[InstSimplify] Fold `getelementptr inbounds null, idx -> null` #130742

diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 4c14dcfb4d75f..dcb17ee8d7b43 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2910,6 +2910,13 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
                           SQ.getWithInstruction(&GEP)))
     return replaceInstUsesWith(GEP, V);
 
+  // getelementptr inbounds null, idx -> null
+  if (auto *BaseC = dyn_cast<Constant>(PtrOp))
+    if (GEP.isInBounds() && BaseC->isNullValue() &&
+        !NullPointerIsDefined(GEP.getFunction(),
+                              GEPType->getPointerAddressSpace()))
+      return replaceInstUsesWith(GEP, Constant::getNullValue(GEPType));
+
   // For vector geps, use the generic demanded vector support.
   // Skip if GEP return type is scalable. The number of elements is unknown at
   // compile-time.
diff --git a/llvm/test/Transforms/InstCombine/getelementptr.ll b/llvm/test/Transforms/InstCombine/getelementptr.ll
index ec03d9a2dae2b..c1bd6806eae86 100644
--- a/llvm/test/Transforms/InstCombine/getelementptr.ll
+++ b/llvm/test/Transforms/InstCombine/getelementptr.ll
@@ -1328,8 +1328,7 @@ define ptr @PR45084_extra_use(i1 %cond, ptr %p) {
 
 define ptr @gep_null_inbounds(i64 %idx) {
 ; CHECK-LABEL: @gep_null_inbounds(
-; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds i8, ptr null, i64 [[IDX:%.*]]
-; CHECK-NEXT:    ret ptr [[GEP]]
+; CHECK-NEXT:    ret ptr null
 ;
   %gep = getelementptr inbounds i8, ptr null, i64 %idx
   ret ptr %gep
@@ -1355,8 +1354,7 @@ define ptr @gep_null_defined(i64 %idx) null_pointer_is_valid {
 
 define ptr @gep_null_inbounds_different_type(i64 %idx1, i64 %idx2) {
 ; CHECK-LABEL: @gep_null_inbounds_different_type(
-; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds [0 x i8], ptr null, i64 0, i64 [[IDX2:%.*]]
-; CHECK-NEXT:    ret ptr [[GEP]]
+; CHECK-NEXT:    ret ptr null
 ;
   %gep = getelementptr inbounds [0 x i8], ptr null, i64 %idx1, i64 %idx2
   ret ptr %gep
@@ -2019,5 +2017,38 @@ define ptr @gep_merge_nusw_const(ptr %p, i64 %idx, i64 %idx2) {
   ret ptr %gep
 }
 
+define <2 x ptr> @gep_inbounds_null_vec(i64 %idx) {
+; CHECK-LABEL: @gep_inbounds_null_vec(
+; CHECK-NEXT:    ret <2 x ptr> zeroinitializer
+;
+  %p = getelementptr inbounds i8, <2 x ptr> zeroinitializer, i64 %idx
+  ret <2 x ptr> %p
+}
+
+define <2 x ptr> @gep_inbounds_null_vec_broadcast(<2 x i64> %idx) {
+; CHECK-LABEL: @gep_inbounds_null_vec_broadcast(
+; CHECK-NEXT:    ret <2 x ptr> zeroinitializer
+;
+  %p = getelementptr inbounds i8, ptr null, <2 x i64> %idx
+  ret <2 x ptr> %p
+}
+
+define ptr @gep_noinbounds_null(i64 %idx) {
+; CHECK-LABEL: @gep_noinbounds_null(
+; CHECK-NEXT:    [[P:%.*]] = getelementptr i8, ptr null, i64 [[IDX:%.*]]
+; CHECK-NEXT:    ret ptr [[P]]
+;
+  %p = getelementptr i8, ptr null, i64 %idx
+  ret ptr %p
+}
+
+define ptr @gep_inbounds_null_null_is_valid(i64 %idx) null_pointer_is_valid {
+; CHECK-LABEL: @gep_inbounds_null_null_is_valid(
+; CHECK-NEXT:    [[P:%.*]] = getelementptr inbounds i8, ptr null, i64 [[IDX:%.*]]
+; CHECK-NEXT:    ret ptr [[P]]
+;
+  %p = getelementptr inbounds i8, ptr null, i64 %idx
+  ret ptr %p
+}
 
