diff options
| author | Ulrich Drepper <drepper@redhat.com> | 2001-05-27 06:45:14 +0000 |
|---|---|---|
| committer | Ulrich Drepper <drepper@redhat.com> | 2001-05-27 06:45:14 +0000 |
| commit | 995a692a486b2e250f598097543dd04135cec327 (patch) | |
| tree | e50617b52eb6477edee1186983cb687e60125c38 /sysdeps/ia64/memmove.S | |
| parent | 17ffa4986db1ce263a8a3c4c322dc237cfaa2777 (diff) | |
Update.
2001-05-22 David Mosberger <davidm@hpl.hp.com>
* sysdeps/ia64/memmove.S: Increase MEMLAT from 6 to 21 for better
performance.
* sysdeps/ia64/memcpy.S: Likewise.
* sysdeps/ia64/bcopy.S: New file.
* sysdeps/ia64/bzero.S: New file (derived from memset.S).
2001-05-26 Ulrich Drepper <drepper@redhat.com>
* sysdeps/ia64/fpu/libm-test-ulps: Add deltas for tanh(-0.7).
Diffstat (limited to 'sysdeps/ia64/memmove.S')
| -rw-r--r-- | sysdeps/ia64/memmove.S | 22 |
1 files changed, 12 insertions, 10 deletions
diff --git a/sysdeps/ia64/memmove.S b/sysdeps/ia64/memmove.S index a3f7edef92..7b9fe2265d 100644 --- a/sysdeps/ia64/memmove.S +++ b/sysdeps/ia64/memmove.S @@ -26,7 +26,7 @@ in2: byte count The core of the function is the memcpy implementation used in memcpy.S. - When bytes have to be copied backwards, only the easy case, when + When bytes have to be copied backwards, only the easy case, when all arguments are multiples of 8, is optimised. In this form, it assumes little endian mode. For big endian mode, @@ -67,10 +67,12 @@ br.ctop.sptk .loop##shift ; \ br.cond.sptk .cpyfew ; /* deal with the remaining bytes */ +#define MEMLAT 21 +#define Nrot (((2*MEMLAT+3) + 7) & ~7) + ENTRY(memmove) .prologue - alloc r2 = ar.pfs, 3, 29, 0, 32 -#include "softpipe.h" + alloc r2 = ar.pfs, 3, Nrot - 3, 0, Nrot .rotr r[MEMLAT + 2], q[MEMLAT + 1] .rotp p[MEMLAT + 2] mov ret0 = in0 // return value = dest @@ -90,7 +92,7 @@ ENTRY(memmove) and tmp4 = 7, tmp3 // tmp4 = (dest | src | len) & 7 cmp.le p6, p0 = dest, src // if dest <= src it's always safe (p6) br.cond.spnt .forward // to copy forward - add tmp3 = src, len;; + add tmp3 = src, len;; cmp.lt p6, p0 = dest, tmp3 // if dest > src && dest < src + len (p6) br.cond.spnt .backward // we have to copy backward @@ -113,7 +115,7 @@ ENTRY(memmove) (p6) br.cond.spnt .restore_and_exit;;// the one-word special case adds adest = 8, dest // set adest one word ahead of dest adds asrc = 8, src ;; // set asrc one word ahead of src - nop.b 0 // get the "golden" alignment for + nop.b 0 // get the "golden" alignment for nop.b 0 // the next loop .l0: (p[0]) ld8 r[0] = [src], 16 @@ -139,8 +141,8 @@ ENTRY(memmove) .l1: // copy -dest % 8 bytes ld1 value = [src], 1 // value = *src++ ;; - st1 [dest] = value, 1 // *dest++ = value - br.cloop.dptk .l1 + st1 [dest] = value, 1 // *dest++ = value + br.cloop.dptk .l1 .dest_aligned: and sh1 = 7, src // sh1 = src % 8 and tmp2 = -8, len // tmp2 = len & -OPSIZ @@ -148,7 +150,7 @@ ENTRY(memmove) shr.u loopcnt = len, 3 // loopcnt = len / 8 and len = 7, len;; // len = len % 8 adds loopcnt = -1, loopcnt // --loopcnt - addl tmp4 = @ltoff(.table), gp + addl tmp4 = @ltoff(.table), gp addl tmp3 = @ltoff(.loop56), gp mov ar.ec = MEMLAT + 1 // set EC mov pr.rot = 1 << 16;; // set rotating predicates @@ -174,7 +176,7 @@ ENTRY(memmove) LOOP(40) LOOP(48) LOOP(56) - + .src_aligned: .l3: (p[0]) ld8 r[0] = [src], 8 @@ -220,7 +222,7 @@ ENTRY(memmove) adds src = -1, src // src points to the last byte adds dest = -1, dest // dest points to the last byte adds loopcnt = -1, len;; // loopcnt = len - 1 - mov ar.lc = loopcnt;; // set the loop counter + mov ar.lc = loopcnt;; // set the loop counter .l6: (p[0]) ld1 r[0] = [src], -1 (p[MEMLAT]) st1 [dest] = r[MEMLAT], -1 |