[llvm-readobj][offload] Fix llvm-readobj --all on MachO by nikic · Pull Request #175912 · llvm/llvm-project

nikic · 2026-01-14T10:08:56Z

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring.

This regressed in #143342.

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring.

llvmbot · 2026-01-14T10:09:31Z

@llvm/pr-subscribers-llvm-binary-utilities

Author: Nikita Popov (nikic)

Changes

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring.

This regressed in #143342.

Full diff: https://github.com/llvm/llvm-project/pull/175912.diff

2 Files Affected:

(modified) llvm/lib/Object/OffloadBundle.cpp (+3-1)
(added) llvm/test/tools/llvm-readobj/MachO/all.test (+34)

diff --git a/llvm/lib/Object/OffloadBundle.cpp b/llvm/lib/Object/OffloadBundle.cpp
index 046cde8640b49..f006298c1f4b9 100644
--- a/llvm/lib/Object/OffloadBundle.cpp
+++ b/llvm/lib/Object/OffloadBundle.cpp
@@ -188,7 +188,9 @@ Error OffloadBundleFatBin::extractBundle(const ObjectFile &Source) {
 
 Error object::extractOffloadBundleFatBinary(
     const ObjectFile &Obj, SmallVectorImpl<OffloadBundleFatBin> &Bundles) {
-  assert((Obj.isELF() || Obj.isCOFF()) && "Invalid file type");
+  // Ignore unsupported object formats.
+  if (!Obj.isELF() && !Obj.isCOFF())
+    return Error::success();
 
   // Iterate through Sections until we find an offload_bundle section.
   for (SectionRef Sec : Obj.sections()) {
diff --git a/llvm/test/tools/llvm-readobj/MachO/all.test b/llvm/test/tools/llvm-readobj/MachO/all.test
new file mode 100644
index 0000000000000..8916d72614cc5
--- /dev/null
+++ b/llvm/test/tools/llvm-readobj/MachO/all.test
@@ -0,0 +1,34 @@
+# RUN: yaml2obj %s | llvm-readobj --all - | FileCheck %s
+
+# Make sure that --all at least doesn't crash.
+
+# CHECK: Format: Mach-O arm
+# CHECK: Arch: arm
+# CHECK: AddressSize: 32bit
+# CHECK: MachHeader {
+# CHECK:   Magic: Magic (0xFEEDFACE)
+# CHECK:   CpuType: Arm (0xC)
+# CHECK:   CpuSubType: CPU_SUBTYPE_ARM_V7 (0x9)
+# CHECK:   FileType: Relocatable (0x1)
+# CHECK:   NumOfLoadCommands: 0
+# CHECK:   SizeOfLoadCommands: 0
+# CHECK:   Flags [ (0x0)
+# CHECK:   ]
+# CHECK: }
+# CHECK: Sections [
+# CHECK: ]
+# CHECK: Relocations [
+# CHECK: ]
+# CHECK: UnwindInfo not implemented.
+# CHECK: Symbols [
+# CHECK: ]
+
+--- !mach-o
+FileHeader:
+  magic:      0xFEEDFACE
+  cputype:    0xC
+  cpusubtype: 0x9
+  filetype:   0x1
+  ncmds:      0
+  sizeofcmds: 0
+  flags:      0x0

jh7370

A part of me is wondering if this is the right place for the fix, but I'd actually prefer @david-salinas to make that decision. I could see a case that the --all option should only enable Offloading for the file formats that support it (either explicitly or via the ObjDumper class hierarchy in llvm-readobj's code).

Not directly related to this PR, but one thing is for certain: the option shouldn't be set by --all for GNU style output (i.e. llvm-readelf), because GNU readelf doesn't have the option. This is in keeping with things like Addrsig.

jh7370 · 2026-01-14T10:29:36Z

llvm/test/tools/llvm-readobj/MachO/all.test

@@ -0,0 +1,34 @@
+# RUN: yaml2obj %s | llvm-readobj --all - | FileCheck %s
+
+# Make sure that --all at least doesn't crash.


Nit: newer llvm-readobj tests use ## for comments to help them stand out from lit/FileCheck directives.

nikic · 2026-01-14T11:13:18Z

A part of me is wondering if this is the right place for the fix, but I'd actually prefer @david-salinas to make that decision. I could see a case that the --all option should only enable Offloading for the file formats that support it (either explicitly or via the ObjDumper class hierarchy in llvm-readobj's code).

Even if it's not part of --all, surely --offloading shouldn't produce a crash either, even if the object format doesn't support offloading?

jh7370 · 2026-01-14T11:55:56Z

True, in which case it should be either the fix you've got here or the class hierarchy rework (where offloading is only implemented for the supported platforms). As I don't maintain the offloading API, I think it's worth getting someone who does to make the decision.

jhuber6 · 2026-01-14T17:43:03Z

I'm torn on whether or not --all should imply offloading, as it's not truly part of the object format. At a minimum we should just ignore this if it's MachO. Theoretically you could get this to work there, we'd just need someone to decide on a section name and correctly emit the linker magic that lets us iterate the linked & merged section. But there's really not any motivation to do so as I don't think there's a single person doing or caring about GPU compute on Apple systems.

jhuber6

I'm fine with this fix, but if people's would rather omit the offloading flag from this mode I'd be fine there as well.

jh7370

I'm fine with this fix, but if people's would rather omit the offloading flag from this mode I'd be fine there as well.

