-Oz is not respected by LSR (-enable-iv-rewrite revealed fluctuations in code size)

[Quuxplusone]

Bugzilla Link	PR12342
Status	NEW
Importance	P enhancement
Reported by	Jörg Sonnenberger (joerg@NetBSD.org)
Reported on	2012-03-24 00:32:01 -0700
Last modified on	2012-03-27 17:30:27 -0700
Version	trunk
Hardware	PC Linux
CC	anton@korobeynikov.info, atrick@apple.com, efriedma@quicinc.com, llvm-bugs@lists.llvm.org
Fixed by commit(s)
Attachments	enable-iv-rewrite.tgz (298267 bytes, application/x-compressed-tar)
Blocks
Blocked by
See also

Created attachment 8262
Files that compile to different sizes and list

Attached is the list of all files from NetBSD's libkern, libsa and libi386 that
differ in code size between -enable-iv-rewrite=false and -enable-iv-
rewrite=true. Some of them are not used in this context. The false.txt /
true.txt list the object sizes as I see it during the release build. compile.sh
*/*.i will compile the files with both options. Big regressions are e.g.
qsort.i, improvements are seen for subr_prf.i

[Quuxplusone]

Attached enable-iv-rewrite.tgz (298267 bytes, application/x-compressed-tar): Files that compile to different sizes and list

[Quuxplusone]

I'm having trouble knowing what to look at here. When I run compile.sh and
simply sum the .o sizes, I get a *larger* size for enable-iv-rewrite=true. I'm
using svn r153254.

false: 54760 bytes
 true: 55508 bytes

* .o sizes
  2596 cd9660.i.false
  2676 cd9660.i.true
   324 dkcksum.i.false
   324 dkcksum.i.true
  5032 dosfs.i.false
  5052 dosfs.i.true
  1948 exec.i.false
  1968 exec.i.true
  4268 ext2fs.i.false
  4264 ext2fs.i.true
  3512 ffsv1.i.false
  3504 ffsv1.i.true
  4156 ffsv2.i.false
  4148 ffsv2.i.true
   592 gets.i.false
   624 gets.i.true
   480 ip_cksum.i.false
   484 ip_cksum.i.true
  3560 lfsv1.i.false
  3552 lfsv1.i.true
  3592 lfsv2.i.false
  3584 lfsv2.i.true
  5200 loadfile_elf32.i.false
  4684 loadfile_elf32.i.true
  4888 loadfile_elf64.i.false
  5488 loadfile_elf64.i.true
  3596 minixfs3.i.false
  3588 minixfs3.i.true
  2248 netif.i.false
  2248 netif.i.true
  2128 qsort.i.false
  2780 qsort.i.true
  1336 strerror.i.false
  1348 strerror.i.true
  2024 subr_prf.i.false
  1920 subr_prf.i.true
  3280 ufs.i.false
  3272 ufs.i.true

[Quuxplusone]

Hm. I thought qsort.i had a large regression, but I might have messed up the list at that point. subr_prf.i is what triggered the investigation: it ended up with 120 Bytes overhead.

[Quuxplusone]

There is a lot of luck involved in getting the "right" codegen for a particular loop nest. Performance results are just as jittery as the size results shown here. The good news is:

• the new default looks like a significant improvement in overall size

• with the new default (-enable-iv-rewrite=false) someone writing their own boot loader could conceivably hand tune loops without llvm undoing the work. If llvm still gets in the way, try -mllvm -disable-lsr (as an unsupported workaround).

I can leave this open because there is an opportunity to make the LSR pass aware of the -Oz option, and this is potentially a good example to motivate it. That said, it's very low on my list of enhancements to add.

[Quuxplusone]

Good enough. Working around the related PR 12348 side steps the issue completely for me, so keeping the example for analysis is good enough.