https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67635
ktkachov at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ktkachov at gcc dot gnu.org
--- Comment #2 from ktkachov at gcc dot gnu.org ---
I had a look at:
unsigned int test_03 (unsigned int x)
{
return x > 0 ? x + 1 : x;
}
The cse1 pass transforms this into something not quite comfortable for ifcvt.
Basically, it transforms the RTL equivalent of:
compare (x, 0);
if (x > 0)
temp := x + 1;
else
temp := x;
return temp;
into:
compare (x, 0);
if (x > 0)
temp := x + 1;
else
temp := 0;
return temp;
which is a valid transformation to make as far as cse is concerned.
Unfortunately, when ifcvt comes along, it's asked to optimise a conditional
select between '0' and 'x + 1', which is not something it can do with an
addc-style pattern.
I initially thought of looking into the condition to see if it contains x (so
that we can deduce that the 'else' arm could actually be replaced with x,
undoing the cse in that case), but the condition rtl contains a comparison with
the T-register, rather than the original comparison with 0.
Perhaps we can consider teaching cse to not transform these kinds of
expressions (c ? x : x + a) if the target has a store_flag/addcc instruction of
the appropriate mode?
But I'm not familiar with the cse code, so I don't know how easy/clean that is
