On 05/28/2014 04:32 PM, Richard Sandiford wrote:
> While working on patches to speed up the handling of constraints,
> I hit some behaviour in ira_get_dup_out_num that looked unintentional:
>
> - the check for output operands was part of the !ignored_p condition
> so would be skipped if the first alternative is disabled/excluded.
>
> - the first disabled/excluded alternative stops all following alternatives
> from being processed, since we get "stuck" in the first part of the
> "if" statement and never increment curr_alt.
>
> This seems to have some effect on the testsuite. E.g. at -O2
> gcc.c-torture/compile/20071117-1.c has changes like:
>
> .LCFI2:
> movq %rsp, %rbx
> subq %rax, %rsp
> - leaq 15(%rsp), %rax
> - andq $-16, %rax
> - movq %rax, %rdi
> + leaq 15(%rsp), %rdi
> + andq $-16, %rdi
> call bar
> xorl %esi, %esi
> movq %rbx, %rsp
>
> There are also some cases where the change introduces a move though.
> E.g. gcc.c-torture/compat/struct-ic.c has:
>
> movabsq $4294967296, %rdx
> addq $8, %rsp
> .LCFI4:
> - andq %rdi, %rax
> + andq %rax, %rdi
> + movq %rdi, %rax
> orq %rdx, %rax
> ret
> .L9:
>
> But AFAICT the patch is what was originally intended.
>
> Tested on x86_64-linux-gnu. OK to install?
>
>
Ok, Richard. Thanks for fixing this.
