[llvm-branch-commits] [llvm] [SROA] Use shared memory metadata propagation (PR #243673)

Yonah Goldberg via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Wed Nov 18 01:13:38 PST 2026


https://github.com/YonahGoldberg updated https://github.com/llvm/llvm-project/pull/243673

>From 91cce4d0a6439498edf0879b56d87a09ba3b95fe Mon Sep 17 00:00:00 2001
From: Yonah Goldberg <ygoldberg at nvidia.com>
Date: Wed, 18 Nov 2026 01:02:28 +0000
Subject: [PATCH 1/2] [SROA] Add memory metadata tests

---
 llvm/test/Transforms/SROA/memory-metadata.ll | 223 +++++++++++++++++++
 1 file changed, 223 insertions(+)
 create mode 100644 llvm/test/Transforms/SROA/memory-metadata.ll

diff --git a/llvm/test/Transforms/SROA/memory-metadata.ll b/llvm/test/Transforms/SROA/memory-metadata.ll
new file mode 100644
index 0000000000000..e65a082e252cb
--- /dev/null
+++ b/llvm/test/Transforms/SROA/memory-metadata.ll
@@ -0,0 +1,223 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes=sroa < %s | FileCheck %s --implicit-check-not=!test.unknown --implicit-check-not=!prof
+
+target datalayout = "e-p:64:64-n8:16:32:64"
+
+define i32 @split_copy(ptr %src) {
+;
+; CHECK-LABEL: define i32 @split_copy(
+; CHECK-SAME: ptr [[SRC:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[A_SROA_0_0_COPYLOAD:%.*]] = load i32, ptr [[SRC]], align 8, !tbaa [[INT_TBAA0:![0-9]+]], !alias.scope [[META4:![0-9]+]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7:![0-9]+]], !llvm.access.group [[META7]], !noalias.addrspace [[META8:![0-9]+]]
+; CHECK-NEXT:    store i32 [[A_SROA_0_0_COPYLOAD]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_2_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
+; CHECK-NEXT:    [[A_SROA_2_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_2_0_SRC_SROA_IDX]], align 4, !tbaa [[FLOAT_TBAA9:![0-9]+]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    store i32 [[A_SROA_2_0_COPYLOAD]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
+; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %a = alloca {i32, i32}, align 8
+  call void @llvm.memcpy.p0.p0.i64(ptr align 8 %a, ptr align 8 %src, i64 8, i1 false), !tbaa.struct !19, !prof !24, !mem.cache_hint !22, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  %p1 = getelementptr {i32, i32}, ptr %a, i32 0, i32 1
+  %x = load volatile i32, ptr %a, align 8
+  %y = load volatile i32, ptr %p1, align 4
+  %r = add i32 %x, %y
+  ret i32 %r
+}
+
+define i32 @rewrite_set(i32 %v) {
+;
+; CHECK-LABEL: define i32 @rewrite_set(
+; CHECK-SAME: i32 [[V:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    store i32 [[V]], ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    store i32 0, ptr [[A_SROA_2]], align 4, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
+; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %a = alloca {i32, i32}, align 8
+  %p1 = getelementptr {i32, i32}, ptr %a, i32 0, i32 1
+  store i32 %v, ptr %a, align 8
+  call void @llvm.memset.p0.i64(ptr align 4 %p1, i8 0, i64 4, i1 false), !tbaa !10, !prof !24, !mem.cache_hint !16, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  %x = load volatile i32, ptr %a, align 8
+  %y = load volatile i32, ptr %p1, align 4
+  %r = add i32 %x, %y
+  ret i32 %r
+}
+
+define i32 @split_aggregate_store({i32, i32} %v) {
+;
+; CHECK-LABEL: define i32 @split_aggregate_store(
+; CHECK-SAME: { i32, i32 } [[V:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[V_FCA_0_EXTRACT:%.*]] = extractvalue { i32, i32 } [[V]], 0
+; CHECK-NEXT:    store i32 [[V_FCA_0_EXTRACT]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[V_FCA_1_EXTRACT:%.*]] = extractvalue { i32, i32 } [[V]], 1
+; CHECK-NEXT:    store i32 [[V_FCA_1_EXTRACT]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
+; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %a = alloca {i32, i32}, align 8
+  store {i32, i32} %v, ptr %a, align 8, !tbaa.struct !19, !invariant.group !0, !mem.cache_hint !18, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  %p1 = getelementptr {i32, i32}, ptr %a, i32 0, i32 1
+  %x = load volatile i32, ptr %a, align 8
+  %y = load volatile i32, ptr %p1, align 4
+  %r = add i32 %x, %y
+  ret i32 %r
+}
+
+define {i32, i32} @split_aggregate_load(ptr %src, ptr %dst) {
+;
+; CHECK-LABEL: define { i32, i32 } @split_aggregate_load(
+; CHECK-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_3:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[A_SROA_0_0_COPYLOAD:%.*]] = load i32, ptr [[SRC]], align 1
+; CHECK-NEXT:    store i32 [[A_SROA_0_0_COPYLOAD]], ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_3_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
+; CHECK-NEXT:    [[A_SROA_3_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_3_0_SRC_SROA_IDX]], align 1
+; CHECK-NEXT:    store i32 [[A_SROA_3_0_COPYLOAD]], ptr [[A_SROA_3]], align 4
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V_FCA_0_LOAD:%.*]] = load i32, ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[V_FCA_0_INSERT:%.*]] = insertvalue { i32, i32 } poison, i32 [[A_SROA_0_0_A_SROA_0_0_V_FCA_0_LOAD]], 0
+; CHECK-NEXT:    [[A_SROA_3_0_A_SROA_3_4_V_FCA_1_LOAD:%.*]] = load i32, ptr [[A_SROA_3]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[V_FCA_1_INSERT:%.*]] = insertvalue { i32, i32 } [[V_FCA_0_INSERT]], i32 [[A_SROA_3_0_A_SROA_3_4_V_FCA_1_LOAD]], 1
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    store volatile i32 [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1]], ptr [[DST]], align 1
+; CHECK-NEXT:    [[A_SROA_3_0_DST_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 4
+; CHECK-NEXT:    [[A_SROA_3_0_A_SROA_3_0_COPYLOAD2:%.*]] = load volatile i32, ptr [[A_SROA_3]], align 4
+; CHECK-NEXT:    store volatile i32 [[A_SROA_3_0_A_SROA_3_0_COPYLOAD2]], ptr [[A_SROA_3_0_DST_SROA_IDX]], align 1
+; CHECK-NEXT:    ret { i32, i32 } [[V_FCA_1_INSERT]]
+;
+  %a = alloca {i32, i32}, align 8
+  call void @llvm.memcpy.p0.p0.i64(ptr %a, ptr %src, i64 8, i1 false)
+  %v = load {i32, i32}, ptr %a, align 8, !tbaa.struct !19, !invariant.load !0, !invariant.group !0, !mem.cache_hint !16, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  call void @llvm.memcpy.p0.p0.i64(ptr %dst, ptr %a, i64 8, i1 true)
+  ret {i32, i32} %v
+}
+
+define double @rewrite_load(ptr %src, ptr %dst) {
+;
+; CHECK-LABEL: define double @rewrite_load(
+; CHECK-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca double, align 8
+; CHECK-NEXT:    [[A_SROA_3:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_0_0_COPYLOAD:%.*]] = load double, ptr [[SRC]], align 1
+; CHECK-NEXT:    store double [[A_SROA_0_0_COPYLOAD]], ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_3_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 8
+; CHECK-NEXT:    [[A_SROA_3_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_3_0_SRC_SROA_IDX]], align 1
+; CHECK-NEXT:    store i32 [[A_SROA_3_0_COPYLOAD]], ptr [[A_SROA_3]], align 8
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V:%.*]] = load double, ptr [[A_SROA_0]], align 8, !fpmath [[META11:![0-9]+]], !invariant.load [[META12:![0-9]+]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META13:![0-9]+]], !llvm.mem.parallel_loop_access [[META7]], !invariant.group [[META12]], !llvm.access.group [[META7]], !noundef [[META12]], !mmra [[META14:![0-9]+]], !noalias.addrspace [[META8]], !nofpclass [[META15:![0-9]+]], !mem.cache_hint [[META16:![0-9]+]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1:%.*]] = load volatile double, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    store volatile double [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1]], ptr [[DST]], align 1
+; CHECK-NEXT:    [[A_SROA_3_0_DST_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 8
+; CHECK-NEXT:    [[A_SROA_3_0_A_SROA_3_0_COPYLOAD2:%.*]] = load volatile i32, ptr [[A_SROA_3]], align 8
+; CHECK-NEXT:    store volatile i32 [[A_SROA_3_0_A_SROA_3_0_COPYLOAD2]], ptr [[A_SROA_3_0_DST_SROA_IDX]], align 1
+; CHECK-NEXT:    ret double [[A_SROA_0_0_A_SROA_0_0_V]]
+;
+  %a = alloca {double, i32}, align 8
+  call void @llvm.memcpy.p0.p0.i64(ptr %a, ptr %src, i64 12, i1 false)
+  %v = load double, ptr %a, align 8, !noundef !0, !nofpclass !15, !fpmath !14, !invariant.load !0, !invariant.group !0, !mem.cache_hint !16, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  call void @llvm.memcpy.p0.p0.i64(ptr %dst, ptr %a, i64 12, i1 true)
+  ret double %v
+}
+
+define ptr @rewrite_pointer_load(ptr %v) {
+;
+; CHECK-LABEL: define ptr @rewrite_pointer_load(
+; CHECK-SAME: ptr [[V:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_1:%.*]] = alloca ptr, align 8
+; CHECK-NEXT:    store ptr [[V]], ptr [[A_SROA_1]], align 8
+; CHECK-NEXT:    [[A_SROA_1_0_A_SROA_1_8_R:%.*]] = load volatile ptr, ptr [[A_SROA_1]], align 8, !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META13]], !llvm.mem.parallel_loop_access [[META7]], !nonnull [[META12]], !dereferenceable [[META18:![0-9]+]], !dereferenceable_or_null [[META18]], !invariant.group [[META12]], !align [[META18]], !llvm.access.group [[META7]], !noundef [[META12]], !mmra [[META14]], !noalias.addrspace [[META8]], !mem.cache_hint [[META16]]
+; CHECK-NEXT:    ret ptr [[A_SROA_1_0_A_SROA_1_8_R]]
+;
+  %a = alloca [2 x ptr], align 8
+  %p1 = getelementptr [2 x ptr], ptr %a, i32 0, i32 1
+  store ptr null, ptr %a, align 8
+  store ptr %v, ptr %p1, align 8
+  %r = load volatile ptr, ptr %p1, align 8, !nonnull !0, !noundef !0, !align !25, !dereferenceable !25, !dereferenceable_or_null !25, !invariant.group !0, !mem.cache_hint !16, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  ret ptr %r
+}
+
+define i32 @presplit_load(ptr %src) {
+;
+; CHECK-LABEL: define i32 @presplit_load(
+; CHECK-SAME: ptr [[SRC:%.*]]) {
+; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
+; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[V1:%.*]] = load i32, ptr [[SRC]], align 8, !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]]
+; CHECK-NEXT:    [[SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
+; CHECK-NEXT:    [[V2:%.*]] = load i32, ptr [[SRC_SROA_IDX]], align 4, !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]]
+; CHECK-NEXT:    store i32 [[V1]], ptr [[A_SROA_0]], align 8, !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    store i32 [[V2]], ptr [[A_SROA_2]], align 4, !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
+; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
+; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %a = alloca {i32, i32}, align 8
+  %v = load i64, ptr %src, align 8, !range !26, !noundef !0, !invariant.load !0, !invariant.group !0, !mem.cache_hint !16, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  store i64 %v, ptr %a, align 8, !mem.cache_hint !18, !alias.scope !6, !noalias !6, !nontemporal !1, !llvm.mem.parallel_loop_access !5, !llvm.access.group !5, !annotation !2, !nosanitize !0, !mmra !3, !noalias.addrspace !4, !test.unknown !0
+  %p1 = getelementptr {i32, i32}, ptr %a, i32 0, i32 1
+  %x = load volatile i32, ptr %a, align 8
+  %y = load volatile i32, ptr %p1, align 4
+  %r = add i32 %x, %y
+  ret i32 %r
+}
+
+!0 = !{}
+!1 = !{i32 1}
+!2 = !{!"preserved memory metadata"}
+!3 = !{!"test", !"relaxed"}
+!4 = !{i32 5, i32 6}
+!5 = distinct !{}
+!6 = !{!7}
+!7 = distinct !{!7, !8, !"scope"}
+!8 = distinct !{!8, i1 false, !"domain"}
+!9 = !{i32 0, i32 100}
+!10 = !{!11, !11, i64 0}
+!11 = !{!"int", !12, i64 0}
+!12 = !{!"omnipotent char", !13, i64 0}
+!13 = !{!"root"}
+!14 = !{float 2.5}
+!15 = !{i32 3}
+!16 = !{i32 0, !17}
+!17 = !{!"nvvm.l1_eviction", !"last"}
+!18 = !{i32 1, !17}
+!19 = !{i64 0, i64 4, !10, i64 4, i64 4, !20}
+!20 = !{!21, !21, i64 0}
+!21 = !{!"float", !12, i64 0}
+!22 = !{i32 0, !17, i32 1, !23}
+!23 = !{!"nvvm.l1_eviction", !"first"}
+!24 = !{!"branch_weights", i32 10}
+!25 = !{i64 8}
+!26 = !{i64 0, i64 100}
+;.
+; CHECK: [[INT_TBAA0]] = !{[[META1:![0-9]+]], [[META1]], i64 0}
+; CHECK: [[META1]] = !{!"int", [[META2:![0-9]+]], i64 0}
+; CHECK: [[META2]] = !{!"omnipotent char", [[META3:![0-9]+]], i64 0}
+; CHECK: [[META3]] = !{!"root"}
+; CHECK: [[META4]] = !{[[META5:![0-9]+]]}
+; CHECK: [[META5]] = distinct !{[[META5]], [[META6:![0-9]+]], !"scope"}
+; CHECK: [[META6]] = distinct !{[[META6]], i1 false, !"domain"}
+; CHECK: [[META7]] = distinct !{}
+; CHECK: [[META8]] = !{i32 5, i32 6}
+; CHECK: [[FLOAT_TBAA9]] = !{[[META10:![0-9]+]], [[META10]], i64 0}
+; CHECK: [[META10]] = !{!"float", [[META2]], i64 0}
+; CHECK: [[META11]] = !{float 2.500000e+00}
+; CHECK: [[META12]] = !{}
+; CHECK: [[META13]] = !{i32 1}
+; CHECK: [[META14]] = !{!"test", !"relaxed"}
+; CHECK: [[META15]] = !{i32 3}
+; CHECK: [[META16]] = !{i32 0, [[META17:![0-9]+]]}
+; CHECK: [[META17]] = !{!"nvvm.l1_eviction", !"last"}
+; CHECK: [[META18]] = !{i64 8}
+;.

