[llvm] [ConstraintElim] Add NUW flag to sub if x >=u y. (PR #218685)
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 25 06:36:09 PDT 2026
https://github.com/fhahn updated https://github.com/llvm/llvm-project/pull/218685
>From 4f2257acb90f6750cf6bca9d21b454933c7ed6b2 Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Mon, 17 Aug 2026 16:48:02 +0100
Subject: [PATCH 1/2] [ConstraintElim] Precommit tests for flags strengthening
---
.../materialize-flags-narrow-pointer-index.ll | 208 ++++++++
.../materialize-flags.ll | 457 ++++++++++++++++++
...onstraint-elimination-materialize-flags.ll | 262 ++++++++++
3 files changed, 927 insertions(+)
create mode 100644 llvm/test/Transforms/ConstraintElimination/materialize-flags-narrow-pointer-index.ll
create mode 100644 llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
create mode 100644 llvm/test/Transforms/PhaseOrdering/constraint-elimination-materialize-flags.ll
diff --git a/llvm/test/Transforms/ConstraintElimination/materialize-flags-narrow-pointer-index.ll b/llvm/test/Transforms/ConstraintElimination/materialize-flags-narrow-pointer-index.ll
new file mode 100644
index 0000000000000..bcdf22c109743
--- /dev/null
+++ b/llvm/test/Transforms/ConstraintElimination/materialize-flags-narrow-pointer-index.ll
@@ -0,0 +1,208 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=constraint-elimination -S %s | FileCheck %s
+
+; The offsets of a GEP are computed in the pointer index type. With a 16 bit
+; index type, a stride or struct member offset that does not fit in 15 bits is
+; negative there, so a non-negative index does not imply a non-negative offset
+; and nuw must not be added.
+
+target datalayout = "p:16:16"
+
+%S.big = type { [40000 x i8], i8 }
+%S.small = type { i16, i8 }
+
+; The stride 40000 is negative as an i16, so the offset for a positive index is
+; negative and the GEP may wrap in the unsigned sense.
+define ptr @gep_no_nuw_stride_negative_in_index_type(ptr %p, i16 %i, i16 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_stride_negative_in_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i16 [[I:%.*]], i16 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i16 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i16 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw [40000 x i8], ptr [[P]], i16 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i16 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i16 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw [40000 x i8], ptr %p, i16 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; Same for the offset of a struct member, which is 40000 for the second field.
+define ptr @gep_no_nuw_struct_offset_negative_in_index_type(ptr %p, i16 %i, i16 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_struct_offset_negative_in_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i16 [[I:%.*]], i16 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i16 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i16 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw [[S_BIG:%.*]], ptr [[P]], i16 [[I]], i32 1
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i16 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i16 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw %S.big, ptr %p, i16 %i, i32 1
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; The stride of the outer array is 160000, which is 28928 and thus non-negative
+; as an i16, but the stride of the inner array is negative.
+define ptr @gep_no_nuw_inner_stride_negative_in_index_type(ptr %p, i16 %i, i16 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_inner_stride_negative_in_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i16 [[I:%.*]], i16 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i16 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i16 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw [4 x [40000 x i8]], ptr [[P]], i16 [[I]], i16 [[J]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i16 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i16 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw [4 x [40000 x i8]], ptr %p, i16 %i, i16 %j
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; The stride fits in the index type, so nuw can be added.
+define ptr @gep_nuw_stride_fits_index_type(ptr %p, i16 %i, i16 %j) {
+; CHECK-LABEL: define ptr @gep_nuw_stride_fits_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i16 [[I:%.*]], i16 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i16 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i16 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i16, ptr [[P]], i16 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i16 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i16 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw i16, ptr %p, i16 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; The struct member offset fits in the index type, so nuw can be added.
+define ptr @gep_nuw_struct_offset_fits_index_type(ptr %p, i16 %i, i16 %j) {
+; CHECK-LABEL: define ptr @gep_nuw_struct_offset_fits_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i16 [[I:%.*]], i16 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i16 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i16 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw [[S_SMALL:%.*]], ptr [[P]], i16 [[I]], i32 1
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i16 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i16 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw %S.small, ptr %p, i16 %i, i32 1
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; An index wider than the index type is truncated. nusw guarantees the
+; truncation preserves the signed value, which together with the index being
+; non-negative means it also preserves the unsigned value.
