/*#************************************************************************#*/ /*#-------------------------------------------------------------------------*/ /*# */ /*# FUNCTION NAME: memset() */ /*# */ /*# PARAMETERS: void* dst; Destination address. */ /*# int c; Value of byte to write. */ /*# int len; Number of bytes to write. */ /*# */ /*# RETURNS: dst. */ /*# */ /*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ /*# Framework taken from memcpy. This routine is */ /*# very sensitive to compiler changes in register allocation. */ /*# Should really be rewritten to avoid this problem. */ /*# */ /*#-------------------------------------------------------------------------*/ /*# */ /*# HISTORY */ /*# */ /*# DATE NAME CHANGES */ /*# ---- ---- ------- */ /*# 990713 HP Tired of watching this function (or */ /*# really, the nonoptimized generic */ /*# implementation) take up 90% of simulator */ /*# output. Measurements needed. */ /*# */ /*#-------------------------------------------------------------------------*/ #include /* No, there's no macro saying 12*4, since it is "hard" to get it into the asm in a good way. Thus better to expose the problem everywhere. */ /* Assuming 1 cycle per dword written or read (ok, not really true), and one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ #define ZERO_BLOCK_SIZE (1*12*4) void *memset(void *pdst, int c, size_t plen) { /* Ok. Now we want the parameters put in special registers. Make sure the compiler is able to make something useful of this. */ register char *return_dst __asm__ ("r10") = pdst; register int n __asm__ ("r12") = plen; register int lc __asm__ ("r11") = c; /* Most apps use memset sanely. Only those memsetting about 3..4 bytes or less get penalized compared to the generic implementation - and that's not really sane use. */ /* Ugh. This is fragile at best. Check with newer GCC releases, if they compile cascaded "x |= x << 8" sanely! */ __asm__("movu.b %0,$r13\n\t" "lslq 8,$r13\n\t" "move.b %0,$r13\n\t" "move.d $r13,%0\n\t" "lslq 16,$r13\n\t" "or.d $r13,%0" : "=r" (lc) : "0" (lc) : "r13"); { register char *dst __asm__ ("r13") = pdst; /* This is NONPORTABLE, but since this whole routine is */ /* grossly nonportable that doesn't matter. */ if (((unsigned long) pdst & 3) != 0 /* Oops! n=0 must be a legal call, regardless of alignment. */ && n >= 3) { if ((unsigned long)dst & 1) { *dst = (char) lc; n--; dst++; } if ((unsigned long)dst & 2) { *(short *)dst = lc; n -= 2; dst += 2; } } /* Now the fun part. For the threshold value of this, check the equation above. */ /* Decide which copying method to use. */ if (n >= ZERO_BLOCK_SIZE) { /* For large copies we use 'movem' */ /* It is not optimal to tell the compiler about clobbering any registers; that will move the saving/restoring of those registers to the function prologue/epilogue, and make non-movem sizes suboptimal. This method is not foolproof; it assumes that the "asm reg" declarations at the beginning of the function really are used here (beware: they may be moved to temporary registers). This way, we do not have to save/move the registers around into temporaries; we can safely use them straight away. If you want to check that the allocation was right; then check the equalities in the first comment. It should say "r13=r13, r12=r12, r11=r11" */ __asm__ volatile (" ;; Check that the following is true (same register names on ;; both sides of equal sign, as in r8=r8): ;; %0=r13, %1=r12, %4=r11 ;; ;; Save the registers we'll clobber in the movem process ;; on the stack. Don't mention them to gcc, it will only be ;; upset. subq 11*4,$sp movem $r10,[$sp] move.d $r11,$r0 move.d $r11,$r1 move.d $r11,$r2 move.d $r11,$r3 move.d $r11,$r4 move.d $r11,$r5 move.d $r11,$r6 move.d $r11,$r7 move.d $r11,$r8 move.d $r11,$r9 move.d $r11,$r10 ;; Now we've got this: ;; r13 - dst ;; r12 - n ;; Update n for the first loop subq 12*4,$r12 0: subq 12*4,$r12 bge 0b movem $r11,[$r13+] addq 12*4,$r12 ;; compensate for last loop underflowing n ;; Restore registers from stack movem [$sp+],$r10" /* Outputs */ : "=r" (dst), "=r" (n) /* Inputs */ : "0" (dst), "1" (n), "r" (lc)); } /* Either we directly starts copying, using dword copying in a loop, or we copy as much as possible with 'movem' and then the last block (<44 bytes) is copied here. This will work since 'movem' will have updated src,dst,n. */ while ( n >= 16 ) { *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; n -= 16; } /* A switch() is definitely the fastest although it takes a LOT of code. * Particularly if you inline code this. */ switch (n) { case 0: break; case 1: *(char*)dst = (char) lc; break; case 2: *(short*)dst = (short) lc; break; case 3: *((short*)dst)++ = (short) lc; *(char*)dst = (char) lc; break; case 4: *((long*)dst)++ = lc; break; case 5: *((long*)dst)++ = lc; *(char*)dst = (char) lc; break; case 6: *((long*)dst)++ = lc; *(short*)dst = (short) lc; break; case 7: *((long*)dst)++ = lc; *((short*)dst)++ = (short) lc; *(char*)dst = (char) lc; break; case 8: *((long*)dst)++ = lc; *((long*)dst)++ = lc; break; case 9: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *(char*)dst = (char) lc; break; case 10: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *(short*)dst = (short) lc; break; case 11: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((short*)dst)++ = (short) lc; *(char*)dst = (char) lc; break; case 12: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; break; case 13: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; *(char*)dst = (char) lc; break; case 14: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; *(short*)dst = (short) lc; break; case 15: *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((long*)dst)++ = lc; *((short*)dst)++ = (short) lc; *(char*)dst = (char) lc; break; } } return return_dst; /* destination pointer. */ } /* memset() */