https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87595
Bug ID: 87595
Summary: __tls_get_addr should be __attribute__((__noplt__))
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: bugdal at aerifal dot cx
Target Milestone: ---
The relevant code seems to be in targets, but this enhancement request applies
to all targets.
__tls_get_addr is sufficiently a bottleneck that many projects (even gcc target
libs) try to bypass it by using initial-exec model. In general, bypassing the
PLT and calling directly through the GOT will save at least an icache line and
1 instruction. On some targets it takes several instruction to get through the
PLT, and also imposes constraints on register allocation (e.g. ebx on i386).
My initial testing shows -fno-plt makes GD TLS access about 8% faster on i386
and no worse on x86_64. I will try to post some reproducible benchmarks as a
follow-up later.