+define ptr @gep_nuw_index_wider_than_index_type(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_nuw_index_wider_than_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i16, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw i16, ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
diff --git a/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll b/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
new file mode 100644
index 0000000000000..6dfa70e044638
--- /dev/null
+++ b/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
@@ -0,0 +1,457 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=constraint-elimination -S %s | FileCheck %s
+
+define i64 @sext_to_zext_nneg(i32 %n, i32 %m) {
+; CHECK-LABEL: define i64 @sext_to_zext_nneg(
+; CHECK-SAME: i32 [[N:%.*]], i32 [[M:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp slt i32 [[N]], [[M]]
+; CHECK-NEXT: br i1 [[C]], label %[[EXIT:.*]], label %[[THEN:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[N]], [[M]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[SUB]] to i64
+; CHECK-NEXT: ret i64 [[EXT]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret i64 0
+;
+entry:
+ %c = icmp slt i32 %n, %m
+ br i1 %c, label %exit, label %then
+
+then:
+ %sub = sub nsw i32 %n, %m
+ %ext = sext i32 %sub to i64
+ ret i64 %ext
+
+exit:
+ ret i64 0
+}
+
+define i64 @sext_to_zext_nneg_negated_condition(i32 %n, i32 %m) {
+; CHECK-LABEL: define i64 @sext_to_zext_nneg_negated_condition(
+; CHECK-SAME: i32 [[N:%.*]], i32 [[M:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp slt i32 [[N]], [[M]]
+; CHECK-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[EXIT:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[N]], [[M]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[SUB]] to i64
+; CHECK-NEXT: ret i64 [[EXT]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret i64 0
+;
+entry:
+ %c = icmp slt i32 %n, %m
+ br i1 %c, label %then, label %exit
+
+then:
+ %sub = sub nsw i32 %n, %m
+ %ext = sext i32 %sub to i64
+ ret i64 %ext
+
+exit:
+ ret i64 0
+}
+
+define i64 @sext_no_conversion_known_via_value_tracking(i32 %n) {
+; CHECK-LABEL: define i64 @sext_no_conversion_known_via_value_tracking(
+; CHECK-SAME: i32 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[AND:%.*]] = and i32 [[N]], 255
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[AND]] to i64
+; CHECK-NEXT: ret i64 [[EXT]]
+;
+entry:
+ %and = and i32 %n, 255
+ %ext = sext i32 %and to i64
+ ret i64 %ext
+}
+
+define i32 @sub_nuw_relational(i32 %a, i32 %b) {
+; CHECK-LABEL: define i32 @sub_nuw_relational(
+; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp uge i32 [[A]], [[B]]
+; CHECK-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[EXIT:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A]], [[B]]
+; CHECK-NEXT: ret i32 [[SUB]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ %c = icmp uge i32 %a, %b
+ br i1 %c, label %then, label %exit
+
+then:
+ %sub = sub i32 %a, %b
+ ret i32 %sub
+
+exit:
+ ret i32 0
+}
+
+define i32 @sub_nuw_transitive(i32 %a, i32 %b, i32 %c) {
+; CHECK-LABEL: define i32 @sub_nuw_transitive(
+; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]], i32 [[C:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp uge i32 [[A]], [[B]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp uge i32 [[B]], [[C]]
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A]], [[C]]
+; CHECK-NEXT: ret i32 [[SUB]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ %c.0 = icmp uge i32 %a, %b
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp uge i32 %b, %c
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %sub = sub i32 %a, %c
+ ret i32 %sub
+
+exit:
+ ret i32 0
+}
+
+; A signed fact does not imply the unsigned no-wrap flag.
+define i32 @sub_no_nuw_signed_fact(i32 %a, i32 %b) {
+; CHECK-LABEL: define i32 @sub_no_nuw_signed_fact(
+; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp sge i32 [[A]], [[B]]
+; CHECK-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[EXIT:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A]], [[B]]
+; CHECK-NEXT: ret i32 [[SUB]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ %c = icmp sge i32 %a, %b
+ br i1 %c, label %then, label %exit
+
+then:
+ %sub = sub i32 %a, %b
+ ret i32 %sub
+
+exit:
+ ret i32 0
+}
+
+define ptr @gep_nuw_transitive(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_nuw_transitive(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw i32, ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+define ptr @gep_no_nuw_unknown_index(ptr %p, i64 %i, i64 %n) {
+; CHECK-LABEL: define ptr @gep_no_nuw_unknown_index(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp sge i64 [[I]], [[N]]
+; CHECK-NEXT: br i1 [[C]], label %[[EXIT:.*]], label %[[THEN:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c = icmp sge i64 %i, %n
+ br i1 %c, label %exit, label %then
+
+then:
+ %gep = getelementptr nusw i32, ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; The negative constant index means the GEP wraps in the unsigned sense.
+define ptr @gep_no_nuw_negative_constant_index(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_negative_constant_index(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: [[GEP_2:%.*]] = getelementptr nusw i32, ptr [[GEP]], i64 -1
+; CHECK-NEXT: ret ptr [[GEP_2]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw i32, ptr %p, i64 %i
+ %gep.2 = getelementptr nusw i32, ptr %gep, i64 -1
+ ret ptr %gep.2
+
+exit:
+ ret ptr null
+}
+
+; Without nusw, nuw cannot be implied by non-negative offsets.
