https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77881
Bug ID: 77881
Summary: [5/6/7 Regression] Non-optimal signed comparison on
x86_64 since r146817
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
int
foo (long long int a, int b)
{
if (a >= 0 || b)
baz ();
}
int
bar (long long int a, int b)
{
if (a < 0 || b)
baz ();
}
on x86_64-linux with -O2 is compiled into following starting with r146817:
testq %rdi, %rdi
jns .L4
testl %esi, %esi
jne .L4
for foo, which looks reasonable, and
shrq $63, %rdi
testb %dil, %dil
jne .L9
testl %esi, %esi
jne .L9
in bar, where I'd expect and we used to emit in the past
testq %rdi, %rdi
js .L9
testl %esi, %esi
jne .L9
or something similar. I wonder if we want a combine pattern that would
transform right shift by precision - 1 followed by comparison of the result
against 0 if the result of the right shift is dead into just signed comparison.
