http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59478
Bug ID: 59478 Summary: Optimize variable access via byte copy Product: gcc Version: 4.9.0 Status: UNCONFIRMED Severity: enhancement Priority: P3 Component: tree-optimization Assignee: unassigned at gcc dot gnu.org Reporter: olegendo at gcc dot gnu.org Target: sh*-*-* This happens at least on SH with trunk rev 205905 (4.9). I'm not sure whether these are target specific or not. Accessing float values as integers can be done in various ways. One way is to do a byte copy... int float_as_int (float val) { char valbytes[sizeof (float)]; __builtin_memcpy (valbytes, &val, sizeof (float)); int result; __builtin_memcpy (&result, valbytes, sizeof (float)); return result; } The above compiled with -m4-single -ml -O2 results in: add #-8,r15 fmov.s fr5,@r15 mov.l @r15,r0 rts add #8,r15 which is not so bad actually, but could be done better by utilizing the fpul register, as it is done when using the union approach: int float_as_int (float val) { union { int i; float f; } tmp; tmp.f = val; return tmp.i; }; compiled with -m4-single -ml -O2: flds fr5,fpul rts sts fpul,r0 It seems that the above could be fixed with a combine pattern, as combine is looking for: Failed to match this instruction: (parallel [ (set (mem/c:SF (plus:SI (reg/f:SI 153 sfp) (const_int -8 [0xfffffffffffffff8])) [3 S4 A32]) (reg:SF 69 fr5 [ val ])) (use (reg/v:PSI 151 )) (set (reg/f:SI 166) (plus:SI (reg/f:SI 153 sfp) (const_int -8 [0xfffffffffffffff8]))) ]) However, this might have some side effects if the location in the stack frame is actually never written. So probably this should be handled earlier during compilation. When writing the mem copy manually, there seems to be another problem: int float_as_int (float val) { char valbytes[sizeof (float)]; for (int i = 0; i < sizeof (float); ++i) valbytes[i] = ((char*)&val)[i]; int result; for (int i = 0; i < sizeof (float); ++i) ((char*)&result)[i] = valbytes[i]; return result; } compiled with -m4-single -ml -O2: add #-8,r15 fmov.s fr5,@r15 // store float at (sfp+0) mov.b @(1,r15),r0 // load 4 bytes from (sfp+0) mov r0,r7 // (loop is unrolled at -O2) mov.b @(2,r15),r0 mov r0,r3 mov.b @(3,r15),r0 mov r0,r2 mov.b @r15,r0 mov.b r0,@(4,r15) // store 4 bytes from (sfp+0) at (sfp+4) mov r7,r0 // (loop is unrolled at -O2) mov.b r0,@(5,r15) mov r3,r0 mov.b r0,@(6,r15) mov r2,r0 mov.b r0,@(7,r15) mov.l @(4,r15),r0 // load int from (sfp+4) rts add #8,r15 compiled with -m4-single -ml -O3: add #-8,r15 // this is the same as using __builtin_memcpy fmov.s fr5,@r15 mov.l @(0,r15),r0 rts add #8,r15 However, when just doing a simple byte wise integer read, the loop is not unrolled at -O2: int read_int (const char* val) { int result; for (int i = 0; i < sizeof (int); ++i) ((char*)&result)[i] = val[i]; return result; } compiled with -m4-single -ml -O2: add #-4,r15 mov #0,r0 mov #0,r2 mov #4,r1 .L3: mov.b @(r0,r4),r3 add #1,r2 dt r1 mov.b r3,@(r0,r15) bf/s .L3 mov r2,r0 mov.l @(0,r15),r0 rts add #4,r15 compiled with -m4-single -ml -O3: add #-4,r15 mov.b @r4,r0 mov.b r0,@r15 mov.b @(1,r4),r0 mov.b r0,@(1,r15) mov.b @(2,r4),r0 mov.b r0,@(2,r15) mov.b @(3,r4),r0 mov.b r0,@(3,r15) mov.l @(0,r15),r0 rts add #4,r15