| Issue |
135353
|
| Summary |
[clang] Clang is unable to generate maddubs_epi16 intrinsics.
|
| Labels |
clang
|
| Assignees |
|
| Reporter |
topitopotin
|
Clang 20.1.0 is unable to generate `vpmaddubsw`. Tested with the following code stub on Compiler Explorer with compiling options `-O2 -march=znver4`
```c
#include <cstdint>
#define MAX(a,b) (((a)>(b))?(a):(b))
#define MIN(a,b) (((a)<(b))?(a):(b))
#define CLIP(low,high,value) MAX((low),MIN((high),(value)))
void maddubs(const uint8_t* const __restrict in_a, const int8_t* const __restrict in_b, int16_t* __restrict result)
{
int d = 64;
for (int i = 0, dst = 0; i < d; i += 2, dst++){
result[dst] = CLIP(INT16_MIN, INT16_MAX, (in_a[i + 0] * in_b[i + 0] + in_a[i + 1] * in_b[i + 1]));
}
}
void madd(const int16_t* const __restrict in_a, const int16_t* const __restrict in_b, int32_t* __restrict result)
{
int d = 32;
for (int i = 0, dst = 0; i < d; i += 2, dst++){
result[dst] = CLIP(INT32_MIN, INT32_MAX, (in_a[i + 0] * in_b[i + 0] + in_a[i + 1] * in_b[i + 1]));
}
}
```
The madd function vectorizes properly with `vpmaddwd` as expected.
https://godbolt.org/z/fMdcP8hjP
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
