| Issue |
172121
|
| Summary |
[compiler-rt] check-builtins is broken when multiple compiler-rt targets are built
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
w2yehia
|
For example, on AIX we build the compiler-rt for 32-bit and 64-bit PowerPC targets.
All LIT tests in compiler-rt/test/builtins/Unit have the following compile+link command:
```
// RUN: %clang_builtins %s %librt -o %t
```
and after LIT substitution (defined in `compiler-rt/test/builtins/Unit/lit.cfg.py`) look like:
```
clang ... -fno-builtin -nodefaultlibs $BUILD/lib/clang/22/lib/<target-triple>/libclang_rt.builtins.a -lc -lm -o ...
```
The path to libclang_rt.builtins.a is computed off the `config.compiler_rt_libdir` field, which has a hardcoded path from cmake:
```
// compiler-rt/test/lit.common.configured.in
...
set_default("compiler_rt_libdir", "@COMPILER_RT_RESOLVED_LIBRARY_OUTPUT_DIR@")
```
So, we end up with using the same, in our case 64-bit, runtime for both 32-bit an 64-bit LIT tests, and that breaks all 32-bit LIT tests.
Other runtimes under compiler-rt, such as `asan` or `profile` don't seem to force the library on the link step (so they leave it to the compiler to choose the correct one from the runtimes built during the build), so this problem seems to be unique to the builtins library.
We considered two solutions:
1) add a fixup logic for `config.compiler_rt_libdir` in `compiler-rt/test/lit.common.configured.in`, which assumes we have a way to compute the correct triple.
2) fix the substitutions so that when `config.enable_per_target_runtime_dir` is ON, we let the compiler choose the runtime (essentially omit `-fno-builtin -nodefaultlibs $BUILD/lib/clang/22/lib/<target-triple>/libclang_rt.builtins.a`).
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