>From 257eca305ab9331cae39856bdaecdc75b65d7a87 Mon Sep 17 00:00:00 2001
From: Yonah Goldberg <ygoldberg at nvidia.com>
Date: Wed, 18 Nov 2026 01:02:34 +0000
Subject: [PATCH 2/2] [SROA] Use shared memory metadata propagation

---
 llvm/lib/Transforms/Scalar/SROA.cpp          | 131 +++++++------------
 llvm/test/Transforms/SROA/memory-metadata.ll |  58 ++++----
 2 files changed, 75 insertions(+), 114 deletions(-)

diff --git a/llvm/lib/Transforms/Scalar/SROA.cpp b/llvm/lib/Transforms/Scalar/SROA.cpp
index 075ac5818082f..d4caaaa22da54 100644
--- a/llvm/lib/Transforms/Scalar/SROA.cpp
+++ b/llvm/lib/Transforms/Scalar/SROA.cpp
@@ -43,6 +43,7 @@
 #include "llvm/Analysis/DomTreeUpdater.h"
 #include "llvm/Analysis/GlobalsModRef.h"
 #include "llvm/Analysis/Loads.h"
+#include "llvm/Analysis/MetadataUtils.h"
 #include "llvm/Analysis/PtrUseVisitor.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/Analysis/VectorUtils.h"
@@ -407,10 +408,16 @@ static void migrateDebugInfo(AllocaInst *OldAlloca, bool IsSplit,
     BaseFragments[getAggregateVariable(DVR)] =
         DVR->getExpression()->getFragmentInfo();
 
