> In that case your patch modification seems to work better since it seems > to remove the cast from the callee and does the sign extension on the > caller: > > (gcc4.5.1 with your patch applied) > > 0001024c <somme>: > 1024c: 10 80 00 04 b 1025c <somme+0x10> > 10250: 82 10 20 00 clr %g1 > 10254: 82 00 60 01 inc %g1 > 10258: 91 3a 20 18 sra %o0, 0x18, %o0 > 1025c: 80 a0 40 09 cmp %g1, %o1 > 10260: 26 bf ff fd bl,a 10254 <somme+0x8> > 10264: 91 2a 20 19 sll %o0, 0x19, %o0 > 10268: 90 02 00 09 add %o0, %o1, %o0 > 1026c: 91 2a 20 10 sll %o0, 0x10, %o0 > 10270: 81 c3 e0 08 retl > 10274: 91 3a 20 10 sra %o0, 0x10, %o0 > > 000102ac <main>: > 102ac: 9d e3 bf 98 save %sp, -104, %sp > 102b0: 82 10 20 01 mov 1, %g1 > 102b4: c2 37 bf fc sth %g1, [ %fp + -4 ] > 102b8: c2 2f bf ff stb %g1, [ %fp + -1 ] > 102bc: c2 2f bf fe stb %g1, [ %fp + -2 ] > 102c0: d0 0f bf ff ldub [ %fp + -1 ], %o0 > 102c4: d2 17 bf fc lduh [ %fp + -4 ], %o1 > 102c8: 91 2a 20 18 sll %o0, 0x18, %o0 > 102cc: 93 2a 60 10 sll %o1, 0x10, %o1 > 102d0: 91 3a 20 18 sra %o0, 0x18, %o0 > 102d4: 7f ff ff de call 1024c <somme> > 102d8: 93 3a 60 10 sra %o1, 0x10, %o1 > 102dc: d0 37 bf fc sth %o0, [ %fp + -4 ] > 102e0: d0 0f bf fe ldub [ %fp + -2 ], %o0 > 102e4: d2 17 bf fc lduh [ %fp + -4 ], %o1 > 102e8: 91 2a 20 18 sll %o0, 0x18, %o0 > 102ec: 93 2a 60 10 sll %o1, 0x10, %o1 > 102f0: 91 3a 20 18 sra %o0, 0x18, %o0 > 102f4: 7f ff ff d6 call 1024c <somme> > ... > > isn't that right? Yes, the code is now correct, but still sub-optimal as the sequence: > 102c4: d2 17 bf fc lduh [ %fp + -4 ], %o1 > 102cc: 93 2a 60 10 sll %o1, 0x10, %o1 > 102d8: 93 3a 60 10 sra %o1, 0x10, %o1 could be reduced to just: ldsh [ %fp + -4 ], %o1 but this requires more work in the back-end. This is actually the bulk of the work to be done as the extension in the callee that TARGET_PROMOTE_PROTOTYPES triggers has very likely a marginal effect overall. Note that x86, MIPS, HP-PA, m68k, IA-64 and many others have the same setting. -- Eric Botcazou