 !0 = !{!"branch_weights", i32 2, i32 10}
diff --git a/llvm/test/Transforms/InstCombine/store.ll b/llvm/test/Transforms/InstCombine/store.ll
index daa40da1828b5..48c63c6f24c72 100644
--- a/llvm/test/Transforms/InstCombine/store.ll
+++ b/llvm/test/Transforms/InstCombine/store.ll
@@ -49,8 +49,7 @@ define void @test2(ptr %P) {
 
 define void @store_at_gep_off_null_inbounds(i64 %offset) {
 ; CHECK-LABEL: @store_at_gep_off_null_inbounds(
-; CHECK-NEXT:    [[PTR:%.*]] = getelementptr inbounds i32, ptr null, i64 [[OFFSET:%.*]]
-; CHECK-NEXT:    store i32 poison, ptr [[PTR]], align 4
+; CHECK-NEXT:    store i32 poison, ptr null, align 4
 ; CHECK-NEXT:    ret void
 ;
   %ptr = getelementptr inbounds i32, ptr null, i64 %offset
diff --git a/llvm/test/Transforms/InstCombine/sub-gep.ll b/llvm/test/Transforms/InstCombine/sub-gep.ll
index 3f8728d3a4381..c86a1a37bd7ad 100644
--- a/llvm/test/Transforms/InstCombine/sub-gep.ll
+++ b/llvm/test/Transforms/InstCombine/sub-gep.ll
@@ -741,7 +741,7 @@ define i64 @nullptrtoint_scalable_c() {
 ; CHECK-NEXT:    ret i64 [[PTR_IDX]]
 ;
 entry:
-  %ptr = getelementptr inbounds <vscale x 4 x i32>, ptr null, i64 8
+  %ptr = getelementptr nusw <vscale x 4 x i32>, ptr null, i64 8
   %ret = ptrtoint ptr %ptr to i64
   ret i64 %ret
 }
@@ -755,7 +755,7 @@ define i64 @nullptrtoint_scalable_x(i64 %x) {
 ; CHECK-NEXT:    ret i64 [[PTR_IDX]]
 ;
 entry:
-  %ptr = getelementptr inbounds <vscale x 4 x i32>, ptr null, i64 %x
+  %ptr = getelementptr nusw <vscale x 4 x i32>, ptr null, i64 %x
   %ret = ptrtoint ptr %ptr to i64
   ret i64 %ret
 }

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

llvm/test/CodeGen/AMDGPU/chain-hi-to-lo.ll

llvm/test/CodeGen/AMDGPU/memcpy-crash-issue63986.ll

nikic

LGTM once the clang changes have landed.

llvm/test/CodeGen/AMDGPU/chain-hi-to-lo.ll

…35305) #130742 (comment)

…R130742 (#135305) llvm/llvm-project#130742 (comment)

dtcxzyw · 2025-04-11T06:52:16Z

Failed Tests

Clang
Clang.CodeGenOpenCL/spir32_target.cl
Clang.CodeGenOpenCL/spir64_target.cl
Clang.CodeGenOpenCL/spirv_target.cl
LLVM
LLVM.CodeGen/Hexagon/64bit_tstbit.ll
LLVM.CodeGen/Hexagon/always-ext.ll
LLVM.CodeGen/Hexagon/autohvx/isel-concat-multiple.ll
LLVM.CodeGen/Hexagon/autohvx/isel-q-legalization-loop.ll
LLVM.CodeGen/Hexagon/autohvx/vector-align-terminator.ll
LLVM.CodeGen/Hexagon/autohvx/vector-align-use-in-different-block.ll
LLVM.CodeGen/Hexagon/reg-scavengebug-2.ll
LLVM.CodeGen/Hexagon/swp-const-tc1.ll
LLVM.CodeGen/PowerPC/loop-instr-form-non-inc.ll
LLVM.CodeGen/PowerPC/p10-spill-crgt.ll
LLVM.CodeGen/PowerPC/pr43527.ll
LLVM.CodeGen/PowerPC/sms-cpy-1.ll
LLVM.CodeGen/PowerPC/sms-phi.ll
LLVM.Transforms/LoopUnroll/PowerPC/p10-respect-unroll-pragma.ll

Needed by #130742.

…id UB. NFC. (#135602) Needed by #130742.

dtcxzyw · 2025-04-30T17:12:03Z

If this is causing a lot of issues can we possibly revert it while the fixes make their way through review, and then re-merge it with the fixes?

Sorry, we cannot add a workaround for the SDL2 case. The base pointer is not a constant null, so that clang doesn't know if it is safe to set inbounds flag. The best solution is to replace the pointer addition with an integer addition.

nikic · 2025-04-30T17:51:56Z

If this is causing a lot of issues can we possibly revert it while the fixes make their way through review, and then re-merge it with the fixes?

@efriedma-quic @nikic I am ok to just add a ubsan check in clang 21. Then we can reland this patch in llvm 22/23.

I'm probably missing something, but don't we already have a ubsan check for this for a long time? https://clang.godbolt.org/z/vGr7qYYjr

jyknight · 2025-04-30T17:56:08Z

Is there a real-world optimization benefit derived from this fold? Or is it just added because it's an obviously-correct simplification?

dtcxzyw · 2025-04-30T17:56:55Z

If this is causing a lot of issues can we possibly revert it while the fixes make their way through review, and then re-merge it with the fixes?

@efriedma-quic @nikic I am ok to just add a ubsan check in clang 21. Then we can reland this patch in llvm 22/23.

I'm probably missing something, but don't we already have a ubsan check for this for a long time? https://clang.godbolt.org/z/vGr7qYYjr

We have a check for pointer arithmetic, but not for member access.

dtcxzyw · 2025-04-30T17:59:18Z

Is there a real-world optimization benefit derived from this fold? Or is it just added because it's an obviously-correct simplification?

See #130734 (comment) and https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2204/files. This patch is useful to eliminate unreachable code introduced by passes like JumpThreading.

nikic · 2025-04-30T18:02:42Z

If this is causing a lot of issues can we possibly revert it while the fixes make their way through review, and then re-merge it with the fixes?

@efriedma-quic @nikic I am ok to just add a ubsan check in clang 21. Then we can reland this patch in llvm 22/23.

I'm probably missing something, but don't we already have a ubsan check for this for a long time? https://clang.godbolt.org/z/vGr7qYYjr

We have a check for pointer arithmetic, but not for member access.

Right, but are any of the problematic cases (beyond the ones we already catch directly in clang) related to member access?

efriedma-quic · 2025-04-30T18:09:13Z

xalancbmk is a member access issue. The glibc issue and the sdl2 issue are pointer arithmetic. I don't think I've seen any other issues reported.

dtcxzyw · 2025-04-30T18:16:20Z

xalancbmk is a member access issue. The glibc issue and the sdl2 issue are pointer arithmetic. I don't think I've seen any other issues reported.

Right. I only checked the member access issue before drafting this patch (hwloc/slurm/luajit/llvm). They have been covered by #130734 and #130952.

DavidTruby · 2025-04-30T23:59:31Z

I didn't realise the recent xalancbmk failures we've been seeing are the same issue here (I hadn't got around to looking into that yet!)
Given the variety of failures I'd like to push a bit harder for a revert until fixes can be found for these failures

dtcxzyw · 2025-05-01T09:02:52Z

We've found that this also breaks SDL2, specifically somewhere in SDL_render_gles2.c. Unfortunately I haven't been able to narrow down the failure yet, I didn't see any obvious use of calculating offsets from nullptr or anything like that. The fix in #137851 didn't work for it.

FYI the fix has been upstreamed: libsdl-org/SDL@14deef9

tblah · 2025-05-01T14:29:52Z

I also think this should be reverted until there is a fix. It will take years until some distributions stop shipping broken versions of glibc.

nikic · 2025-05-01T14:39:27Z

I'm ok with reverting this until the __PTR_ALIGN workarounds in clang have landed. We should reapply directly after that though, to ensure more of these issues get caught as soon as possible.

dtcxzyw · 2025-05-01T16:45:28Z

I am preparing for a revert patch.

llvm#130742)"

#130742)" (#138168) Revert #130742 for now to avoid breaking glibc failures until the workaround patches are landed.

This works around undefined behavior that is now being exploited by clang. See llvm/llvm-project#130742

…130742) Proof: https://alive2.llvm.org/ce/z/5ZkPx- See also llvm#130734 for the motivation.

llvm#130742)" (llvm#138168) Revert llvm#130742 for now to avoid breaking glibc failures until the workaround patches are landed.

