https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121038
Bug ID: 121038
Summary: autoprofiledbootstrap is broken in few ways
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: bootstrap
Assignee: unassigned at gcc dot gnu.org
Reporter: hubicka at gcc dot gnu.org
Target Milestone: ---
Lets track the problems here. Currently
1) autoprofiledbootstrap fails for me at 256 core machine since perf runs out
of memory
Workaround is:
diff --git a/Makefile.tpl b/Makefile.tpl
index 21cd95e1bc8..cbf80544cc1 100644
--- a/Makefile.tpl
+++ b/Makefile.tpl
@@ -416,7 +416,7 @@ MAKEINFO = @MAKEINFO@
EXPECT = @EXPECT@
RUNTEST = @RUNTEST@
-AUTO_PROFILE = gcc-auto-profile --all -c 10000000 -o perf.data
+AUTO_PROFILE = gcc-auto-profile --all -c 10000000 -o perf.data -m10
# This just becomes part of the MAKEINFO definition passed down to
# sub-makes. It lets flags be given on the command line while still
2) if --with-build-config=bootstrap-lto is used then the middle-end is not
trained at all, since we do not produce auto-profile of lto1. Fix is posted at
https://gcc.gnu.org/pipermail/gcc-patches/2025-July/689329.html
3) autorpfoiledbootstrap seems to build stage1 with host compiler and within
perf it builds stage2 and uses produced profile for stage3. This means that
profile is mixing profile of unoptimized binary buildt by host compiler and
profile of stage2 compiler that is used to build runtime afterwards. This is
wrong; we should always train on compiler build by stage1 compiler and not
stage1 compiler itself
Workaround is to always build previously built compiler and use -O2 for stage1
flags.
4) there is no debug compare for autorpofiledbootstrap and profiledbootstrap
5) I think --all -c 10000000 to gcc-auto-profile is kind of wrong. I do not
see why we need to profile kernel (--all) and also I think count should be
prime number. It seems it is relatively low count for size of the train run.