-  // The new inst needs a DIAssignID unique metadata tag (if OldInst has
-  // one). It shouldn't already have one: assert this assumption.
-  assert(!Inst->getMetadata(LLVMContext::MD_DIAssignID));
-  DIAssignID *NewID = nullptr;
+  // Reuse an assignment ID copied for an unsplit store at the same address.
+  auto *NewID =
+      cast_or_null<DIAssignID>(Inst->getMetadata(LLVMContext::MD_DIAssignID));
+  if (NewID) {
+    assert(!IsSplit &&
+           NewID == OldInst->getMetadata(LLVMContext::MD_DIAssignID) &&
+           "unexpected assignment ID on replacement instruction");
+    // Deleting OldInst must not delete the records now linked to Inst.
+    OldInst->setMetadata(LLVMContext::MD_DIAssignID, nullptr);
+  }
   auto &Ctx = Inst->getContext();
   DIBuilder DIB(*OldInst->getModule(), /*AllowUnresolved*/ false);
   assert(OldAlloca->isStaticAlloca());
@@ -477,7 +484,7 @@ static void migrateDebugInfo(AllocaInst *OldAlloca, bool IsSplit,
     } else {
       // The store is not split, simply steal the existing dbg_assign.
       NewAssign = DbgAssign;
-      NewAssign->setAssignId(NewID); // FIXME: Can we avoid generating new IDs?
+      NewAssign->setAssignId(NewID);
       NewAssign->setAddress(Dest);
       if (Value)
         NewAssign->replaceVariableLocationOp(0u, Value);
@@ -3503,8 +3510,8 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     LoadInst *Load =
         IRB.CreateAlignedLoad(NewAllocaTy, &NewAI, NewAI.getAlign(), "load");
 
-    Load->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
-                            LLVMContext::MD_access_group});
+    copyMetadataForMemoryAccess(
+        *Load, LI, int64_t(NewAllocaBeginOffset) - int64_t(BeginOffset));
     return extractVector(IRB, Load, BeginIndex, EndIndex, "vec");
   }
 
@@ -3537,8 +3544,6 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     Value *OldOp = LI.getOperand(0);
     assert(OldOp == OldPtr);
 
-    AAMDNodes AATags = LI.getAAMetadata();
-
     unsigned AS = LI.getPointerAddressSpace();
 
     Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
@@ -3564,15 +3569,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
       if (NewLI->isAtomic())
         NewLI->setAlignment(LI.getAlign());
 
-      // Copy any metadata that is valid for the new load. This may require
-      // conversion to a different kind of metadata, e.g. !nonnull might change
-      // to !range or vice versa.
-      copyMetadataForLoad(*NewLI, LI);
-
-      // Do this after copyMetadataForLoad() to preserve the TBAA shift.
-      if (AATags)
-        NewLI->setAAMetadata(AATags.adjustForAccess(
-            NewBeginOffset - BeginOffset, NewLI->getType(), DL));
+      copyMetadataForMemoryAccess(*NewLI, LI, NewBeginOffset - BeginOffset);
 
       // Try to preserve nonnull metadata
       V = NewLI;
@@ -3594,15 +3591,9 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
           IRB.CreateAlignedLoad(TargetTy, getNewAllocaSlicePtr(IRB, LTy),
                                 getSliceAlign(), LI.isVolatile(), LI.getName());
 
-      if (AATags)
-        NewLI->setAAMetadata(AATags.adjustForAccess(
-            NewBeginOffset - BeginOffset, NewLI->getType(), DL));
-
       if (LI.isVolatile())
         NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
-      NewLI->copyMetadata(LI, {LLVMContext::MD_mem_parallel_loop_access,
-                               LLVMContext::MD_access_group});
-
+      copyMetadataForMemoryAccess(*NewLI, LI, NewBeginOffset - BeginOffset);
       V = NewLI;
       IsPtrAdjusted = true;
     }
@@ -3645,8 +3636,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     return !LI.isVolatile() && !IsPtrAdjusted;
   }
 
-  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
-                                  AAMDNodes AATags) {
+  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
     // Capture V for the purpose of debug-info accounting once it's converted
     // to a vector store.
     Value *OrigV = V;
@@ -3669,11 +3659,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
       V = insertVector(IRB, Old, V, BeginIndex, "vec");
     }
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
-    Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
-                             LLVMContext::MD_access_group});
-    if (AATags)
-      Store->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                  V->getType(), DL));
+    copyMetadataForMemoryAccess(*Store, SI, NewBeginOffset - BeginOffset);
     Pass.DeadInsts.push_back(&SI);
 
     // NOTE: Careful to use OrigV rather than V.
@@ -3683,7 +3669,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     return true;
   }
 
-  bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
+  bool rewriteIntegerStore(Value *V, StoreInst &SI) {
     assert(IntTy && "We cannot extract an integer from the alloca");
     assert(!SI.isVolatile());
     if (DL.getTypeSizeInBits(V->getType()).getFixedValue() !=
@@ -3697,11 +3683,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     }
     V = IRB.CreateBitPreservingCastChain(DL, V, NewAllocaTy);
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
-    Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
-                             LLVMContext::MD_access_group});
-    if (AATags)
-      Store->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                  V->getType(), DL));
+    copyMetadataForMemoryAccess(*Store, SI, NewBeginOffset - BeginOffset);
 
     migrateDebugInfo(&OldAI, IsSplit, NewBeginOffset * 8, SliceSize * 8, &SI,
                      Store, Store->getPointerOperand(),
@@ -3717,7 +3699,6 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     Value *OldOp = SI.getOperand(1);
     assert(OldOp == OldPtr);
 
-    AAMDNodes AATags = SI.getAAMetadata();
     Value *V = SI.getValueOperand();
 
     // Strip all inbounds GEPs and pointer casts to try to dig out any root
@@ -3739,9 +3720,9 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     }
 
     if (VecTy)
-      return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
+      return rewriteVectorizedStoreInst(V, SI, OldOp);
     if (IntTy && V->getType()->isIntegerTy())
-      return rewriteIntegerStore(V, SI, AATags);
+      return rewriteIntegerStore(V, SI);
 
     StoreInst *NewSI;
     if (NewBeginOffset == NewAllocaBeginOffset &&
@@ -3759,11 +3740,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
       NewSI =
           IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(), SI.isVolatile());
     }
-    NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
-                             LLVMContext::MD_access_group});
-    if (AATags)
-      NewSI->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                  V->getType(), DL));
+    copyMetadataForMemoryAccess(*NewSI, SI, NewBeginOffset - BeginOffset);
     if (SI.isVolatile())
       NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
     if (NewSI->isAtomic())
@@ -3942,11 +3919,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     Value *NewPtr = getPtrToNewAI(II.getDestAddressSpace(), II.isVolatile());
     StoreInst *New =
         IRB.CreateAlignedStore(V, NewPtr, NewAI.getAlign(), II.isVolatile());
-    New->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
-                           LLVMContext::MD_access_group});
-    if (AATags)
-      New->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                V->getType(), DL));
+    copyMetadataForMemoryAccess(*New, II, NewBeginOffset - BeginOffset);
 
     migrateDebugInfo(&OldAI, IsSplit, NewBeginOffset * 8, SliceSize * 8, &II,
                      New, New->getPointerOperand(), V, DL);
@@ -4144,11 +4117,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
     } else {
       LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
                                              II.isVolatile(), "copyload");
-      Load->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
-                              LLVMContext::MD_access_group});
-      if (AATags)
-        Load->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                   Load->getType(), DL));
+      copyMetadataForMemoryAccess(*Load, II, NewBeginOffset - BeginOffset);
       Src = Load;
     }
 
@@ -4167,11 +4136,7 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
 
     StoreInst *Store = cast<StoreInst>(
         IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
-    Store->copyMetadata(II, {LLVMContext::MD_mem_parallel_loop_access,
-                             LLVMContext::MD_access_group});
-    if (AATags)
-      Store->setAAMetadata(AATags.adjustForAccess(NewBeginOffset - BeginOffset,
-                                                  Src->getType(), DL));
+    copyMetadataForMemoryAccess(*Store, II, NewBeginOffset - BeginOffset);
 
     APInt Offset(DL.getIndexTypeSizeInBits(DstPtr->getType()), 0);
     if (IsDest) {
@@ -4455,7 +4420,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
   };
 
   struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
-    AAMDNodes AATags;
+    LoadInst *AggLoad;
     // A vector to hold the split components that we want to emit
     // separate fake uses for.
     SmallVector<Value *, 4> Components;
@@ -4464,11 +4429,11 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
     SmallVector<Instruction *, 1> FakeUses;
 
     LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
-                   AAMDNodes AATags, Align BaseAlign, const DataLayout &DL,
+                   LoadInst *AggLoad, Align BaseAlign, const DataLayout &DL,
                    IRBuilderTy &IRB)
         : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign, DL,
                                      IRB),
-          AATags(AATags) {}
+          AggLoad(AggLoad) {}
 
     /// Emit a leaf load of a single value. This is called at the leaves of the
     /// recursive emission to actually load values.
@@ -4482,10 +4447,12 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
 
       APInt Offset(
           DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
-      if (AATags &&
-          GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset))
-        Load->setAAMetadata(
-            AATags.adjustForAccess(Offset.getZExtValue(), Load->getType(), DL));
+      bool KnownOffset =
+          GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset);
+      copyMetadataForMemoryAccess(
+          *Load, *AggLoad,
+          KnownOffset ? std::optional<int64_t>(Offset.getZExtValue())
+                      : std::nullopt);
       // Record the load so we can generate a fake use for this aggregate
       // component.
       Components.push_back(Load);
@@ -4521,7 +4488,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
 
