The following block of code appears to produce an un-needed memcpy on both
Intel and PowerPC platforms. There is no aliasing or side-effect that I can
think of that could possibly force such copying to occur -- the problem seems
to be that gcc is not aware of the lifetime of large structures kept on stack.
The full source code:
class TV
{
private:
float truth;
float confidence;
int stuff[444];
public:
TV(void);
float getT(void);
};
extern TV my_tv_maker(float tr);
extern void other(TV *);
float my_subr(float tr)
{
TV tv;
other (&tv); // force constructor TV::TV to run first
tv = my_tv_maker(434.23); // over-write previous tv.
return tv.getT();
}
PowerPC assembly, created with gcc -S -O2 -c
.L._Z7my_subrf:
.LFB2:
mflr 0
.LCFI0:
std 28,-32(1)
.LCFI1:
std 29,-24(1)
.LCFI2:
std 0,16(1)
.LCFI3:
stdu 1,-3728(1) make room for two instances of TV on stack
.LCFI4:
addi 29,1,112 one instance of TV
addi 28,1,1904 second instance of TV
mr 3,29
bl _ZN2TVC1Ev call constructor on instance 1
nop
mr 3,29
bl _Z5otherP2TV call other() on instance 1
nop
lfs 1,....@toc(2)
mr 3,28
bl _Z11my_tv_makerf call my_tv_make on instance 2
nop
mr 4,28
mr 3,29
li 5,1784
bl memcpy copy instance 2 over to 1! waste of CPU!
nop
mr 3,29
bl _ZN2TV4getTEv call method on instance 1
nop
addi 1,1,3728
ld 0,16(1)
ld 28,-32(1)
ld 29,-24(1)
mtlr 0
blr
The missed optimizations are: -- two copies of the instance are not needed; the
copy is not needed either. For large structures, this can be a significant
time-waster.
Exactly the same problem shows up in Intel as well:
_Z7my_subrf:
.LFB2:
pushl %ebp
.LCFI0:
movl %esp, %ebp
.LCFI1:
subl $3608, %esp
.LCFI2:
movl %ebx, -8(%ebp)
.LCFI3:
leal -1792(%ebp), %ebx instance 1 of TV
movl %esi, -4(%ebp)
.LCFI4:
leal -3592(%ebp), %esi instance 2 of TV
movl %ebx, (%esp)
call _ZN2TVC1Ev call constructor on instance 1
movl %ebx, (%esp)
call _Z5otherP2TV call other() on instance 1
movl %esi, (%esp)
movl $0x43d91d71, 4(%esp)
call _Z11my_tv_makerf call my_tv_maker on instance 2
subl $4, %esp
movl %esi, 4(%esp)
movl %ebx, (%esp)
movl $1784, 8(%esp)
call memcpy copy instance 2 to instance 1
movl %ebx, (%esp)
call _ZN2TV4getTEv call getT() on instance 1
movl -8(%ebp), %ebx
movl -4(%ebp), %esi
movl %ebp, %esp
popl %ebp
ret
--
Summary: missed optimization: un-needed copy of structure.
Product: gcc
Version: 4.1.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: linasvepstas at gmail dot com
GCC build triplet: powerpc64-unknown-linux-gnu
GCC host triplet: powerpc64-unknown-linux-gnu
GCC target triplet: powerpc64-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39081