…130742) Proof: https://alive2.llvm.org/ce/z/5ZkPx- See also llvm#130734 for the motivation.

llvm#130742)" (llvm#138168) Revert llvm#130742 for now to avoid breaking glibc failures until the workaround patches are landed.

This commit (and its corresponding merge commit) broke several codegen tests. ``` commit 4b59b95 Merge: c5ca745 c37b254 Author: git apple-llvm automerger <am@git-apple-llvm> Date: Fri May 2 23:22:29 2025 -0700 Merge commit 'c37b2549ff01' from llvm.org/main into next commit c37b254 error: cannot run gpg: No such file or directory Author: Yingwei Zheng <[email protected]> Date: Fri May 2 05:21:59 2025 +0800 Revert "[InstSimplify] Fold getelementptr inbounds null, idx -> null (llvm#130742)" (llvm#138168) Revert llvm#130742 for now to avoid breaking glibc failures until the workaround patches are landed. ``` This commit re-generates the failing codegen tests. rdar://150697626

llvm#130742)" (llvm#138168) Revert llvm#130742 for now to avoid breaking glibc failures until the workaround patches are landed.

This was referenced Mar 11, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

pre-commit: PR130742 dtcxzyw/llvm-opt-benchmark#2204

Closed

dtcxzyw mentioned this pull request Mar 12, 2025

[Clang][CodeGen] Do not set inbounds flag for struct GEP with null base pointers #130734

Merged

dtcxzyw requested a review from efriedma-quic March 12, 2025 08:38

dtcxzyw marked this pull request as ready for review March 12, 2025 12:21

dtcxzyw requested a review from nikic as a code owner March 12, 2025 12:21

llvmbot added llvm:instcombine llvm:transforms labels Mar 12, 2025

nikic reviewed Mar 12, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp Outdated Show resolved Hide resolved

llvmbot added backend:AMDGPU llvm:analysis labels Mar 12, 2025

arsenm reviewed Mar 12, 2025

View reviewed changes

llvm/test/CodeGen/AMDGPU/chain-hi-to-lo.ll Outdated Show resolved Hide resolved

llvm/test/CodeGen/AMDGPU/memcpy-crash-issue63986.ll Outdated Show resolved Hide resolved

nikic approved these changes Mar 13, 2025

View reviewed changes

llvm/test/CodeGen/AMDGPU/chain-hi-to-lo.ll Outdated Show resolved Hide resolved

efriedma-quic mentioned this pull request Mar 21, 2025

-fwrapv-pointer doesn't affect member accesses #132449

Closed

dtcxzyw force-pushed the perf/gep-null-inbounds branch from 9ec97f4 to 1870109 Compare April 11, 2025 02:53

dtcxzyw mentioned this pull request Apr 11, 2025

[AMDGPU][InstCombine][InstSimplify] Pre-commit tests for PR130742 #135305

Merged

dtcxzyw added a commit that referenced this pull request Apr 11, 2025

[AMDGPU][InstCombine][InstSimplify] Pre-commit tests for PR130742 (#1…

db27a0a

…35305) #130742 (comment)

dtcxzyw force-pushed the perf/gep-null-inbounds branch from 1870109 to 471662e Compare April 11, 2025 05:02

llvm-sync bot pushed a commit to arm/arm-toolchain that referenced this pull request Apr 11, 2025

Automerge: [AMDGPU][InstCombine][InstSimplify] Pre-commit tests for P…

8a19e5b

…R130742 (#135305) llvm/llvm-project#130742 (comment)

This was referenced Apr 14, 2025

[Clang][CodeGen][OpenCL] Add -fno-delete-null-pointer-checks to avoid UB. NFC. #135602

Merged

[Hexagon] Pre-commit tests for PR130742. NFC. #135604

Merged

[PowerPC] Pre-commit tests for PR130742. NFC. #135606

Merged

dtcxzyw changed the title ~~[InstCombine] Fold getelementptr inbounds null, idx -> null~~ [InstSimplify] Fold getelementptr inbounds null, idx -> null Apr 15, 2025

dtcxzyw added a commit that referenced this pull request Apr 17, 2025

[Hexagon] Pre-commit tests for PR130742. NFC. (#135604)

78b37ca

Needed by #130742.

