-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 05/30/11 04:55, Bernd Schmidt wrote: > On 05/27/2011 07:49 PM, Jeff Law wrote: >> Updated based on some comments from Bernd; specifically the other use of >> delete_dead_insn has been removed. >> >> WRT the assembly differences on MIPS Bernd referred to; what ultimately >> caused this problem were two dead insns that had been previously >> eliminated by reload were still in the insn stream and inhibited an >> if-conversion which resulted in slightly different assembly code, but >> shouldn't have had any significant impact on the performance or size of >> the resulting code. The dead insns were deleted by the post-reload DCE >> pass (which obviously runs after post-reload if conversion). >> >> If we really wanted to get those insns out of the stream, we could flag >> when reload deleted insns which might result in dead code remaining in >> the stream, then conditionally run DCE immediately after reload. I >> didn't think this was worth doing right now, but if someone objects I >> can certainly look into it. > > Patch is OK (still no codegen changes on x86 as far as I can tell), but > I'd appreciate the DCE thing as a followup. I suspect it's going to be just as easy to go ahead and add the DCE cleanup now rather than waiting. It should be a relatively trivial tweak to delete_dead_insn to set a flag which we'd check at the end of reload. It's #2 in the queue :-)
jeff -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.11 (GNU/Linux) Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/ iQEcBAEBAgAGBQJN5SmlAAoJEBRtltQi2kC7RfwH/RV9OsQIwMeKOHUOgOkpvQZx DIjRjlGTld+ndDN7sjgt0bIVLBITdR2K1CcVyPMDCZ2qZB0yiuDbXDQOZJkdrq89 55BUKrK7h5KLI5Ad8z5cGsyeK5nY3TUXmuPsdl0Pk/Lbu5m+EfxxxuQw0EljBhY9 UOasXcQ8VAXut0k4banTNFrY07TCM3gHMN2qRKTGiabtq/gh17ez/DejTVt3+2Ru 1Qi9XUVpL1mLwWsrqhI4JIe/zhpJS4eJizvP4Dywp+0pdplSPJVV9tybFWiojdfP zt/4IQD0YHosTdxIEPyLWZBuszyG0TDmsQYyt29cXYzQeZSge+T1CF9TfEcX1wY= =tSMU -----END PGP SIGNATURE-----