+define ptr @gep_no_nusw(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_no_nusw(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr i32, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr i32, ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+define ptr @gep_no_nuw_constant_index(ptr %p) {
+; CHECK-LABEL: define ptr @gep_no_nuw_constant_index(
+; CHECK-SAME: ptr [[P:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C:%.*]] = icmp ne ptr [[P]], null
+; CHECK-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[EXIT:.*]]
+; CHECK: [[THEN]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 3
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c = icmp ne ptr %p, null
+ br i1 %c, label %then, label %exit
+
+then:
+ %gep = getelementptr nusw i32, ptr %p, i64 3
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+define i64 @no_facts(i32 %n, i32 %m, ptr %p) {
+; CHECK-LABEL: define i64 @no_facts(
+; CHECK-SAME: i32 [[N:%.*]], i32 [[M:%.*]], ptr [[P:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[N]], [[M]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[SUB]] to i64
+; CHECK-NEXT: [[SUB_2:%.*]] = sub i32 [[N]], [[M]]
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i32 [[SUB_2]]
+; CHECK-NEXT: store i32 0, ptr [[GEP]], align 4
+; CHECK-NEXT: ret i64 [[EXT]]
+;
+entry:
+ %sub = sub nsw i32 %n, %m
+ %ext = sext i32 %sub to i64
+ %sub.2 = sub i32 %n, %m
+ %gep = getelementptr nusw i32, ptr %p, i32 %sub.2
+ store i32 0, ptr %gep, align 4
+ ret i64 %ext
+}
+
+define void @induction_start_bound(ptr %p, i32 %start, i32 %n) {
+; CHECK-LABEL: define void @induction_start_bound(
+; CHECK-SAME: ptr [[P:%.*]], i32 [[START:%.*]], i32 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[START]], %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[IV]], [[START]]
+; CHECK-NEXT: [[IDX:%.*]] = sext i32 [[SUB]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 [[IDX]]
+; CHECK-NEXT: store i32 0, ptr [[GEP]], align 4
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i32 [[IV]], 1
+; CHECK-NEXT: [[EC:%.*]] = icmp slt i32 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC]], label %[[LOOP]], label %[[EXIT:.*]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ]
+ %sub = sub nsw i32 %iv, %start
+ %idx = sext i32 %sub to i64
+ %gep = getelementptr nusw i32, ptr %p, i64 %idx
+ store i32 0, ptr %gep, align 4
+ %iv.next = add nsw i32 %iv, 1
+ %ec = icmp slt i32 %iv.next, %n
+ br i1 %ec, label %loop, label %exit
+
+exit:
+ ret void
+}
+
+define void @induction_start_bound_ne_latch(ptr %p, i32 %start, i32 %n) {
+; CHECK-LABEL: define void @induction_start_bound_ne_latch(
+; CHECK-SAME: ptr [[P:%.*]], i32 [[START:%.*]], i32 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: br label %[[LOOP_HEADER:.*]]
+; CHECK: [[LOOP_HEADER]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[START]], %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
+; CHECK-NEXT: [[EC:%.*]] = icmp ne i32 [[IV]], [[N]]
+; CHECK-NEXT: br i1 [[EC]], label %[[LOOP_BODY:.*]], label %[[EXIT:.*]]
+; CHECK: [[LOOP_BODY]]:
+; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[IV]], [[START]]
+; CHECK-NEXT: [[IDX:%.*]] = sext i32 [[SUB]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i32, ptr [[P]], i64 [[IDX]]
+; CHECK-NEXT: store i32 0, ptr [[GEP]], align 4
+; CHECK-NEXT: br label %[[LOOP_LATCH]]
+; CHECK: [[LOOP_LATCH]]:
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i32 [[IV]], 1
+; CHECK-NEXT: br label %[[LOOP_HEADER]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ br label %loop.header
+
+loop.header:
+ %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop.latch ]
+ %ec = icmp ne i32 %iv, %n
+ br i1 %ec, label %loop.body, label %exit
+
+loop.body:
+ %sub = sub nsw i32 %iv, %start
+ %idx = sext i32 %sub to i64
+ %gep = getelementptr nusw i32, ptr %p, i64 %idx
+ store i32 0, ptr %gep, align 4
+ br label %loop.latch
+
+loop.latch:
+ %iv.next = add nsw i32 %iv, 1
+ br label %loop.header
+
+exit:
+ ret void
+}
+
+; The stride 2^63 is negative when interpreted in the pointer index type, so a
+; non-negative index does not imply a non-negative offset and nuw must not be
+; added.
+define ptr @gep_no_nuw_stride_negative_in_index_type(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_stride_negative_in_index_type(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw [9223372036854775808 x i8], ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw [9223372036854775808 x i8], ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
+
+; The sign of a scalable stride in the pointer index type is not known.
