http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57650
Bug ID: 57650
Summary: Suboptimal code after TRUTH_AND_EXPR is changed into
BIT_AND_EXPR
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
On:
int baz (int);
int
foo (char *x, int y, int z)
{
if (y > z && x)
return baz (1);
if (x)
baz (2);
return 0;
}
int
bar (char *x, int y, int z)
{
if (x && y > z)
return baz (1);
if (x)
baz (2);
return 0;
}
one of the functions (depending on reassoc decission to keep or swap the
BIT_AND_EXPR arguments) ends up with really bad generated code, will do
testq %rdi, %rdi
setne %al
...
compare y with z, jump based on that, then test %al (separately in both
branches). Would be much better if tree optimizers could figure out that if
one of the BIT_AND_EXPR operands is also used as condition for a conditional
jump nearby, that it perhaps could jump thread this (essentially generate the
same code as
if (x)
{
if (y > z)
return bar (1);
bar (2);
}
would).
