instruction *InstCombinerImpl::visitGEPOfGEP(GetElementPtrInst &GEP,
APInt NewTrueVal = *ConstOffset + *TrueVal;
APInt NewFalseVal = *ConstOffset + *FalseVal;
- Constant *NewTrue = Builder.getInt(NewTrueVal);
- Constant *NewFalse = Builder.getInt(NewFalseVal);
- // Consider vector splat.
- if (auto *VT = dyn_cast<VectorType>(Select->getType())) {
- NewTrue = ConstantVector::getSplat(VT->getElementCount(), NewTrue);
- NewFalse = ConstantVector::getSplat(VT->getElementCount(), NewFalse);
- }
+ Constant *NewTrue = ConstantInt::get(Select->getType(), NewTrueVal);
+ Constant *NewFalse = ConstantInt::get(Select->getType(), NewFalseVal);
Value *NewSelect = Builder.CreateSelect(Cond, NewTrue, NewFalse);
GEPNoWrapFlags Flags =
getMergedGEPNoWrapFlags(*Src, *cast<GEPOperator>(&GEP));
diff --git a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
index f3b4758fa5cc8..35397eaf4438e 100644
--- a/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
+++ b/llvm/test/Transforms/InstCombine/gep-fold-chained-const-select.ll
@@ -46,8 +46,8 @@ define ptr @src_inbounds_nuw(i32 %arg0, ptr %arg1) {
define ptr @src_swap(i32 %arg0, ptr %arg1) {
; CHECK-LABEL: @src_swap(
; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
-; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 63252, i64 29452
-; CHECK-NEXT: [[V0:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[V2]]
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[V0:%.*]] = getelementptr i8, ptr [[ARG1:%.*]], i64 [[TMP1]]
; CHECK-NEXT: ret ptr [[V0]]
;
%v1 = icmp sgt i32 %arg0, 3
@@ -157,6 +157,24 @@ define ptr @src_fail_source_gep_multiple_use(i32 %arg0, ptr %arg1, ptr %arg2) {
ret ptr %v3
}
+define ptr @src_fail_source_gep_multiple_swap(i32 %arg0, ptr %arg1) {
+; CHECK-LABEL: @src_fail_source_gep_multiple_swap(
+; CHECK-NEXT: [[V1:%.*]] = icmp sgt i32 [[ARG0:%.*]], 3
+; CHECK-NEXT: [[V2:%.*]] = select i1 [[V1]], i64 55104, i64 21304
+; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds nuw i8, ptr [[ARG1:%.*]], i64 [[V2]]
+; CHECK-NEXT: call void @use(ptr nonnull [[GEP1]])
+; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[V1]], i64 63252, i64 29452
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr i8, ptr [[ARG1]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP2]]
+;
+ %v1 = icmp sgt i32 %arg0, 3
+ %v2 = select i1 %v1, i64 55104, i64 21304
+ %gep1 = getelementptr inbounds nuw i8, ptr %arg1, i64 %v2
+ call void @use(ptr %gep1)
+ %gep2 = getelementptr i8, ptr %gep1, i64 8148
+ ret ptr %gep2
+}
+
; TODO: constant vector index
define <2 x ptr> @fail_vector_const_vector_select(i32 %arg0, <2 x ptr> %arg1) {
More information about the llvm-commits
mailing list