+define ptr @gep_no_nuw_scalable_stride(ptr %p, i64 %i, i64 %j) {
+; CHECK-LABEL: define ptr @gep_no_nuw_scalable_stride(
+; CHECK-SAME: ptr [[P:%.*]], i64 [[I:%.*]], i64 [[J:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[C_0:%.*]] = icmp sgt i64 [[I]], [[J]]
+; CHECK-NEXT: br i1 [[C_0]], label %[[BB_1:.*]], label %[[EXIT:.*]]
+; CHECK: [[BB_1]]:
+; CHECK-NEXT: [[C_1:%.*]] = icmp sgt i64 [[J]], 0
+; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
+; CHECK: [[BB_2]]:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw <vscale x 2 x i8>, ptr [[P]], i64 [[I]]
+; CHECK-NEXT: ret ptr [[GEP]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret ptr null
+;
+entry:
+ %c.0 = icmp sgt i64 %i, %j
+ br i1 %c.0, label %bb.1, label %exit
+
+bb.1:
+ %c.1 = icmp sgt i64 %j, 0
+ br i1 %c.1, label %bb.2, label %exit
+
+bb.2:
+ %gep = getelementptr nusw <vscale x 2 x i8>, ptr %p, i64 %i
+ ret ptr %gep
+
+exit:
+ ret ptr null
+}
diff --git a/llvm/test/Transforms/PhaseOrdering/constraint-elimination-materialize-flags.ll b/llvm/test/Transforms/PhaseOrdering/constraint-elimination-materialize-flags.ll
new file mode 100644
index 0000000000000..b94b0e6a384df
--- /dev/null
+++ b/llvm/test/Transforms/PhaseOrdering/constraint-elimination-materialize-flags.ll
@@ -0,0 +1,262 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -passes='default<O3>' -S %s | FileCheck %s
+
+define void @index_mask_removed1(ptr %A, i16 %start, i16 %v, i16 %n) {
+; CHECK-LABEL: define void @index_mask_removed1(
+; CHECK-SAME: ptr nofree writeonly captures(none) [[A:%.*]], i16 [[START:%.*]], i16 [[V:%.*]], i16 [[N:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[PRE:%.*]] = icmp slt i16 [[START]], [[V]]
+; CHECK-NEXT: [[EC_NOT1:%.*]] = icmp eq i16 [[START]], [[N]]
+; CHECK-NEXT: [[OR_COND:%.*]] = select i1 [[PRE]], i1 true, i1 [[EC_NOT1]]
+; CHECK-NEXT: br i1 [[OR_COND]], label %[[EXIT:.*]], label %[[LOOP_BODY_PREHEADER:.*]]
+; CHECK: [[LOOP_BODY_PREHEADER]]:
+; CHECK-NEXT: [[TMP0:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP1:%.*]] = add i16 [[N]], [[TMP0]]
+; CHECK-NEXT: [[TMP2:%.*]] = zext i16 [[TMP1]] to i32
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i32 [[TMP2]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i16 [[TMP1]], 11
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[LOOP_BODY_PREHEADER3:.*]], label %[[VECTOR_SCEVCHECK:.*]]
+; CHECK: [[VECTOR_SCEVCHECK]]:
+; CHECK-NEXT: [[TMP4:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP5:%.*]] = add i16 [[N]], [[TMP4]]
+; CHECK-NEXT: [[TMP6:%.*]] = sext i16 [[V]] to i32
+; CHECK-NEXT: [[TMP7:%.*]] = zext i16 [[TMP5]] to i32
+; CHECK-NEXT: [[TMP8:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP9:%.*]] = sext i16 [[TMP8]] to i32
+; CHECK-NEXT: [[TMP10:%.*]] = add nsw i32 [[TMP9]], [[TMP6]]
+; CHECK-NEXT: [[TMP11:%.*]] = icmp ult i32 [[TMP10]], [[TMP7]]
+; CHECK-NEXT: br i1 [[TMP11]], label %[[LOOP_BODY_PREHEADER3]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_VEC:%.*]] = and i32 [[TMP3]], 131068
+; CHECK-NEXT: [[TMP12:%.*]] = trunc i32 [[N_VEC]] to i16
+; CHECK-NEXT: [[TMP13:%.*]] = add i16 [[START]], [[TMP12]]
+; CHECK-NEXT: [[INVARIANT_OP:%.*]] = sub i16 [[START]], [[V]]
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP14:%.*]] = trunc i32 [[INDEX]] to i16
+; CHECK-NEXT: [[DOTREASS:%.*]] = add i16 [[TMP14]], [[INVARIANT_OP]]
+; CHECK-NEXT: [[TMP15:%.*]] = sext i16 [[DOTREASS]] to i64
+; CHECK-NEXT: [[TMP16:%.*]] = and i64 [[TMP15]], 4294967295
+; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[TMP16]]
+; CHECK-NEXT: store <4 x i8> splat (i8 1), ptr [[TMP17]], align 1
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
+; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP18]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP3]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT]], label %[[LOOP_BODY_PREHEADER3]]
+; CHECK: [[LOOP_BODY_PREHEADER3]]:
+; CHECK-NEXT: [[IV2_PH:%.*]] = phi i16 [ [[START]], %[[VECTOR_SCEVCHECK]] ], [ [[START]], %[[LOOP_BODY_PREHEADER]] ], [ [[TMP13]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: br label %[[LOOP_BODY:.*]]
+; CHECK: [[LOOP_BODY]]:
+; CHECK-NEXT: [[IV2:%.*]] = phi i16 [ [[IV_NEXT:%.*]], %[[LOOP_BODY]] ], [ [[IV2_PH]], %[[LOOP_BODY_PREHEADER3]] ]