I think removing it from the --all mode for GNU output is the right thing to do (could you or @david-salinas or someone else with offloading ownership make that change, please?), but as @nikic has already highlighted, that's not enough to fix the issue, so if you're happy with this change, I am too. LGTM.

nikic · 2026-01-15T10:56:16Z

/cherry-pick 7abe5b7

llvmbot · 2026-01-15T11:03:09Z

/pull-request #176133

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring. This regressed in llvm#143342. (cherry picked from commit 7abe5b7)

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring. This regressed in llvm#143342.

Update to LLVM 22 Scheduled release date: Feb 24 1.94 becomes stable: Mar 5 Changes: * Update to rc2, with one patch to work around our outdated illumos sysroot (rust-lang/llvm-project@41256ab). * Update the host toolchain as well, otherwise we lose cross-language LTO, in particular for jemalloc. * Adjust one loongarch assembly test. The split into r and s variants is based on the suggestion in #151134. Depends on: * [x] #151410 * [ ] #150756 * [x] llvm/llvm-project#175190 * [x] llvm/llvm-project#175912 * [x] llvm/llvm-project#175965 * [x] llvm/llvm-project#176195 * [x] llvm/llvm-project#157073 * [x] llvm/llvm-project#176421 * [x] llvm/llvm-project#176925 * [x] llvm/llvm-project#177187

Update to LLVM 22 Scheduled release date: Feb 24 1.94 becomes stable: Mar 5 Changes: * Update to rc2, with one patch to work around our outdated illumos sysroot (rust-lang/llvm-project@41256ab). * Update the host toolchain as well, otherwise we lose cross-language LTO, in particular for jemalloc. * Adjust one loongarch assembly test. The split into r and s variants is based on the suggestion in rust-lang/rust#151134. Depends on: * [x] rust-lang/rust#151410 * [ ] rust-lang/rust#150756 * [x] llvm/llvm-project#175190 * [x] llvm/llvm-project#175912 * [x] llvm/llvm-project#175965 * [x] llvm/llvm-project#176195 * [x] llvm/llvm-project#157073 * [x] llvm/llvm-project#176421 * [x] llvm/llvm-project#176925 * [x] llvm/llvm-project#177187

[llvm-readobj][offload] Fix llvm-readobj --all on MachO

4c654f4

Currently running llvm-readobj --all on any MachO object asserts, because it implies --offload, and the --offload extraction includes an assert on the object format. Instead, we should be silently ignoring.

nikic requested review from david-salinas and jh7370 January 14, 2026 10:08

llvmbot added the llvm:binary-utilities label Jan 14, 2026

nikic mentioned this pull request Jan 14, 2026

Update to LLVM 22 rust-lang/rust#150722

Merged

10 tasks

jh7370 reviewed Jan 14, 2026

View reviewed changes

Adjust comment style

573e05b

jhuber6 approved these changes Jan 14, 2026

View reviewed changes

jh7370 approved these changes Jan 15, 2026

View reviewed changes

nikic merged commit 7abe5b7 into llvm:main Jan 15, 2026
11 checks passed

nikic deleted the llvm-readobj-macho branch January 15, 2026 10:55

nikic added this to the LLVM 22.x Release milestone Jan 15, 2026

github-project-automation bot added this to LLVM Release Status Jan 15, 2026

github-project-automation bot moved this to Needs Triage in LLVM Release Status Jan 15, 2026

github-project-automation bot moved this from Needs Triage to Done in LLVM Release Status Jan 15, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[llvm-readobj][offload] Fix llvm-readobj --all on MachO#175912

[llvm-readobj][offload] Fix llvm-readobj --all on MachO#175912
nikic merged 2 commits intollvm:mainfrom
nikic:llvm-readobj-macho

nikic commented Jan 14, 2026

Uh oh!

llvmbot commented Jan 14, 2026

Uh oh!

jh7370 left a comment

Uh oh!

jh7370 Jan 14, 2026

Uh oh!

nikic commented Jan 14, 2026

Uh oh!

jh7370 commented Jan 14, 2026

Uh oh!

jhuber6 commented Jan 14, 2026 •

edited

Loading

Uh oh!

jhuber6 left a comment

Uh oh!

jh7370 left a comment

Uh oh!

Uh oh!

nikic commented Jan 15, 2026

Uh oh!

llvmbot commented Jan 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		@@ -0,0 +1,34 @@
		# RUN: yaml2obj %s \| llvm-readobj --all - \| FileCheck %s

		# Make sure that --all at least doesn't crash.

Conversation

nikic commented Jan 14, 2026

Uh oh!

llvmbot commented Jan 14, 2026

Uh oh!

jh7370 left a comment

Choose a reason for hiding this comment

Uh oh!

jh7370 Jan 14, 2026

Choose a reason for hiding this comment

Uh oh!

nikic commented Jan 14, 2026

Uh oh!

jh7370 commented Jan 14, 2026

Uh oh!

jhuber6 commented Jan 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jhuber6 left a comment

Choose a reason for hiding this comment

Uh oh!

jh7370 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

nikic commented Jan 15, 2026

Uh oh!

llvmbot commented Jan 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

jhuber6 commented Jan 14, 2026 •

edited

Loading