http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55488
Bug #: 55488
Summary: Implement cold calls in tsan run-time
Classification: Unclassified
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: sanitizer
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: konstantin.s.serebry...@gmail.com
CC: do...@gcc.gnu.org, dvyu...@gcc.gnu.org,
ja...@gcc.gnu.org, k...@gcc.gnu.org, w...@gcc.gnu.org
The hottest functions in tsan run-time make two cold calls: __tsan_report_race
and __tsan_trace_switch
If these calls are implemented via regular calling convention,
they ruin the performance since the compiler creates too many spills.
So, we've manually implemented a cold-call calling convention using a separate
assembly file, libsanitizer/tsan/tsan_rtl_amd64.S
Currently, this hack is disabled in GCC causing the tsan run-time to be slower
than it could be:
libsanitizer/tsan/tsan_rtl.h:
#if 0 && TSAN_DEBUG == 0
...
#define HACKY_CALL(f) \
We need to enable building/linking the file libsanitizer/tsan/tsan_rtl_amd64.S
and enable the HACKY_CALL in libsanitizer/tsan/tsan_rtl.h. This will eliminate
the only difference left between the gcc version and upstream.
Or maybe gcc has another way to implement a cold call?