+; CHECK-NEXT: [[OFF:%.*]] = sub nsw i16 [[IV2]], [[V]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i16 [[OFF]] to i64
+; CHECK-NEXT: [[MASK:%.*]] = and i64 [[EXT]], 4294967295
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[MASK]]
+; CHECK-NEXT: store i8 1, ptr [[GEP]], align 1
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i16 [[IV2]], 1
+; CHECK-NEXT: [[EC_NOT:%.*]] = icmp eq i16 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC_NOT]], label %[[EXIT]], label %[[LOOP_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %pre = icmp slt i16 %start, %v
+ br i1 %pre, label %exit, label %loop.header
+
+loop.header:
+ %iv = phi i16 [ %start, %entry ], [ %iv.next, %loop.latch ]
+ %ec = icmp ne i16 %iv, %n
+ br i1 %ec, label %loop.body, label %exit
+
+loop.body:
+ %off = sub nsw i16 %iv, %v
+ %ext = sext i16 %off to i64
+ %mask = and i64 %ext, 4294967295
+ %gep = getelementptr inbounds i8, ptr %A, i64 %mask
+ store i8 1, ptr %gep, align 1
+ br label %loop.latch
+
+loop.latch:
+ %iv.next = add nsw i16 %iv, 1
+ br label %loop.header
+
+exit:
+ ret void
+}
+
+define void @index_mask_removed2(ptr %A, i16 %start, i16 %v, i16 %n) {
+; CHECK-LABEL: define void @index_mask_removed2(
+; CHECK-SAME: ptr nofree writeonly captures(none) [[A:%.*]], i16 [[START:%.*]], i16 [[V:%.*]], i16 [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[PRE:%.*]] = icmp slt i16 [[START]], [[V]]
+; CHECK-NEXT: [[EC_NOT1:%.*]] = icmp eq i16 [[START]], [[N]]
+; CHECK-NEXT: [[OR_COND:%.*]] = select i1 [[PRE]], i1 true, i1 [[EC_NOT1]]
+; CHECK-NEXT: br i1 [[OR_COND]], label %[[EXIT:.*]], label %[[LOOP_BODY_PREHEADER:.*]]
+; CHECK: [[LOOP_BODY_PREHEADER]]:
+; CHECK-NEXT: [[TMP0:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP1:%.*]] = add i16 [[N]], [[TMP0]]
+; CHECK-NEXT: [[TMP2:%.*]] = zext i16 [[TMP1]] to i32
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i32 [[TMP2]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i16 [[TMP1]], 11
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[LOOP_BODY_PREHEADER3:.*]], label %[[VECTOR_SCEVCHECK:.*]]
+; CHECK: [[VECTOR_SCEVCHECK]]:
+; CHECK-NEXT: [[TMP4:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP5:%.*]] = add i16 [[N]], [[TMP4]]
+; CHECK-NEXT: [[TMP6:%.*]] = sext i16 [[V]] to i32
+; CHECK-NEXT: [[TMP7:%.*]] = zext i16 [[TMP5]] to i32
+; CHECK-NEXT: [[TMP8:%.*]] = xor i16 [[START]], -1
+; CHECK-NEXT: [[TMP9:%.*]] = sext i16 [[TMP8]] to i32
+; CHECK-NEXT: [[TMP10:%.*]] = add nsw i32 [[TMP9]], [[TMP6]]
+; CHECK-NEXT: [[TMP11:%.*]] = icmp ult i32 [[TMP10]], [[TMP7]]
+; CHECK-NEXT: br i1 [[TMP11]], label %[[LOOP_BODY_PREHEADER3]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_VEC:%.*]] = and i32 [[TMP3]], 131068
+; CHECK-NEXT: [[TMP12:%.*]] = trunc i32 [[N_VEC]] to i16
+; CHECK-NEXT: [[TMP13:%.*]] = add i16 [[START]], [[TMP12]]
+; CHECK-NEXT: [[INVARIANT_OP:%.*]] = sub i16 [[START]], [[V]]
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP14:%.*]] = trunc i32 [[INDEX]] to i16
+; CHECK-NEXT: [[DOTREASS:%.*]] = add i16 [[TMP14]], [[INVARIANT_OP]]
+; CHECK-NEXT: [[TMP15:%.*]] = sext i16 [[DOTREASS]] to i64
+; CHECK-NEXT: [[TMP16:%.*]] = and i64 [[TMP15]], 4294967295
+; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[TMP16]]
+; CHECK-NEXT: store <4 x i8> splat (i8 1), ptr [[TMP17]], align 1
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
+; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP18]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP3]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT]], label %[[LOOP_BODY_PREHEADER3]]
+; CHECK: [[LOOP_BODY_PREHEADER3]]:
+; CHECK-NEXT: [[IV2_PH:%.*]] = phi i16 [ [[START]], %[[VECTOR_SCEVCHECK]] ], [ [[START]], %[[LOOP_BODY_PREHEADER]] ], [ [[TMP13]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: br label %[[LOOP_BODY:.*]]
+; CHECK: [[LOOP_BODY]]:
+; CHECK-NEXT: [[IV2:%.*]] = phi i16 [ [[IV_NEXT:%.*]], %[[LOOP_BODY]] ], [ [[IV2_PH]], %[[LOOP_BODY_PREHEADER3]] ]
+; CHECK-NEXT: [[OFF:%.*]] = sub nsw i16 [[IV2]], [[V]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i16 [[OFF]] to i64
+; CHECK-NEXT: [[MASK:%.*]] = and i64 [[EXT]], 4294967295
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[MASK]]
+; CHECK-NEXT: store i8 1, ptr [[GEP]], align 1
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i16 [[IV2]], 1
