| Issue |
146887
|
| Summary |
llc crashes with llvm.aarch64.sve.whilelt lowering to invalid <vscale x 1 x s8> type
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
Dasor
|
I've lowered a matmul kernel from MLIR to LLVM IR successfully, using SVE vector operations. However, when attempting to compile the resulting LLVM IR to assembly using llc, I encountered a segmentation fault. The crash appears to happen in the section of the code that uses the `llvm.aarch64.sve.whilelt.nxv4i1.i64` intrinsic.
I'm not entirely sure what's causing this, but I noticed that at some point during the compilation, a SVE vector of type `<vscale x 1 x s8>` appears in the IR. I'm not an SVE expert, but the use of a single-lane vector with 8-bit elements looks suspicious and possibly invalid in this context.
I've isolated a minimal example that reproduces the crash on Compiler Explorer (Godbolt):
https://godbolt.org/z/KWhvcMK61
The code doesn’t do anything meaningful, it's just a reduced version of the original kernel, trimmed to the part that triggers the crash.
Let me know if there's anything else I can provide to help debug this, thanks in advance!
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