dtcxzyw added a commit that referenced this pull request Apr 17, 2025

[Clang][CodeGen][OpenCL] Add -fno-delete-null-pointer-checks to avo…

8e67d8f

…id UB. NFC. (#135602) Needed by #130742.

dschuff mentioned this pull request May 1, 2025

emscripten: Fix undefined behavior in opengles2 renderer libsdl-org/SDL#12581

Merged

dtcxzyw added a commit to dtcxzyw/llvm-project that referenced this pull request May 1, 2025

Revert "[InstSimplify] Fold getelementptr inbounds null, idx -> null (

8b4527b

llvm#130742)"

dtcxzyw mentioned this pull request May 1, 2025

Revert "[InstSimplify] Fold getelementptr inbounds null, idx -> null (#130742)" #138168

Merged

efriedma-quic pushed a commit that referenced this pull request May 1, 2025

Revert "[InstSimplify] Fold getelementptr inbounds null, idx -> null (

c37b254

#130742)" (#138168) Revert #130742 for now to avoid breaking glibc failures until the workaround patches are landed.

dschuff added a commit to dschuff/emscripten that referenced this pull request May 1, 2025

Build libSDL with -fwrapv-pointer

acbf6b3

This works around undefined behavior that is now being exploited by clang. See llvm/llvm-project#130742

dschuff added a commit to dschuff/emscripten that referenced this pull request May 1, 2025

Build libSDL with -fwrapv-pointer

8e8beec

This works around undefined behavior that is now being exploited by clang. See llvm/llvm-project#130742

dschuff mentioned this pull request May 1, 2025

Build libSDL with -fwrapv-pointer emscripten-core/emscripten#24243

Merged

dschuff added a commit to emscripten-core/emscripten that referenced this pull request May 2, 2025

Build libSDL with -fwrapv-pointer (#24243)

1ae691e

This works around undefined behavior that is now being exploited by clang. See llvm/llvm-project#130742

IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025

[InstSimplify] Fold getelementptr inbounds null, idx -> null (llvm#…

6232a3a

…130742) Proof: https://alive2.llvm.org/ce/z/5ZkPx- See also llvm#130734 for the motivation.

IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025

[InstSimplify] Fold getelementptr inbounds null, idx -> null (llvm#…

625a978

…130742) Proof: https://alive2.llvm.org/ce/z/5ZkPx- See also llvm#130734 for the motivation.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[InstSimplify] Fold `getelementptr inbounds null, idx -> null` #130742

[InstSimplify] Fold `getelementptr inbounds null, idx -> null` #130742

dtcxzyw commented Mar 11, 2025

dtcxzyw commented Mar 12, 2025

dtcxzyw commented Mar 12, 2025

dtcxzyw commented Mar 12, 2025

llvmbot commented Mar 12, 2025 •

edited

Loading

nikic left a comment

dtcxzyw commented Apr 11, 2025

dtcxzyw commented Apr 30, 2025

nikic commented Apr 30, 2025

jyknight commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

nikic commented Apr 30, 2025

efriedma-quic commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

DavidTruby commented Apr 30, 2025

dtcxzyw commented May 1, 2025

tblah commented May 1, 2025

nikic commented May 1, 2025

dtcxzyw commented May 1, 2025

[InstSimplify] Fold getelementptr inbounds null, idx -> null #130742

[InstSimplify] Fold getelementptr inbounds null, idx -> null #130742

Conversation

dtcxzyw commented Mar 11, 2025

dtcxzyw commented Mar 12, 2025

dtcxzyw commented Mar 12, 2025

dtcxzyw commented Mar 12, 2025

llvmbot commented Mar 12, 2025 • edited Loading

nikic left a comment

Choose a reason for hiding this comment

dtcxzyw commented Apr 11, 2025

dtcxzyw commented Apr 30, 2025

nikic commented Apr 30, 2025

jyknight commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

nikic commented Apr 30, 2025

efriedma-quic commented Apr 30, 2025

dtcxzyw commented Apr 30, 2025

DavidTruby commented Apr 30, 2025

dtcxzyw commented May 1, 2025

tblah commented May 1, 2025

nikic commented May 1, 2025

dtcxzyw commented May 1, 2025

[InstSimplify] Fold `getelementptr inbounds null, idx -> null` #130742

[InstSimplify] Fold `getelementptr inbounds null, idx -> null` #130742

llvmbot commented Mar 12, 2025 •

edited

Loading