https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89829
Bug ID: 89829
Summary: incorrect profile data is used during
profiledbootstrap
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: other
Assignee: unassigned at gcc dot gnu.org
Reporter: belyshev at depni dot sinp.msu.ru
CC: marxin at gcc dot gnu.org
Target Milestone: ---
Looking at the *.gcda timestamps, I noticed that incorrect data (the one from
stageprofile libgcc build) is used, and correct one (from stagetrain) is
discarded.
The fix is just to revert r254150 which effectively undid 4-stage
profiledbootstrap to a previous status quo where only libgcc training data was
used.
Also note that the r254150 confusion probably stems from the fact that there
are two stagefeedback-start:: rules in the Makefile.tpl, where the first one is
generated from stage[+id+]-start:: template. The corresponding recipes are
executed both, and they are not independent as the first one renames directory
structure by doing 'make stage'. This (ab)use of double-colon rules is not
recommended by the GNU make documentation, see
https://www.gnu.org/software/make/manual/html_node/Double_002dColon.html#Double_002dColon
.