+; CHECK-NEXT: [[EC_NOT:%.*]] = icmp eq i16 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC_NOT]], label %[[EXIT]], label %[[LOOP_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %pre = icmp slt i16 %start, %v
+ br i1 %pre, label %exit, label %loop.header
+
+loop.header:
+ %iv = phi i16 [ %start, %entry ], [ %iv.next, %loop.latch ]
+ %ec = icmp ne i16 %iv, %n
+ br i1 %ec, label %loop.body, label %exit
+
+loop.body:
+ %off = sub nsw i16 %iv, %v
+ %ext = sext i16 %off to i64
+ %mask = and i64 %ext, 4294967295
+ %gep = getelementptr inbounds i8, ptr %A, i64 %mask
+ store i8 1, ptr %gep, align 1
+ br label %loop.latch
+
+loop.latch:
+ %iv.next = add nsw i16 %iv, 1
+ br label %loop.header
+
+exit:
+ ret void
+}
+
+define void @index_not_widened1(ptr %A, i32 %start, i32 %v, i32 %n) {
+; CHECK-LABEL: define void @index_not_widened1(
+; CHECK-SAME: ptr nofree writeonly captures(none) [[A:%.*]], i32 [[START:%.*]], i32 [[V:%.*]], i32 [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[PRE:%.*]] = icmp slt i32 [[START]], [[V]]
+; CHECK-NEXT: [[EC_NOT1:%.*]] = icmp eq i32 [[START]], [[N]]
+; CHECK-NEXT: [[OR_COND:%.*]] = select i1 [[PRE]], i1 true, i1 [[EC_NOT1]]
+; CHECK-NEXT: br i1 [[OR_COND]], label %[[EXIT:.*]], label %[[LOOP_BODY:.*]]
+; CHECK: [[LOOP_BODY]]:
+; CHECK-NEXT: [[IV2:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP_BODY]] ], [ [[START]], %[[ENTRY]] ]
+; CHECK-NEXT: [[OFF:%.*]] = sub nsw i32 [[IV2]], [[V]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[OFF]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds [4 x i8], ptr [[A]], i64 [[EXT]]
+; CHECK-NEXT: store i32 [[IV2]], ptr [[GEP]], align 4
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i32 [[IV2]], 1
+; CHECK-NEXT: [[EC_NOT:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC_NOT]], label %[[EXIT]], label %[[LOOP_BODY]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %pre = icmp slt i32 %start, %v
+ br i1 %pre, label %exit, label %loop.header
+
+loop.header:
+ %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop.latch ]
+ %ec = icmp ne i32 %iv, %n
+ br i1 %ec, label %loop.body, label %exit
+
+loop.body:
+ %off = sub nsw i32 %iv, %v
+ %ext = sext i32 %off to i64
+ %gep = getelementptr inbounds [4 x i8], ptr %A, i64 %ext
+ store i32 %iv, ptr %gep, align 4
+ br label %loop.latch
+
+loop.latch:
+ %iv.next = add nsw i32 %iv, 1
+ br label %loop.header
+
+exit:
+ ret void
+}
+
+define void @index_not_widened2(ptr %A, i32 %start, i32 %v, i32 %n) {
+; CHECK-LABEL: define void @index_not_widened2(
+; CHECK-SAME: ptr nofree writeonly captures(none) [[A:%.*]], i32 [[START:%.*]], i32 [[V:%.*]], i32 [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[PRE:%.*]] = icmp slt i32 [[START]], [[V]]
+; CHECK-NEXT: [[EC_NOT1:%.*]] = icmp eq i32 [[START]], [[N]]
+; CHECK-NEXT: [[OR_COND:%.*]] = select i1 [[PRE]], i1 true, i1 [[EC_NOT1]]
+; CHECK-NEXT: br i1 [[OR_COND]], label %[[EXIT:.*]], label %[[LOOP_BODY:.*]]
+; CHECK: [[LOOP_BODY]]:
+; CHECK-NEXT: [[IV2:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP_BODY]] ], [ [[START]], %[[ENTRY]] ]
+; CHECK-NEXT: [[OFF:%.*]] = sub nsw i32 [[IV2]], [[V]]
+; CHECK-NEXT: [[EXT:%.*]] = sext i32 [[OFF]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds [4 x i8], ptr [[A]], i64 [[EXT]]
+; CHECK-NEXT: store i32 [[IV2]], ptr [[GEP]], align 4
+; CHECK-NEXT: [[IV_NEXT]] = add nsw i32 [[IV2]], 1
+; CHECK-NEXT: [[EC_NOT:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC_NOT]], label %[[EXIT]], label %[[LOOP_BODY]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %pre = icmp slt i32 %start, %v
+ br i1 %pre, label %exit, label %loop.header
+
+loop.header:
+ %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop.latch ]
+ %ec = icmp ne i32 %iv, %n
+ br i1 %ec, label %loop.body, label %exit
+
+loop.body:
+ %off = sub nsw i32 %iv, %v
+ %ext = sext i32 %off to i64
+ %gep = getelementptr inbounds [4 x i8], ptr %A, i64 %ext
+ store i32 %iv, ptr %gep, align 4
+ br label %loop.latch
+
+loop.latch:
+ %iv.next = add nsw i32 %iv, 1
+ br label %loop.header
+
+exit:
+ ret void
+}
>From 2f684e13a45b0490c081d369db175beb1c3b1eba Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Mon, 17 Aug 2026 18:45:55 +0100
Subject: [PATCH 2/2] [ConstraintElim] Strengthen flags on sub.
