Paul Hua <paul.hua...@gmail.com> writes:
> The gcc.c-torture/execute/scal-to-vec1.c trigger a gcc ICE bug.
>
> It's a mistake in define_expand vec_setv4hi in loongson.md file.
>
> 375 (define_expand "vec_setv4hi"
> 376 [(set (match_operand:V4HI 0 "register_operand" "=f")
> 377 (unspec:V4HI [(match_operand:V4HI 1 "register_operand" "f")
> 378 (match_operand:HI 2 "register_operand" "f")
> 379 (match_operand:SI 3 "const_0_to_3_operand"
> "")]
> 380 UNSPEC_LOONGSON_PINSRH))]
> 381 "TARGET_HARD_FLOAT && TARGET_LOONGSON_VECTORS"
> 382 {
> 383 rtx ext = gen_reg_rtx (SImode);
> 384 emit_move_insn (ext, gen_lowpart (SImode, operands[1]));
> 385 operands[1] = ext;
> 386 })
>
> The line 384 gen_lowpart the operands[1], should be gen_lowpart
> operands[2], cause the operands[2] are HImode.
>
>
> The attached patch fixed this bug.
>
> Bootstrapped and reg-tested on mips64el-unknown-linux-gnu.
> Ok for commit ?
>
>
> -------------------------------
Hi Paul,
This looks good, just one issue with the changelog entry. The entry
would go in the gcc/ChangeLog file and the path is then relative to
the gcc/ directory. The PR should be referenced as target/<num>:
2018-03-24 Chenghua Xu <paul.hua...@gmail.com>
PR target/86076
* config/mips/loongson.md (vec_setv4hi): Gen_lowpart for
operands[2] instead of operands[1].
OK to commit, thanks for finding and fixing. This has been broken
since 2011!
Matthew
