runtime: fix performance regression in morestack_noctxt on ppc64

In the fix for 54332 the MOVD R1, R1 instruction was added to
morestack_noctxt function to set the SPWRITE bit. However, the
instruction MOVD R1, R1 results in or r1,r1,r1 which is a special
instruction on ppc64 architecture as it changes the thread priority
and can negatively impact performance in some cases.
More details on such similar nops can be found in Power ISA v3.1
Book II on Power ISA Virtual Environment architecture in the chapter
on Program Priority Registers and Or instructions.
Replacing this by OR R0, R1 has the same affect on setting SPWRITE as
needed by the first fix but does not affect thread priority and
hence does not cause the degradation in performance

Hash65536-64           2.81GB/s ±10%  16.69GB/s ± 0%  +494.44%
Fixes #57741

Change-Id: Ib912e3716c6afd277994d6c1c5b2891f82225d50
Reviewed-on: https://go-review.googlesource.com/c/go/+/461597
Reviewed-by: Benny Siegert <bsiegert@gmail.com>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Auto-Submit: Benny Siegert <bsiegert@gmail.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
This commit is contained in:
Archana R 2023-01-11 11:45:28 -06:00 committed by Benny Siegert
parent ebb572d82f
commit 1c65b69bd1
1 changed files with 4 additions and 1 deletions

View File

@ -339,8 +339,11 @@ TEXT runtime·morestack_noctxt(SB),NOSPLIT|NOFRAME,$0-0
// the caller doesn't save LR on stack but passes it as a
// register (R5), and the unwinder currently doesn't understand.
// Make it SPWRITE to stop unwinding. (See issue 54332)
MOVD R1, R1
// Use OR R0, R1 instead of MOVD R1, R1 as the MOVD instruction
// has a special affect on Power8,9,10 by lowering the thread
// priority and causing a slowdown in execution time
OR R0, R1
MOVD R0, R11
BR runtime·morestack(SB)