---
.../Scalar/ConstraintElimination.cpp | 49 +++++++++++++++++--
...nduction-condition-in-loop-exit-postinc.ll | 2 +-
.../materialize-flags.ll | 4 +-
.../reproducer-remarks.ll | 2 +-
.../Transforms/ConstraintElimination/sub.ll | 2 +-
5 files changed, 51 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Transforms/Scalar/ConstraintElimination.cpp b/llvm/lib/Transforms/Scalar/ConstraintElimination.cpp
index 054ab3b45e108..f3d5c3c6cb698 100644
--- a/llvm/lib/Transforms/Scalar/ConstraintElimination.cpp
+++ b/llvm/lib/Transforms/Scalar/ConstraintElimination.cpp
@@ -101,7 +101,7 @@ struct FactOrCheck {
InstFact, /// A fact that holds after Inst executed (e.g. an assume or
/// min/mix intrinsic.
InstCheck, /// An instruction to simplify (e.g. an overflow math
- /// intrinsics).
+ /// intrinsics) or whose flags may be strengthened.
UseCheck /// An use of a compare instruction to simplify.
};
@@ -146,8 +146,8 @@ struct FactOrCheck {
return FactOrCheck(DTN, U);
}
- static FactOrCheck getCheck(DomTreeNode *DTN, CallInst *CI) {
- return FactOrCheck(EntryTy::InstCheck, DTN, CI);
+ static FactOrCheck getCheck(DomTreeNode *DTN, Instruction *I) {
+ return FactOrCheck(EntryTy::InstCheck, DTN, I);
}
bool isCheck() const {
@@ -1291,6 +1291,40 @@ static bool getConstraintFromMemoryAccess(GetElementPtrInst &GEP,
return true;
}
+/// Returns true if \p I is a candidate whose poison-generating flags may be
+/// strengthened using the constraint systems.
+static bool canStrengthenFlags(Instruction *I) {
+ switch (I->getOpcode()) {
+ case Instruction::Sub:
+ // A - B does not wrap unsigned, if A >=u B. Constant operands are handled
+ // by CorrelatedValuePropagation using ranges.
+ return I->getType()->isIntegerTy() && !I->hasNoUnsignedWrap() &&
+ !isa<Constant>(I->getOperand(1));
+ default:
+ return false;
+ }
+}
+
+/// Try to strengthen \p I's poison generating flags using \p Info. Returns
+/// true if \p I was modified.
+static bool tryToStrengthenFlags(Instruction *I, ConstraintInfo &Info,
+ SmallVectorImpl<Instruction *> &ToRemove) {
+ assert(canStrengthenFlags(I) && "not a candidate for flag strengthening");
+
+ switch (I->getOpcode()) {
+ case Instruction::Sub: {
+ // Op0 - Op1 does not wrap unsigned, if Op0 >=u Op1.
+ if (!Info.doesHold(CmpInst::ICMP_UGE, I->getOperand(0), I->getOperand(1)))
+ return false;
+ LLVM_DEBUG(dbgs() << "Adding nuw to " << *I << "\n");
+ I->setHasNoUnsignedWrap();
+ return true;
+ }
+ default:
+ return false;
+ }
+}
+
void State::addInfoFor(BasicBlock &BB) {
addBoundsForHeaderInductions(BB);
addInfoForInductions(BB);
@@ -1404,6 +1438,11 @@ void State::addInfoFor(BasicBlock &BB) {
WorkList.push_back(FactOrCheck::getInstFact(DT.getNode(&BB), BO));
}
+ // Queue instructions whose flags may be strengthened based on the facts
+ // that hold on entry to BB.
