https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87254
Bug ID: 87254
Summary: Inlining clones in trivial wrappers
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: ipa
Assignee: unassigned at gcc dot gnu.org
Reporter: glisse at gcc dot gnu.org
CC: marxin at gcc dot gnu.org
Target Milestone: ---
int h();
static int f(int j,double*a){
if(j)
{
double t[1024];
for(int i=0;i<1024;++i){t[i]=i;}
return t[(int)*a];
}
else
return h();
}
int g(double*a){return f(0,a);}
Compiled with -O3.
The code in f is random crap, the key elements to reproduce are:
- f is not inlined in einline,
- f is cloned in IPA-CP,
- the clone of f is not inlined in g.
(so a simpler reproducer may be possible with -fdisable-tree-einline --param
ipa-cp-eval-threshold=10)
The first 2 steps are normal, but the last one seems wrong to me. There is
nothing to lose by inlining f's clone into the trivial wrapper that is its only
caller. The inline dump complains that "--param large-stack-frame-growth limit
reached", but changing this param has no effect, I have to change
"large-stack-frame" instead to convince it to inline.
Just in case, before that I was investigating
static int f(int j,double*a){
if(j)
{
double t[1024];
for(int i=0;i<1024;++i){t[i]=i;}
return t[(int)*a];
}
else
{
float t[1024];
for(int i=0;i<1024;++i){t[i]=i;}
return t[(int)*a];
}
}
int g(double*a){return f(0,a);}
int h(double*a){return f(1,a);}
with -O3 -fdisable-tree-einline --param ipa-cp-eval-threshold=10 .
In the past, I have seen such trivial left-over wrappers with -flto on GMP (
https://gmplib.org/list-archives/gmp-devel/2018-April/004874.html ), and I was
reminded of it by https://stackoverflow.com/q/51152215/1918193 .
