https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70731
Bug ID: 70731
Summary: With -Ofast, reorder floating-point and integer
operands to minimize conversions
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: josh at joshtriplett dot org
Target Milestone: ---
[Not sure if this goes in the 'c' component or some appropriate optimizer
component.]
Given options like -ffast-math (and specifically, probably -fassociative-math),
GCC should reorder operations with a series of floating-point and integer
operands to minimize the number of conversions.
For example, consider the following code (based loosely on a real optimization
issue observed in zopfli, see
http://roartindon.blogspot.com/2016/04/boosting-zopfli-performance.html):
double test1(int lbits, double lsym, int dbits, double dsym)
{
return lsym + lbits + dsym + dbits;
}
double test2(int lbits, double lsym, int dbits, double dsym)
{
return lbits + dbits + lsym + dsym;
}
gcc on x86-64, with either -O3 or -Ofast, produces two cvtsi2sd instructions
for the first, and just one cvtsi2sd for the second. (-Ofast does simplify
some of the surrounding instructions, but not the conversion.) Given
-fassociative-math, implied by -ffast-math and thus by -Ofast, I think GCC
could reorder the additions in test1 to only do one conversion.
