http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46208
Summary: TARGET_PROMOTE_PROTOTYPES for SPARC V8
Product: gcc
Version: 4.5.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: jorge.pe...@invia.fr
Based on Ian Taylor's recommendation (http://goo.gl/yAoL)
******************* Problem summary (http://goo.gl/Mcrm)
When a caller function calls a callee function with short or char
arguments, the arguments are casted twice: inside the caller function
and inside the callee function, see the example. It is a waste of
performance in code density and speed!
Example: sparc-elf-gcc -o test.elf test.c -Os
short somme(char a, short b){
int i;
for (i=0; i<b; i++){
a+=a;
}
return a+b;
}
short somme2(char a, short b){
int i;
for (i=0; i<b; i++){
a+=b;
}
return a+2*b;
}
int main(){
volatile short b=1;
volatile char a=1, c=1;
b=somme(a,b);
b=somme(c,b);
b=somme(a,c);
b=somme2(a,b);
b=somme2(c,b);
b=somme2(a,c);
return 0;
}
0001024c <somme>:
1024c: 85 2a 60 10 sll %o1, 0x10, %g2
10250: 82 10 20 00 clr %g1
10254: 10 80 00 03 b 10260 <somme+0x14>
10258: 85 38 a0 10 sra %g2, 0x10, %g2
1025c: 82 00 60 01 inc %g1
10260: 80 a0 40 02 cmp %g1, %g2
10264: 26 bf ff fe bl,a 1025c <somme+0x10>
10268: 91 2a 20 01 sll %o0, 1, %o0
1026c: 91 2a 20 18 sll %o0, 0x18, %o0
10270: 91 3a 20 18 sra %o0, 0x18, %o0
10274: 81 c3 e0 08 retl
10278: 90 02 40 08 add %o1, %o0, %o0
000102b4 <main>:
102b4: 9d e3 bf 98 save %sp, -104, %sp
102b8: 82 10 20 01 mov 1, %g1
102bc: c2 37 bf fc sth %g1, [ %fp + -4 ]
102c0: c2 2f bf ff stb %g1, [ %fp + -1 ]
102c4: c2 2f bf fe stb %g1, [ %fp + -2 ]
102c8: d0 0f bf ff ldub [ %fp + -1 ], %o0
102cc: d2 17 bf fc lduh [ %fp + -4 ], %o1
102d0: 91 2a 20 18 sll %o0, 0x18, %o0
102d4: 93 2a 60 10 sll %o1, 0x10, %o1
102d8: 91 3a 20 18 sra %o0, 0x18, %o0
102dc: 7f ff ff dc call 1024c <somme>
102e0: 93 3a 60 10 sra %o1, 0x10, %o1
...
**************** Fix proposal (http://goo.gl/0mJ1)
we modified the function *TARGET_PROMOTE_PROTOTYPES
*(http://goo.gl/2pQQ) and set it to the default value which is FALSE.
static bool
sparc_promote_prototypes (const_tree fntype ATTRIBUTE_UNUSED)
{
return TARGET_ARCH32 ? true : false;
}
was replaced by
static bool
sparc_promote_prototypes (const_tree fntype ATTRIBUTE_UNUSED)
{
return false;
}
Then we re-compiled GCC 4.5.1 for SPARC V8 and the results for the
previously mentioned test.c are
short somme(char a, short b){
int i;
for (i=0; i<b; i++){
a+=a;
}
return a+b;
}
short somme2(char a, short b){
int i;
for (i=0; i<b; i++){
a+=b;
}
return a+2*b;
}
int main(){
volatile short b=1;
volatile char a=1, c=1;
b=somme(a,b);
b=somme(c,b);
b=somme(a,c);
b=somme2(a,b);
b=somme2(c,b);
b=somme2(a,c);
b=somme2(a,c*2);
b=somme2(a,c*3);
b=somme2(a,c*4);
return 0;
}
0001024c <somme>:
1024c: 85 2a 60 10 sll %o1, 0x10, %g2
10250: 82 10 20 00 clr %g1
10254: 10 80 00 03 b 10260 <somme+0x14>
10258: 85 38 a0 10 sra %g2, 0x10, %g2
1025c: 82 00 60 01 inc %g1
10260: 80 a0 40 02 cmp %g1, %g2
10264: 26 bf ff fe bl,a 1025c <somme+0x10>
10268: 91 2a 20 01 sll %o0, 1, %o0
1026c: 91 2a 20 18 sll %o0, 0x18, %o0