     // We have an aggregate being loaded, split it apart.
     LLVM_DEBUG(dbgs() << "    original: " << LI << "\n");
-    LoadOpSplitter Splitter(&LI, *U, LI.getType(), LI.getAAMetadata(),
+    LoadOpSplitter Splitter(&LI, *U, LI.getType(), &LI,
                             getAdjustedAlignment(&LI, 0), DL, IRB);
     Splitter.recordFakeUses(LI);
     Value *V = PoisonValue::get(LI.getType());
@@ -4535,12 +4502,11 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
 
   struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
     StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
-                    AAMDNodes AATags, StoreInst *AggStore, Align BaseAlign,
-                    const DataLayout &DL, IRBuilderTy &IRB)
+                    StoreInst *AggStore, Align BaseAlign, const DataLayout &DL,
+                    IRBuilderTy &IRB)
         : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
                                       DL, IRB),
-          AATags(AATags), AggStore(AggStore) {}
-    AAMDNodes AATags;
+          AggStore(AggStore) {}
     StoreInst *AggStore;
     /// Emit a leaf store of a single value. This is called at the leaves of the
     /// recursive emission to actually produce stores.
@@ -4560,10 +4526,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
       APInt Offset(
           DL.getIndexSizeInBits(Ptr->getType()->getPointerAddressSpace()), 0);
       GEPOperator::accumulateConstantOffset(BaseTy, GEPIndices, DL, Offset);
-      if (AATags) {
-        Store->setAAMetadata(AATags.adjustForAccess(
-            Offset.getZExtValue(), ExtractValue->getType(), DL));
-      }
+      copyMetadataForMemoryAccess(*Store, *AggStore, Offset.getZExtValue());
 
       // migrateDebugInfo requires the base Alloca. Walk to it from this gep.
       // If we cannot (because there's an intervening non-const or unbounded
@@ -4595,7 +4558,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
 
     // We have an aggregate being stored, split it apart.
     LLVM_DEBUG(dbgs() << "    original: " << SI << "\n");
-    StoreOpSplitter Splitter(&SI, *U, V->getType(), SI.getAAMetadata(), &SI,
+    StoreOpSplitter Splitter(&SI, *U, V->getType(), &SI,
                              getAdjustedAlignment(&SI, 0), DL, IRB);
     Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
     Visited.erase(&SI);
@@ -5043,8 +5006,7 @@ class SliceSplitter {
                        PartPtrTy, BasePtr->getName() + "."),
         getAdjustedAlignment(SplitI, It.getOffset()),
         /*IsVolatile*/ false, SplitI->getName());
-    PLoad->copyMetadata(*SplitI, {LLVMContext::MD_mem_parallel_loop_access,
-                                  LLVMContext::MD_access_group});
+    copyMetadataForMemoryAccess(*PLoad, *SplitI, It.getOffset());
     return PLoad;
   }
 
@@ -5059,13 +5021,8 @@ class SliceSplitter {
                        PartPtrTy, BasePtr->getName() + "."),
         getAdjustedAlignment(SplitI, It.getOffset()),
         /*IsVolatile*/ false);
-    PStore->copyMetadata(*SplitI, {LLVMContext::MD_mem_parallel_loop_access,
-                                   LLVMContext::MD_access_group,
-                                   LLVMContext::MD_DIAssignID});
-    AAMDNodes AATags = SplitI->getAAMetadata();
-    if (AATags)
-      PStore->setAAMetadata(
-          AATags.adjustForAccess(It.getOffset(), StoreVal->getType(), DL));
+    copyMetadataForMemoryAccess(*PStore, *SplitI, It.getOffset());
+    PStore->copyMetadata(*SplitI, LLVMContext::MD_DIAssignID);
 
