firefox fails to build with ICE in src/gl.cc with clang-19 #110374
Labels
Comments
|
there were some reports of musttail errors in various issues over the last week, ccing @ostannard as he is working on them. I'm happy to test any patches |
|
@llvm/issue-subscribers-backend-arm Author: None (stefson)
from the build log, I think this is the error:
here is the full compressed build log: build.log.gz and here is the preprocessed sources, and the reproducer: preprocess.tar.gz |
This error doesn't look obviously related to the musttail bugs I have been working on, but if you think it's relevant then you could test with #109943. |
|
yeah I think you're right, the only part of firefox using I could need some help with reducing, never had an issue like that before with llvm - is there a guide somewhere on what to do? |
|
There's a tool called creduce which is very useful for reducing compiler test cases: https://github.com/csmith-project/creduce. Unfortunately it's documentation site has been down for a while, but archive.org has it: https://web.archive.org/web/20201204141133/https://embed.cs.utah.edu/creduce/using/ |
|
One thing |
|
That looks like a really big case that might take a while to reduce from C. With a mixture of -emit-llvm, running it through llc and isolating the function that was going wrong I managed to get it down to this: It fails with |
|
I've compiled a llvm-20 toolchain to check if the error was fixed recently, and it fails as before: |
|
I've compiled weekly snapshots of the 19.x branch for the last days, and have the regression range down to one week: |
|
I narrowed it down to 6b76c1e. |
|
Cc: @nikic |
|
There is a reduced example here: https://godbolt.org/z/aYYTGxWWv |
BaiXilin
pushed a commit
to BaiXilin/llvm-fix-vnni-instr-types
that referenced
this issue
Jan 12, 2025
With range and undef metadata on a call we can have vector AssertZExt generated on a target with no vector operations. The AssertZExt needs to scalarize to a normal `AssertZext tin, ValueType`. I have added AssertSext too, although I do not have a test case. Fixes llvm#110374
tru
pushed a commit
to llvmbot/llvm-project
that referenced
this issue
Jan 13, 2025
With range and undef metadata on a call we can have vector AssertZExt generated on a target with no vector operations. The AssertZExt needs to scalarize to a normal `AssertZext tin, ValueType`. I have added AssertSext too, although I do not have a test case. Fixes llvm#110374 (cherry picked from commit ab9a80a)
tru
pushed a commit
to llvmbot/llvm-project
that referenced
this issue
Jan 14, 2025
With range and undef metadata on a call we can have vector AssertZExt generated on a target with no vector operations. The AssertZExt needs to scalarize to a normal `AssertZext tin, ValueType`. I have added AssertSext too, although I do not have a test case. Fixes llvm#110374 (cherry picked from commit ab9a80a)
tru
pushed a commit
to llvmbot/llvm-project
that referenced
this issue
Jan 14, 2025
With range and undef metadata on a call we can have vector AssertZExt generated on a target with no vector operations. The AssertZExt needs to scalarize to a normal `AssertZext tin, ValueType`. I have added AssertSext too, although I do not have a test case. Fixes llvm#110374 (cherry picked from commit ab9a80a)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
from the build log, I think this is the error:
here is the full compressed build log: build.log.gz
and here is the preprocessed sources, and the reproducer: preprocess.tar.gz
The text was updated successfully, but these errors were encountered: