On Thu, Feb 20, 2025 at 01:11:25PM -0800, Andrew Pinski wrote:
> So what is happening here is that after r15-268-g9dbff9c05520a7,
> a move instruction still exists after combine and the register
> allocator choses different register allocation order for the xor
> and because the input operand of lzcntq is not the same as output
> operand, there is an extra xor that happens (due to an errata).
>
> This fixes the testcase by using loading from a pointer instead
> of a function argument directly. The register allocator has more
> freedom since the load has no hard register associated with it (rdi)
> so it can be in eax register right away.
>
> Tested for both -m32 and -m64 on x86_64-linux-gnu.
>
> gcc/testsuite/ChangeLog:
>
> PR testsuite/115028
> * gcc.target/i386/pr101950-2.c:
Please describe what you changed here.
Otherwise LGTM.
>
> Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
> ---
> gcc/testsuite/gcc.target/i386/pr101950-2.c | 13 +++++++++----
> 1 file changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr101950-2.c
> b/gcc/testsuite/gcc.target/i386/pr101950-2.c
> index 896f1b46414..ccc361e3a46 100644
> --- a/gcc/testsuite/gcc.target/i386/pr101950-2.c
> +++ b/gcc/testsuite/gcc.target/i386/pr101950-2.c
> @@ -6,14 +6,19 @@
> /* { dg-final { scan-assembler-times "\txor\[ql]\t" 2 } } */
> /* { dg-final { scan-assembler-times "\tsar\[ql]\t|\tcltd" 2 } } */
>
> +/* Use pointers to avoid register allocation difference due to argument
> + and return register being different and the difference in selecting eax
> + for one the result of the xor vs selecting rdi due to the order of the
> + shift vs the not shift. */
> +
> int
> -foo (long x)
> +foo (long *x)
> {
> - return __builtin_clrsbl (x);
> + return __builtin_clrsbl (*x);
> }
>
> int
> -bar (int x)
> +bar (int *x)
> {
> - return __builtin_clrsb (x);
> + return __builtin_clrsb (*x);
> }
> --
> 2.43.0
Jakub
