http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53189
Bug #: 53189
Summary: DImode and/or/not/xor optimized poorly in
core-registers
Classification: Unclassified
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
AssignedTo: [email protected]
ReportedBy: [email protected]
Target: arm
The following code does not optimize well on current trunk.
--------------
void bar (long long, long long);
void
foo (long long a)
{
bar (a&1, a);
}
--------------
Compiled with "-O2 -mfpu=vfpv3 -mthumb" gives:
--------------
foo:
mov r2, r0
mov r3, r1
movs r0, #1
movs r1, #0
ands r0, r0, r2
ands r1, r1, r3
b bar
--------------
As you can see there are many missed optimizations here:
1. Failure to notice that r1 will always be zero.
2. Failure to use immediate constant "#1" with "ands".
I'd expect output like this:
mov r2, r0
mov r3, r1
ands r0, r0, #1
mov r1, #0
b bar
The problem is two-fold:
First, adddi3 does not expand to two instructions so the two parts of the
operation cannot be optimized independently.
Second, adddi3 does not allow immediate constants so the expander is forced to
put the constants in registers.
As a general rule, if NEON or IWMMXT is not in use then DImode operations
should be decomposed from expand. If NEON/IWMMXT is available then
decomposition should be delayed until after reload, and the splitters should
attempt to produce optimal sequences in as many cases as possible. (Ideally, we
would be able to make the decision long before register allocation, but we're
not there yet.)
