| Message ID | 201105250030.07953.ebotcazou@adacore.com |
|---|---|
| State | New |
| Headers | show |
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 05/24/11 16:30, Eric Botcazou wrote: > The new save slot sharing algorithm has a documented limitation: > > Future work: > > In the fallback case we should iterate backwards across all possible > modes for the save, choosing the largest available one instead of > falling back to the smallest mode immediately. (eg TF -> DF -> SF). That's not new -- I wrote that circa 1992/1993. The whole point behind the changes from ~1992 was to try and use DFmode insns when caller-saving FP regs on the sparc. However, I didn't think it was worth the effort to deal with that case (TF -> DF -> SF). The code tries to build a save/restore insn of MOVE_MAX_WORDS size and if that fails, then it drops back to WORD_SIZE IIRC. > that is annoying for the SPARC when it comes to floating-point code because the > floating-point registers are single (SF) but there is a fully-fledged support > for double (DF) arithmetics in the architecture. So saving registers on an > individual basis really pessimizes here. For example, the size of the object > generated for the Ada unit a-nlcefu.ads at -O2 decreases from 96080 to 95088 > bytes when you pass -fno-ira-share-save-slots. It's the new slot sharing code that doesn't have support for saving larger hunks. Having written the original code to handle larger saves specifically to help sparc, I can certainly understand why the new code is causing you grief :-) > > Experiments have shown that the impact on integer code is null in terms of code > size and negligible in terms of stack usage (-fstack-usage reports 8/16 bytes > increase for most functions). Therefore this patch disables the option by > default for the SPARC. Boostrapped/regtested on SPARC/Solaris, applied on the > mainline and 4.6 branch. > > > Jeff, I'd like to apply it to the 4.5 branch as well, but I need your patch: > > 2011-01-21 Jeff Law <law@redhat.com> > > PR rtl-optimization/41619 > * caller-save.c (setup_save_areas): Break out code to determine > which hard regs are live across calls by examining the reload chains > so that it is always used. > Eliminate code which checked REG_N_CALLS_CROSSED. > > Do you have any objections to me backporting it to the branch? No objections at all. I don't believe there were any follow-up patches and all the change did was make more of the paths through caller-save consistent in how they determined what needed to be saved. jeff -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.11 (GNU/Linux) Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/ iQEcBAEBAgAGBQJN3SihAAoJEBRtltQi2kC7bnsH/10X0fDwfHtt8b16js5nZHZ8 n+f6TPlAuAu1vJ5h4YI7afybMMfBfHAbvTLwtD+f37boreTQU1wizVH4JLC4GgMS KP9vB48mK/wHli0Hze37QAxVcQt8CPCr3d1fJtpVp6CNUp1gzLWkqT2GjUmxTxfX M3qQ7wRot0cfvVDx8upOj3Yr9tih/c/vIm5ez49s8fzha2acSpEB0vFFj3gcx3EO C3Mgu6z1ZVskIP5KOUIV/2EhtXHMoC4dxsodurfvtGafK5gmbaqVSipzZlKj4BSg Oc4XPKAy07/cSxQhx94pYFB8+Jr7TC99Yubgq2v2gJitf+99AW9MCWmnEvLfX2A= =w6DV -----END PGP SIGNATURE-----
> It's the new slot sharing code that doesn't have support for saving > larger hunks. Having written the original code to handle larger saves > specifically to help sparc, I can certainly understand why the new code > is causing you grief :-) Thanks for the historical perspective. :-) > No objections at all. I don't believe there were any follow-up patches > and all the change did was make more of the paths through caller-save > consistent in how they determined what needed to be saved. OK, now backported after bootstrapping/regtesting on i586-suse-linux.
Index: config/sparc/sparc.c =================================================================== --- config/sparc/sparc.c (revision 174058) +++ config/sparc/sparc.c (working copy) @@ -933,6 +933,12 @@ sparc_option_override (void) ? 64 : 32), global_options.x_param_values, global_options_set.x_param_values); + + /* Disable save slot sharing for call-clobbered registers by default. + The IRA sharing algorithm works on single registers only and this + pessimizes for double floating-point registers. */ + if (!global_options_set.x_flag_ira_share_save_slots) + flag_ira_share_save_slots = 0; } /* Miscellaneous utilities. */
The new save slot sharing algorithm has a documented limitation: Future work: In the fallback case we should iterate backwards across all possible modes for the save, choosing the largest available one instead of falling back to the smallest mode immediately. (eg TF -> DF -> SF). that is annoying for the SPARC when it comes to floating-point code because the floating-point registers are single (SF) but there is a fully-fledged support for double (DF) arithmetics in the architecture. So saving registers on an individual basis really pessimizes here. For example, the size of the object generated for the Ada unit a-nlcefu.ads at -O2 decreases from 96080 to 95088 bytes when you pass -fno-ira-share-save-slots. Experiments have shown that the impact on integer code is null in terms of code size and negligible in terms of stack usage (-fstack-usage reports 8/16 bytes increase for most functions). Therefore this patch disables the option by default for the SPARC. Boostrapped/regtested on SPARC/Solaris, applied on the mainline and 4.6 branch. Jeff, I'd like to apply it to the 4.5 branch as well, but I need your patch: 2011-01-21 Jeff Law <law@redhat.com> PR rtl-optimization/41619 * caller-save.c (setup_save_areas): Break out code to determine which hard regs are live across calls by examining the reload chains so that it is always used. Eliminate code which checked REG_N_CALLS_CROSSED. Do you have any objections to me backporting it to the branch? 2011-05-24 Eric Botcazou <ebotcazou@adacore.com> * config/sparc/sparc.c (sparc_option_override): If not set by the user, force flag_ira_share_save_slots to 0.