On 11/28/18 2:47 PM, Vladimir Makarov wrote:
> The patch I committed today recently for
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88207
>
> creates a new regression for pr34256.c. 2 moves is expected but gcc
> with the patch generates 3 moves. I think now RA generates the right code.
>
> We have the following code before RA
>
> (insn 7 6 13 2 (set (reg:V2SI 88)
> (plus:V2SI (reg:V2SI 89 [ x ])
> (mem/c:V2SI (symbol_ref:DI ("y") [flags 0x2] <var_decl
> 0x7faad08b5b40 y>) [1 y+0 S8 A64])))"mmintrin.h":306:18 1115 {*mmx_addv2si3}
>
> (expr_list:REG_DEAD (reg:V2SI 89 [ x ])
> (nil)))
> (insn 13 7 14 2 (set (reg/i:DI 0 ax)
> (subreg:DI (reg:V2SI 88) 0)) "pr34256.c":11:1 66 {*movdi_internal}
> (expr_list:REG_DEAD (reg:V2SI 88)
>
> The test is expected to assign mmx reg to pseudo 88 but gcc with the
> patch assigns memory to it. The cost of mmx to general reg move is 13,
> while overall cost of mmx to mem and mem to general moves is 10. So IRA
> now chooses memory for pseudo 88 according to the minimal cost.
>
> Now, if we want still assign mmx reg to the pseudo 88, we should change
> the costs in machine-dependent x86 code. But I think it might create
> other unexpected code generation. As mmx is basically not used nowadays
> the test is not important, I just propose the following patch.
>
> Is it ok for the trunk?
>
>
>
>
> pr88207-2.patch
>
> Index: testsuite/ChangeLog
> ===================================================================
> --- testsuite/ChangeLog (revision 266582)
> +++ testsuite/ChangeLog (working copy)
> @@ -1,3 +1,7 @@
> +2018-11-28 Vladimir Makarov <vmaka...@redhat.com>
> +
> + * gcc.target/i386/pr34256.c: Adjust the number of expected moves.
OK.
jeff
