https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91116
Bug ID: 91116
Summary: bad register choices for rs6000 -m32
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: segher at gcc dot gnu.org
Target Milestone: ---
In the new testcase pr88233.c, which is
typedef struct { double a[2]; } A;
A
foo (const A *a)
{
return *a;
}
we currently get as generated code for -m32
addi 10,4,4
lfiwzx 10,0,4
addi 9,3,12
lfiwzx 11,0,10
addi 10,4,8
lfiwzx 12,0,10
addi 10,4,12
stfiwx 10,0,3
lfiwzx 0,0,10
addi 10,3,4
stfiwx 11,0,10
addi 10,3,8
stfiwx 12,0,10
stfiwx 0,0,9
blr
Expand decides to do this as four SImode copies, which isn't such a great
idea, of course; but RA thinks it is cost 0 to put a SImode in an FP or
altivec register. That won't fly.
