https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70471
Bug ID: 70471
Summary: Superfluous move instructions in floating-point
instruction sequence
Product: gcc
Version: 5.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: schnetter at gmail dot com
Target Milestone: ---
In a function consisting of a long chain of floating-point operations, GCC
5.3.0 on Darwin 15.4.0 targeting an "Intel(R) Core(TM) i7-4980HQ CPU @ 2.80GHz"
with options
-march=native
-Ofast
-fopenmp
generates this sequence (line 6302 ff):
vmovapd %ymm4, %ymm8
vmovapd %ymm4, -10192(%rbp)
vsubpd %ymm1, %ymm8, %ymm0
vmovapd -4496(%rbp), %ymm8
Am I right with my interpretation that the first vmovapd is strictly
superfluous? The register %ymm8 is set on the first line, is overwritten on the
last line, and is used only once in the third line, where the original value
%ymm4 is still available.
Additional comments:
This is not the only place where inspection by eye indicates superfluous move
instructions; these seem to occur in many places. As you see below, both the
input and the generated code are quite long. Most relevant for performance is
likely the cache footprint. Thus, superfluous instructions are worrisome.
Details:
$ uname -a
Darwin Redshift 15.4.0 Darwin Kernel Version 15.4.0: Fri Feb 26 22:08:05 PST
2016; root:xnu-3248.40.184~3/RELEASE_X86_64 x86_64
$
/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/bin/g++
-v
Reading specs from
/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/lib64/gcc/x86_64-apple-darwin15.4.0/5.3.0/specs
COLLECT_GCC=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/bin/g++
COLLECT_LTO_WRAPPER=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/libexec/gcc/x86_64-apple-darwin15.4.0/5.3.0/lto-wrapper
Target: x86_64-apple-darwin15.4.0
Configured with:
/Users/eschnett/src/spack/var/spack/stage/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/gcc-5.3.0/configure
--prefix=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42
--libdir=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/lib64
--disable-multilib --enable-languages=fortran,c,java,objc,c++
--with-mpc=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/mpc-1.0.3-kg7pswhyszxa6vbgohqjhy2pywb76gpc
--with-mpfr=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/mpfr-3.1.4-taib2hirt72ggnirqb2brytc4cvp2igf
--with-gmp=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gmp-6.1.0-ld7rtqn2neg3z47mzg2vnexqeet4pz3i
--enable-lto --with-quad
--with-isl=/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/isl-0.14-cn4dbzoocjsf2a5jwamnhnverh2hwccr
Thread model: posix
gcc version 5.3.0 (GCC)
Compiled with:
$
/Users/eschnett/src/spack/opt/spack/darwin-x86_64/gcc-4.2.1/gcc-5.3.0-j3ujojmhirf6t2mi5enfosb6545duy42/bin/g++
-fopenmp -march=native -std=gnu++11 -Ofast -S
ML_BSSN_FD4_EvolutionInterior.cc.i
Pre-processed input "ML_BSSN_FD4_EvolutionInterior.cc.i" (3.5 MByte):
https://gist.github.com/eschnett/10bf0b2b1977348f3e15ae29db871bb0
Compiler output "ML_BSSN_FD4_EvolutionInterior.cc.s" (470 kByte):
https://gist.github.com/eschnett/79d31fe08e9588d28763a9ad5c77ccfa
