Take the following testcase (either on spu-elf or powerpc-linux-gnu with
-maltivec):
#define vector __attribute__((__vector_size__(16) ))
typedef vector float vec_float4;
typedef struct {
vec_float4 data;
} VecFloat4;
typedef struct {
vec_float4 a;
vec_float4 b;
} VecFloat4x2;
VecFloat4 test1(VecFloat4 a, VecFloat4 b)
{
a.data = a.data+b.data;
return a;
}
VecFloat4x2 test2(VecFloat4x2 data)
{
data.a = data.a+data.a;
data.b = data.b+data.b;
return data;
}
----- cut -----
Right now we do (for spu-elf, it is a similar issue for PPC):
_Z5test211VecFloat4x2:
hbr .L5,$lr
stqd $sp,-128($sp)
ai $sp,$sp,-128
stqd $3,64($sp)
stqd $4,80($sp)
lqd $5,80($sp)
lqd $4,64($sp)
fa $2,$5,$5
fa $3,$4,$4
stqd $2,48($sp)
stqd $3,32($sp)
lqd $4,48($sp)
lqd $3,32($sp)
ai $sp,$sp,128
.L5:
bi $lr
---- cut ----
With the patch which I will attach, we get:
_Z5test211VecFloat4x2:
fa $2,$3,$3
hbr .L5,$lr
stqd $sp,-128($sp)
ai $sp,$sp,-128
nop 127
stqd $3,64($sp)
fa $3,$4,$4
stqd $4,80($sp)
nop 127
stqd $2,32($sp)
ori $4,$3,0
stqd $3,48($sp)
ori $3,$2,0
ai $sp,$sp,128
.L5:
bi $lr
----------- cut ------
Notice how the loads are gone.
Note dse could do the same.
--
Summary: postreload can handle the case where the memory
locations use different modes
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: rtl-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: pinskia at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33790
