[Bug target/118188] aarch64: worse code with -mtune=grace (vs -mtune=generic) in TSVC s4115

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118188

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-12-29
            Summary|[15 regression] aarch64:    |aarch64: worse code with
                   |30% regression in TSVC      |-mtune=grace (vs
                   |s4115 since                 |-mtune=generic) in TSVC
                   |r15-5565-gdbc38dd9e96a99    |s4115
             Status|UNCONFIRMED                 |NEW
   Target Milestone|15.0                        |---
     Ever confirmed|0                           |1

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> This was a costing model change so this is a target issue.

Definitely since the generic armv9 cost model produces the expected behavior.
that is `-Ofast -march=armv9-a -mcpu=generic` still produces the full SVE code
while `-Ofast -mcpu=grace` does not.


Also it is not a regression in that GCC 14.2.0 produced the non-SVE code with
-mcpu=neoverse-v2 (grace), so did 13.2.0 and 12.2.0 while the generic tuning
didn't.