https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82981
Bug ID: 82981
Summary: unnecessary __multi3 call for mips64r6 linux kernel
Product: gcc
Version: 7.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wilson at gcc dot gnu.org
Target Milestone: ---
Given the testcase
unsigned long func(unsigned long a, unsigned long b)
{
return a > (~0UL) / b;
}
compiled with -march=mips64r6 -mabi=64 -mexplicit-relocs -O2 we end up with a
call to __multi3 which is inefficient and inconvenient.
The testcase gets converted to
_1 = MUL_OVERFLOW (a_4(D), b_5(D));
_2 = IMAGPART_EXPR <_1>;
There are no mulv patterns in the mips port, so it tries mulditi3 which fails,
and then calls __multi3.
Mips64r6 does have a widening DImode multiply which should have been used
instead. The problem is that the *mulditi3 pattern is missing mips64r6
support.
Alternatively, the expander should try using a muldi3_highpart pattern when the
mulditi3 pattern doesn't work. Especially when the highpart is the only part
we need as in this example. The mips64r6 multi3_highpart is present.
Or alternatively, a mulvti3 pattern should be added to the mips port.
See for instance the thread at
https://www.linux-mips.org/archives/linux-mips/2017-08/msg00041.html