     return PStore;
   }
diff --git a/llvm/test/Transforms/SROA/memory-metadata.ll b/llvm/test/Transforms/SROA/memory-metadata.ll
index e65a082e252cb..64c7a812324e7 100644
--- a/llvm/test/Transforms/SROA/memory-metadata.ll
+++ b/llvm/test/Transforms/SROA/memory-metadata.ll
@@ -9,11 +9,11 @@ define i32 @split_copy(ptr %src) {
 ; CHECK-SAME: ptr [[SRC:%.*]]) {
 ; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
 ; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
-; CHECK-NEXT:    [[A_SROA_0_0_COPYLOAD:%.*]] = load i32, ptr [[SRC]], align 8, !tbaa [[INT_TBAA0:![0-9]+]], !alias.scope [[META4:![0-9]+]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7:![0-9]+]], !llvm.access.group [[META7]], !noalias.addrspace [[META8:![0-9]+]]
-; CHECK-NEXT:    store i32 [[A_SROA_0_0_COPYLOAD]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_COPYLOAD:%.*]] = load i32, ptr [[SRC]], align 8, !tbaa [[INT_TBAA0:![0-9]+]], !alias.scope [[META4:![0-9]+]], !noalias [[META4]], !nontemporal [[META7:![0-9]+]], !llvm.mem.parallel_loop_access [[META8:![0-9]+]], !llvm.access.group [[META8]], !annotation [[META9:![0-9]+]], !nosanitize [[META10:![0-9]+]], !mmra [[META11:![0-9]+]], !noalias.addrspace [[META12:![0-9]+]], !mem.cache_hint [[META13:![0-9]+]]
+; CHECK-NEXT:    store i32 [[A_SROA_0_0_COPYLOAD]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15:![0-9]+]]
 ; CHECK-NEXT:    [[A_SROA_2_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
-; CHECK-NEXT:    [[A_SROA_2_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_2_0_SRC_SROA_IDX]], align 4, !tbaa [[FLOAT_TBAA9:![0-9]+]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
-; CHECK-NEXT:    store i32 [[A_SROA_2_0_COPYLOAD]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_2_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_2_0_SRC_SROA_IDX]], align 4, !tbaa [[FLOAT_TBAA17:![0-9]+]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META13]]
+; CHECK-NEXT:    store i32 [[A_SROA_2_0_COPYLOAD]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA17]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
 ; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
@@ -35,7 +35,7 @@ define i32 @rewrite_set(i32 %v) {
 ; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
 ; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
 ; CHECK-NEXT:    store i32 [[V]], ptr [[A_SROA_0]], align 8
-; CHECK-NEXT:    store i32 0, ptr [[A_SROA_2]], align 4, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    store i32 0, ptr [[A_SROA_2]], align 4, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
 ; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
@@ -58,9 +58,9 @@ define i32 @split_aggregate_store({i32, i32} %v) {
 ; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
 ; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
 ; CHECK-NEXT:    [[V_FCA_0_EXTRACT:%.*]] = extractvalue { i32, i32 } [[V]], 0
-; CHECK-NEXT:    store i32 [[V_FCA_0_EXTRACT]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    store i32 [[V_FCA_0_EXTRACT]], ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
 ; CHECK-NEXT:    [[V_FCA_1_EXTRACT:%.*]] = extractvalue { i32, i32 } [[V]], 1
-; CHECK-NEXT:    store i32 [[V_FCA_1_EXTRACT]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    store i32 [[V_FCA_1_EXTRACT]], ptr [[A_SROA_2]], align 4, !tbaa [[FLOAT_TBAA17]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
 ; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
@@ -86,9 +86,9 @@ define {i32, i32} @split_aggregate_load(ptr %src, ptr %dst) {
 ; CHECK-NEXT:    [[A_SROA_3_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
 ; CHECK-NEXT:    [[A_SROA_3_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_3_0_SRC_SROA_IDX]], align 1
 ; CHECK-NEXT:    store i32 [[A_SROA_3_0_COPYLOAD]], ptr [[A_SROA_3]], align 4
-; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V_FCA_0_LOAD:%.*]] = load i32, ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V_FCA_0_LOAD:%.*]] = load i32, ptr [[A_SROA_0]], align 8, !tbaa [[INT_TBAA0]], !invariant.load [[META10]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META19:![0-9]+]]
 ; CHECK-NEXT:    [[V_FCA_0_INSERT:%.*]] = insertvalue { i32, i32 } poison, i32 [[A_SROA_0_0_A_SROA_0_0_V_FCA_0_LOAD]], 0
-; CHECK-NEXT:    [[A_SROA_3_0_A_SROA_3_4_V_FCA_1_LOAD:%.*]] = load i32, ptr [[A_SROA_3]], align 4, !tbaa [[FLOAT_TBAA9]], !alias.scope [[META4]], !noalias [[META4]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[A_SROA_3_0_A_SROA_3_4_V_FCA_1_LOAD:%.*]] = load i32, ptr [[A_SROA_3]], align 4, !tbaa [[FLOAT_TBAA17]], !invariant.load [[META10]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META19]]
 ; CHECK-NEXT:    [[V_FCA_1_INSERT:%.*]] = insertvalue { i32, i32 } [[V_FCA_0_INSERT]], i32 [[A_SROA_3_0_A_SROA_3_4_V_FCA_1_LOAD]], 1
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    store volatile i32 [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1]], ptr [[DST]], align 1
@@ -115,7 +115,7 @@ define double @rewrite_load(ptr %src, ptr %dst) {
 ; CHECK-NEXT:    [[A_SROA_3_0_SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 8
 ; CHECK-NEXT:    [[A_SROA_3_0_COPYLOAD:%.*]] = load i32, ptr [[A_SROA_3_0_SRC_SROA_IDX]], align 1
 ; CHECK-NEXT:    store i32 [[A_SROA_3_0_COPYLOAD]], ptr [[A_SROA_3]], align 8
-; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V:%.*]] = load double, ptr [[A_SROA_0]], align 8, !fpmath [[META11:![0-9]+]], !invariant.load [[META12:![0-9]+]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META13:![0-9]+]], !llvm.mem.parallel_loop_access [[META7]], !invariant.group [[META12]], !llvm.access.group [[META7]], !noundef [[META12]], !mmra [[META14:![0-9]+]], !noalias.addrspace [[META8]], !nofpclass [[META15:![0-9]+]], !mem.cache_hint [[META16:![0-9]+]]
+; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_V:%.*]] = load double, ptr [[A_SROA_0]], align 8, !fpmath [[META20:![0-9]+]], !invariant.load [[META10]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !noundef [[META10]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !nofpclass [[META21:![0-9]+]], !mem.cache_hint [[META19]]
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1:%.*]] = load volatile double, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    store volatile double [[A_SROA_0_0_A_SROA_0_0_COPYLOAD1]], ptr [[DST]], align 1
 ; CHECK-NEXT:    [[A_SROA_3_0_DST_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 8
@@ -136,7 +136,7 @@ define ptr @rewrite_pointer_load(ptr %v) {
 ; CHECK-SAME: ptr [[V:%.*]]) {
 ; CHECK-NEXT:    [[A_SROA_1:%.*]] = alloca ptr, align 8
 ; CHECK-NEXT:    store ptr [[V]], ptr [[A_SROA_1]], align 8
-; CHECK-NEXT:    [[A_SROA_1_0_A_SROA_1_8_R:%.*]] = load volatile ptr, ptr [[A_SROA_1]], align 8, !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META13]], !llvm.mem.parallel_loop_access [[META7]], !nonnull [[META12]], !dereferenceable [[META18:![0-9]+]], !dereferenceable_or_null [[META18]], !invariant.group [[META12]], !align [[META18]], !llvm.access.group [[META7]], !noundef [[META12]], !mmra [[META14]], !noalias.addrspace [[META8]], !mem.cache_hint [[META16]]
+; CHECK-NEXT:    [[A_SROA_1_0_A_SROA_1_8_R:%.*]] = load volatile ptr, ptr [[A_SROA_1]], align 8, !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !nonnull [[META10]], !dereferenceable [[META22:![0-9]+]], !dereferenceable_or_null [[META22]], !align [[META22]], !llvm.access.group [[META8]], !noundef [[META10]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META19]]
 ; CHECK-NEXT:    ret ptr [[A_SROA_1_0_A_SROA_1_8_R]]
 ;
   %a = alloca [2 x ptr], align 8
@@ -153,11 +153,11 @@ define i32 @presplit_load(ptr %src) {
 ; CHECK-SAME: ptr [[SRC:%.*]]) {
 ; CHECK-NEXT:    [[A_SROA_0:%.*]] = alloca i32, align 8
 ; CHECK-NEXT:    [[A_SROA_2:%.*]] = alloca i32, align 4
-; CHECK-NEXT:    [[V1:%.*]] = load i32, ptr [[SRC]], align 8, !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]]
+; CHECK-NEXT:    [[V1:%.*]] = load i32, ptr [[SRC]], align 8, !invariant.load [[META10]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !invariant.group [[META10]], !llvm.access.group [[META8]], !noundef [[META10]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META19]]
 ; CHECK-NEXT:    [[SRC_SROA_IDX:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 4
-; CHECK-NEXT:    [[V2:%.*]] = load i32, ptr [[SRC_SROA_IDX]], align 4, !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]]
-; CHECK-NEXT:    store i32 [[V1]], ptr [[A_SROA_0]], align 8, !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
-; CHECK-NEXT:    store i32 [[V2]], ptr [[A_SROA_2]], align 4, !alias.scope [[META4]], !noalias [[META4]], !llvm.mem.parallel_loop_access [[META7]], !llvm.access.group [[META7]], !noalias.addrspace [[META8]]
+; CHECK-NEXT:    [[V2:%.*]] = load i32, ptr [[SRC_SROA_IDX]], align 4, !invariant.load [[META10]], !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !noundef [[META10]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META19]]
+; CHECK-NEXT:    store i32 [[V1]], ptr [[A_SROA_0]], align 8, !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
+; CHECK-NEXT:    store i32 [[V2]], ptr [[A_SROA_2]], align 4, !alias.scope [[META4]], !noalias [[META4]], !nontemporal [[META7]], !llvm.mem.parallel_loop_access [[META8]], !llvm.access.group [[META8]], !annotation [[META9]], !nosanitize [[META10]], !mmra [[META11]], !noalias.addrspace [[META12]], !mem.cache_hint [[META15]]
 ; CHECK-NEXT:    [[A_SROA_0_0_A_SROA_0_0_X:%.*]] = load volatile i32, ptr [[A_SROA_0]], align 8
 ; CHECK-NEXT:    [[A_SROA_2_0_A_SROA_2_4_Y:%.*]] = load volatile i32, ptr [[A_SROA_2]], align 4
 ; CHECK-NEXT:    [[R:%.*]] = add i32 [[A_SROA_0_0_A_SROA_0_0_X]], [[A_SROA_2_0_A_SROA_2_4_Y]]
@@ -208,16 +208,20 @@ define i32 @presplit_load(ptr %src) {
 ; CHECK: [[META4]] = !{[[META5:![0-9]+]]}
 ; CHECK: [[META5]] = distinct !{[[META5]], [[META6:![0-9]+]], !"scope"}
 ; CHECK: [[META6]] = distinct !{[[META6]], i1 false, !"domain"}
-; CHECK: [[META7]] = distinct !{}
-; CHECK: [[META8]] = !{i32 5, i32 6}
-; CHECK: [[FLOAT_TBAA9]] = !{[[META10:![0-9]+]], [[META10]], i64 0}
-; CHECK: [[META10]] = !{!"float", [[META2]], i64 0}
-; CHECK: [[META11]] = !{float 2.500000e+00}
-; CHECK: [[META12]] = !{}
-; CHECK: [[META13]] = !{i32 1}
-; CHECK: [[META14]] = !{!"test", !"relaxed"}
-; CHECK: [[META15]] = !{i32 3}
-; CHECK: [[META16]] = !{i32 0, [[META17:![0-9]+]]}
-; CHECK: [[META17]] = !{!"nvvm.l1_eviction", !"last"}
-; CHECK: [[META18]] = !{i64 8}
+; CHECK: [[META7]] = !{i32 1}
+; CHECK: [[META8]] = distinct !{}
+; CHECK: [[META9]] = !{!"preserved memory metadata"}
+; CHECK: [[META10]] = !{}
+; CHECK: [[META11]] = !{!"test", !"relaxed"}
+; CHECK: [[META12]] = !{i32 5, i32 6}
+; CHECK: [[META13]] = !{i32 0, [[META14:![0-9]+]]}
+; CHECK: [[META14]] = !{!"nvvm.l1_eviction", !"first"}
+; CHECK: [[META15]] = !{i32 1, [[META16:![0-9]+]]}
+; CHECK: [[META16]] = !{!"nvvm.l1_eviction", !"last"}
+; CHECK: [[FLOAT_TBAA17]] = !{[[META18:![0-9]+]], [[META18]], i64 0}
+; CHECK: [[META18]] = !{!"float", [[META2]], i64 0}
+; CHECK: [[META19]] = !{i32 0, [[META16]]}
+; CHECK: [[META20]] = !{float 2.500000e+00}
+; CHECK: [[META21]] = !{i32 3}
+; CHECK: [[META22]] = !{i64 8}
 ;.



More information about the llvm-branch-commits mailing list