http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47430
Summary: [4.6 Regression] Random PCH related bootstrap failures
on powerpc64-linux
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: pch
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ja...@gcc.gnu.org
CC: b...@gcc.gnu.org, pa...@gcc.gnu.org
I saw during the weekend two bootstrap failures in libstdc++.
The problem seems to be that powerpc64 uses very large pages (64KB), address
space randomization and the fact that for building stdtr1c++.h.gch and
extc++.h.gch uses stdc++.h.gch.
When building e.g. stdtr1c++.h.gch, stdc++.h.gch already likely exists and thus
it first mmaps stdc++.h.gch at the desired address, which is likely
TRY_EMPTY_VM_SPACE (OT, for __powerpc__ && __LP64__, shouldn't it use some
larger value, say 0x6000000000ULL ?), thus when attempting to determine best VM
slot for stdtr1c++.h.gch, that slot is already taken by stdc++.h.gch and thus
can't be used. While there is some code to attempt to handle address space
randomization, perhaps 32MB works well for 4KB pages, but is too small for
larger pages or something, but it certainly fails from time to time on ppc64
and some shared library is mmapped at that point.
Potential ways how to fix this, one or more of:
1) in host-linux.c define more than one TRY_EMPTY_VM_SPACE value for each OS,
and if mmap at TRY_EMPTY_VM_SPACE fails, try TRY_EMPTY_VM_SPACE2 which is say
0x70000000 for most 32-bit hosts and some GB appart from TRY_EMPTY_VM_SPACE for
64-bit hosts.
2) make sure other PCH is not read during compilation that is writing PCH
3) change the logic for address space randomization in host-linux.c, use larger
buffer if page size is > 4KB
4) change libstdc++ makefiles to make sure stdc++.h.gch does not exist when
building the other *.gch files, use explicit ordering in the Makefile that it
is first rm -f'ed, then the other gch files built, then stdc++.h.gch built
Preferences?