10270: 91 3a 20 18 sra %o0, 0x18, %o0
10274: 81 c3 e0 08 retl
10278: 90 02 00 09 add %o0, %o1, %o0
0001027c <somme2>:
1027c: 87 2a 60 10 sll %o1, 0x10, %g3
10280: 82 10 20 00 clr %g1
10284: 87 38 e0 10 sra %g3, 0x10, %g3
10288: 10 80 00 03 b 10294 <somme2+0x18>
1028c: 84 10 00 09 mov %o1, %g2
10290: 82 00 60 01 inc %g1
10294: 80 a0 40 03 cmp %g1, %g3
10298: 26 bf ff fe bl,a 10290 <somme2+0x14>
1029c: 90 02 00 02 add %o0, %g2, %o0
102a0: 93 2a 60 01 sll %o1, 1, %o1
102a4: 91 2a 20 18 sll %o0, 0x18, %o0
102a8: 91 3a 20 18 sra %o0, 0x18, %o0
102ac: 81 c3 e0 08 retl
102b0: 90 02 40 08 add %o1, %o0, %o0
000102b4 <main>:
102b4: 9d e3 bf 98 save %sp, -104, %sp
102b8: 82 10 20 01 mov 1, %g1
102bc: c2 37 bf fc sth %g1, [ %fp + -4 ]
102c0: c2 2f bf ff stb %g1, [ %fp + -1 ]
102c4: c2 2f bf fe stb %g1, [ %fp + -2 ]
102c8: d0 0f bf ff ldub [ %fp + -1 ], %o0
102cc: d2 17 bf fc lduh [ %fp + -4 ], %o1
102d0: 7f ff ff df call 1024c <somme>
102d4: 01 00 00 00 nop
102d8: d0 37 bf fc sth %o0, [ %fp + -4 ]
102dc: d0 0f bf fe ldub [ %fp + -2 ], %o0
102e0: d2 17 bf fc lduh [ %fp + -4 ], %o1
102e4: 7f ff ff da call 1024c <somme>
102e8: 01 00 00 00 nop
102ec: d0 37 bf fc sth %o0, [ %fp + -4 ]
102f0: d0 0f bf ff ldub [ %fp + -1 ], %o0
102f4: d2 0f bf fe ldub [ %fp + -2 ], %o1
102f8: 93 2a 60 18 sll %o1, 0x18, %o1
102fc: 7f ff ff d4 call 1024c <somme>
10300: 93 3a 60 18 sra %o1, 0x18, %o1
10304: d0 37 bf fc sth %o0, [ %fp + -4 ]
10308: d0 0f bf ff ldub [ %fp + -1 ], %o0
1030c: d2 17 bf fc lduh [ %fp + -4 ], %o1
10310: 7f ff ff db call 1027c <somme2>
10314: 01 00 00 00 nop
10318: d0 37 bf fc sth %o0, [ %fp + -4 ]
1031c: d0 0f bf fe ldub [ %fp + -2 ], %o0
10320: d2 17 bf fc lduh [ %fp + -4 ], %o1
10324: 7f ff ff d6 call 1027c <somme2>
10328: 01 00 00 00 nop
1032c: d0 37 bf fc sth %o0, [ %fp + -4 ]
10330: d0 0f bf ff ldub [ %fp + -1 ], %o0
10334: d2 0f bf fe ldub [ %fp + -2 ], %o1
10338: 93 2a 60 18 sll %o1, 0x18, %o1
1033c: 7f ff ff d0 call 1027c <somme2>
10340: 93 3a 60 18 sra %o1, 0x18, %o1
10344: d0 37 bf fc sth %o0, [ %fp + -4 ]
10348: d0 0f bf ff ldub [ %fp + -1 ], %o0
1034c: d2 0f bf fe ldub [ %fp + -2 ], %o1
10350: 93 2a 60 18 sll %o1, 0x18, %o1
10354: 7f ff ff ca call 1027c <somme2>
10358: 93 3a 60 17 sra %o1, 0x17, %o1
1035c: d0 37 bf fc sth %o0, [ %fp + -4 ]
10360: d0 0f bf ff ldub [ %fp + -1 ], %o0
10364: c2 0f bf fe ldub [ %fp + -2 ], %g1
10368: 83 28 60 18 sll %g1, 0x18, %g1
1036c: 83 38 60 18 sra %g1, 0x18, %g1
10370: 93 28 60 01 sll %g1, 1, %o1
10374: 7f ff ff c2 call 1027c <somme2>
10378: 92 02 40 01 add %o1, %g1, %o1
1037c: b0 10 20 00 clr %i0
10380: d0 37 bf fc sth %o0, [ %fp + -4 ]
10384: d0 0f bf ff ldub [ %fp + -1 ], %o0
10388: d2 0f bf fe ldub [ %fp + -2 ], %o1
1038c: 93 2a 60 18 sll %o1, 0x18, %o1
10390: 7f ff ff bb call 1027c <somme2>
10394: 93 3a 60 16 sra %o1, 0x16, %o1
10398: d0 37 bf fc sth %o0, [ %fp + -4 ]
1039c: 81 c7 e0 08 ret
103a0: 81 e8 00 00 restore
this reduced the size of the main by 96 bytes which is huge taking into
account that this is a dummy example.
Thanks for your help,
Best regards,
Jorge Perez
