[LLVMbugs] [Bug 22024] Unnecessary vxorps instructions
bugzilla-daemon at llvm.org
bugzilla-daemon at llvm.org
Mon Jan 5 10:36:35 PST 2015
http://llvm.org/bugs/show_bug.cgi?id=22024
Quentin Colombet <qcolombet at apple.com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
CC| |qcolombet at apple.com
Resolution|--- |WONTFIX
--- Comment #4 from Quentin Colombet <qcolombet at apple.com> ---
As far as I know, xorps A,A,A are good for performance because it tells that A
is zeroed. In this particular case, it even helps to break a scheduling
dependency. Indeed, the out of order engine would see two independent
computation:
vcvtsi2ssq %rcx, %xmm0, %xmm0
vmovss %xmm0, (%rsi)
and
vcvtsi2ssq %rdx, %xmm0, %xmm0
vmovss %xmm0, (%rdi)
Instead of one big computation:
vcvtsi2ssq %rcx, %xmm0, %xmm0
vmovss %xmm0, (%rsi)
vcvtsi2ssq %rdx, %xmm0, %xmm0 # <-- we do not know that the xmm0
given as input is independent of the previous convert.
vmovss %xmm0, (%rdi)
If you actually see a performance problem, feel free to reopen.
Thanks.
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20150105/d3d2bc61/attachment.html>
More information about the llvm-bugs
mailing list