On 10/02/2025 08:37, Jin Ma wrote:
On Sun, 09 Feb 2025 14:04:00 +0800, Jin Ma wrote:
PR target/118601
Ok for trunk?
Best regards,
Jin Ma
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_use_by_pieces_infrastructure_p):
Exclude XTheadVector.
Reported-by: Edwin Lu <e...@rivosinc.com>
---
gcc/config/riscv/riscv.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 819e1538741..e5776aa0fbe 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -13826,7 +13826,7 @@ riscv_use_by_pieces_infrastructure_p (unsigned
HOST_WIDE_INT size,
/* For set/clear with size > UNITS_PER_WORD, by pieces uses vector
broadcasts
with UNITS_PER_WORD size pieces. Use setmem<mode> instead which can use
bigger chunks. */
- if (TARGET_VECTOR && stringop_strategy & STRATEGY_VECTOR
+ if (TARGET_VECTOR && !TARGET_XTHEADVECTOR && stringop_strategy &
STRATEGY_VECTOR
&& (op == CLEAR_BY_PIECES || op == SET_BY_PIECES)
&& speed_p && size > UNITS_PER_WORD)
return false;
`riscv_vector::expand_vec_setmem` generates the unrecognizable
instruction and your patch avoids calling it in some, but not all,
cases. Here is a case that still ICEs with `-march=rv32gc_xtheadvector
-mabi=ilp32d` and `-march=rv64gc_xtheadvector -mabi=lp64d` after
applying your patch:
```
void foo1_16 (void *p)
{
__builtin_memset (p, 1, 16);
}
```
I suggest returning `false` in `riscv_vector::expand_vec_setmem` for
`TARGET_XTHEADVECTOR` or trying to generate something that is valid for
`TARGET_XTHEADVECTOR`. If you do bail out of
`riscv_vector::expand_vec_setmem` then you probably want to keep your
existing change too so that by pieces is still used for smaller lengths
rather than a libcall.
--
2.25.1
