| Issue |
168998
|
| Summary |
[Clang] [AArch64] Runtime CPU feature checks ignored or unused with -march=native or -mcpu=native
|
| Labels |
clang
|
| Assignees |
|
| Reporter |
StDymphna
|
After #160410 fixes host feature detection, LLVM itself correctly avoids illegal instructions on affected hardware during a program's runtime. However, for the compilation phase, Clang still uses static feature sets from AArch64Features.td based solely on CPU name detection.
The issue is presumed to be getAArch64ArchFeaturesFromMcpu() in clang/lib/Driver/ToolChains/Arch/AArch64.cpp either doesn't call or ignores getHostCPUFeatures() results, and only uses static feature sets in AArch64Features.td that are based solely on CPU name/ID. This affects mostly Qualcomm CPUs, as they rebrand ARM Cortex CPUs models that should have SVE, and via firmware, fusing, or some other mechanism, they disable SVE; it also potentially affects OSs configured without SVE support (e.g. Linux kernel with CONFIG_ARM64_SVE=N).
Reproduction:
```c
#include <stdio.h>
int main(void) {
int n = 4096;
float a[n];
float b[n];
int indices[n];
for (int i = 0; i < n; i++) {
a[i] = (float)i;
b[i] = 0.0f;
indices[i] = (n - 1) - i;
}
for (int i = 0; i < n; i++) {
b[i] = a[indices[i]] * 2.0f;
}
printf("b[0] = %f, b[%d] = %f\n", b[0], n - 1, b[n - 1]);
return 0;
}
```
```bash
clang -o example example.c -mcpu=native -O3 -ffp-model=fast
./example # SIGILL on affected systems
```
Related: #130509, #149370
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