+ if (canStrengthenFlags(&I))
+ WorkList.push_back(FactOrCheck::getCheck(DT.getNode(&BB), &I));
+
GuaranteedToExecute &= isGuaranteedToTransferExecutionToSuccessor(&I);
}
@@ -2158,6 +2197,10 @@ static bool eliminateConstraints(Function &F, DominatorTree &DT, LoopInfo &LI,
Instruction *Inst = CB.getInstructionToSimplify();
if (!Inst)
continue;
+ if (canStrengthenFlags(Inst)) {
+ Changed |= tryToStrengthenFlags(Inst, Info, ToRemove);
+ continue;
+ }
LLVM_DEBUG(dbgs() << "Processing condition to simplify: " << *Inst
<< "\n");
if (auto *II = dyn_cast<WithOverflowInst>(Inst)) {
diff --git a/llvm/test/Transforms/ConstraintElimination/induction-condition-in-loop-exit-postinc.ll b/llvm/test/Transforms/ConstraintElimination/induction-condition-in-loop-exit-postinc.ll
index 3d509158c523f..1fa5344c9911d 100644
--- a/llvm/test/Transforms/ConstraintElimination/induction-condition-in-loop-exit-postinc.ll
+++ b/llvm/test/Transforms/ConstraintElimination/induction-condition-in-loop-exit-postinc.ll
@@ -997,7 +997,7 @@ define i64 @latch_postdec_umin_clamp(ptr %s, i64 %n) {
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], 0
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]]
; CHECK: [[IF_FOUND]]:
-; CHECK-NEXT: [[IDX:%.*]] = sub i64 [[N]], [[IV]]
+; CHECK-NEXT: [[IDX:%.*]] = sub nuw i64 [[N]], [[IV]]
; CHECK-NEXT: br label %[[EXIT]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: [[RES:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[N]], %[[LOOP_LATCH]] ], [ [[IDX]], %[[IF_FOUND]] ]
diff --git a/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll b/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
index 6dfa70e044638..d27ff1b5e3068 100644
--- a/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
+++ b/llvm/test/Transforms/ConstraintElimination/materialize-flags.ll
@@ -74,7 +74,7 @@ define i32 @sub_nuw_relational(i32 %a, i32 %b) {
; CHECK-NEXT: [[C:%.*]] = icmp uge i32 [[A]], [[B]]
; CHECK-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[EXIT:.*]]
; CHECK: [[THEN]]:
-; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A]], [[B]]
+; CHECK-NEXT: [[SUB:%.*]] = sub nuw i32 [[A]], [[B]]
; CHECK-NEXT: ret i32 [[SUB]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: ret i32 0
@@ -101,7 +101,7 @@ define i32 @sub_nuw_transitive(i32 %a, i32 %b, i32 %c) {
; CHECK-NEXT: [[C_1:%.*]] = icmp uge i32 [[B]], [[C]]
; CHECK-NEXT: br i1 [[C_1]], label %[[BB_2:.*]], label %[[EXIT]]
; CHECK: [[BB_2]]:
-; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A]], [[C]]
+; CHECK-NEXT: [[SUB:%.*]] = sub nuw i32 [[A]], [[C]]
; CHECK-NEXT: ret i32 [[SUB]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: ret i32 0
diff --git a/llvm/test/Transforms/ConstraintElimination/reproducer-remarks.ll b/llvm/test/Transforms/ConstraintElimination/reproducer-remarks.ll
index 44dc4e0f690cc..5fa4e2e9f8c06 100644
--- a/llvm/test/Transforms/ConstraintElimination/reproducer-remarks.ll
+++ b/llvm/test/Transforms/ConstraintElimination/reproducer-remarks.ll
@@ -271,7 +271,7 @@ define i1 @shared_operand_no_const(i8 %a, i8 %b) {
; CHECK-NEXT: entry:
; CHECK-NEXT: %0 = icmp ule i8 %b, %a
; CHECK-NEXT: call void @llvm.assume(i1 %0)
-; CHECK-NEXT: %sub = sub i8 %a, %b
+; CHECK-NEXT: %sub = sub nuw i8 %a, %b
; CHECK-NEXT: %sub.2 = sub nuw i8 %sub, 0
; CHECK-NEXT: %c.5 = icmp ult i8 %sub.2, %sub
; CHECK-NEXT: ret i1 %c.5
diff --git a/llvm/test/Transforms/ConstraintElimination/sub.ll b/llvm/test/Transforms/ConstraintElimination/sub.ll
index df759e30365ec..751a6d72b4e6b 100644
--- a/llvm/test/Transforms/ConstraintElimination/sub.ll
+++ b/llvm/test/Transforms/ConstraintElimination/sub.ll
@@ -253,7 +253,7 @@ define i1 @sub_no_wrap_flags_ule(i8 %a, i8 %b) {
; CHECK-NEXT: call void @use(i1 [[C_NEG]])
; CHECK-NEXT: [[PRECOND:%.*]] = icmp ule i8 [[B]], [[A]]
; CHECK-NEXT: call void @llvm.assume(i1 [[PRECOND]])
-; CHECK-NEXT: [[SUB:%.*]] = sub i8 [[A]], [[B]]
+; CHECK-NEXT: [[SUB:%.*]] = sub nuw i8 [[A]], [[B]]
; CHECK-NEXT: ret i1 true
;
entry:
More information about the llvm-commits
mailing list