commit 1f8023e18a6dc7b950826810b392fdd46bcb0d45
Author: Mattias Andrée <[email protected]>
AuthorDate: Wed May 4 14:33:27 2016 +0200
Commit: Mattias Andrée <[email protected]>
CommitDate: Wed May 4 14:33:27 2016 +0200
Optimise zand, zor, and zxor
Signed-off-by: Mattias Andrée <[email protected]>
diff --git a/src/internals.h b/src/internals.h
index 841346f..f1a2f78 100644
--- a/src/internals.h
+++ b/src/internals.h
@@ -343,3 +343,31 @@ zfree_temp(z_t a)
}
/* } */
+
+#define ZMEM_OP(a, b, c, n, OP) \
+ do { \
+ zahl_char_t *a__ = (a); \
+ const zahl_char_t *b__ = (b); \
+ const zahl_char_t *c__ = (c); \
+ size_t i__, n__ = (n); \
+ if (n__ <= 4) { \
+ if (n__ >= 1) \
+ a__[0] = b__[0] OP c__[0]; \
+ if (n__ >= 2) \
+ a__[1] = b__[1] OP c__[1]; \
+ if (n__ >= 3) \
+ a__[2] = b__[2] OP c__[2]; \
+ if (n__ >= 4) \
+ a__[3] = b__[3] OP c__[3]; \
+ } else { \
+ for (i__ = 0; (i__ += 4) < n__;) { \
+ a__[i__ - 1] = b__[i__ - 1] OP c__[i__ - 1]; \
+ a__[i__ - 2] = b__[i__ - 2] OP c__[i__ - 2]; \
+ a__[i__ - 3] = b__[i__ - 3] OP c__[i__ - 3]; \
+ a__[i__ - 4] = b__[i__ - 4] OP c__[i__ - 4]; \
+ } \
+ if (i__ > n__) \
+ for (i__ -= 4; i__ < n__; i__++) \
+ a__[i__] = b__[i__] OP c__[i__]; \
+ } \
+ } while (0)
diff --git a/src/zand.c b/src/zand.c
index 6d3a184..45d1ed5 100644
--- a/src/zand.c
+++ b/src/zand.c
@@ -2,25 +2,6 @@
#include "internals.h"
-O2 static inline void
-zand_impl_3(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b, size_t n)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] &= b[i];
-}
-
-static inline void
-zand_impl_4(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b,
- register const zahl_char_t *restrict c, size_t n)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] = b[i] & c[i];
-}
-
void
zand(z_t a, z_t b, z_t c)
{
@@ -36,12 +17,12 @@ zand(z_t a, z_t b, z_t c)
a->used = MIN(b->used, c->used);
if (a == b) {
- zand_impl_3(a->chars, c->chars, a->used);
+ ZMEM_OP(a->chars, a->chars, c->chars, a->used, &);
} else if (unlikely(a == c)) {
- zand_impl_3(a->chars, b->chars, a->used);
+ ZMEM_OP(a->chars, a->chars, b->chars, a->used, &);
} else {
ENSURE_SIZE(a, a->used);
- zand_impl_4(a->chars, b->chars, c->chars, a->used);
+ ZMEM_OP(a->chars, b->chars, c->chars, a->used, &);
}
TRIM_AND_SIGN(a, zpositive1(b, c) * 2 - 1);
diff --git a/src/zor.c b/src/zor.c
index 2152c7f..ed952ca 100644
--- a/src/zor.c
+++ b/src/zor.c
@@ -2,27 +2,6 @@
#include "internals.h"
-O2 static inline void
-zor_impl_3(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b, size_t n)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] |= b[i];
-}
-
-static inline void
-zor_impl_5(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b, size_t n,
- register const zahl_char_t *restrict c, size_t m)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] = b[i] | c[i];
- for (; i < m; i++)
- a[i] = c[i];
-}
-
void
zor(z_t a, z_t b, z_t c)
{
@@ -40,17 +19,19 @@ zor(z_t a, z_t b, z_t c)
ENSURE_SIZE(a, m);
if (a == b) {
- zor_impl_3(a->chars, c->chars, n);
+ ZMEM_OP(a->chars, a->chars, c->chars, n, |);
if (a->used < c->used)
zmemcpy_range(a->chars, c->chars, n, m);
} else if (unlikely(a == c)) {
- zor_impl_3(a->chars, b->chars, n);
+ ZMEM_OP(a->chars, a->chars, b->chars, n, |);
if (a->used < b->used)
zmemcpy_range(a->chars, b->chars, n, m);
} else if (m == b->used) {
- zor_impl_5(a->chars, c->chars, n, b->chars, m);
+ ZMEM_OP(a->chars, c->chars, b->chars, n, |);
+ zmemcpy_range(a->chars, b->chars, n, m);
} else {
- zor_impl_5(a->chars, b->chars, n, c->chars, m);
+ ZMEM_OP(a->chars, b->chars, c->chars, n, |);
+ zmemcpy_range(a->chars, c->chars, n, m);
}
a->used = m;
diff --git a/src/zxor.c b/src/zxor.c
index 601a947..3f2ca5c 100644
--- a/src/zxor.c
+++ b/src/zxor.c
@@ -2,27 +2,6 @@
#include "internals.h"
-O2 static inline void
-zxor_impl_3(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b, size_t n)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] ^= b[i];
-}
-
-static inline void
-zxor_impl_5(register zahl_char_t *restrict a,
- register const zahl_char_t *restrict b, size_t n,
- register const zahl_char_t *restrict c, size_t m)
-{
- size_t i;
- for (i = 0; i < n; i++)
- a[i] = b[i] ^ c[i];
- for (; i < m; i++)
- a[i] = c[i];
-}
-
void
zxor(z_t a, z_t b, z_t c)
{
@@ -47,17 +26,19 @@ zxor(z_t a, z_t b, z_t c)
ENSURE_SIZE(a, m);
if (a == b) {
- zxor_impl_3(a->chars, cc, n);
+ ZMEM_OP(a->chars, a->chars, cc, n, ^);
if (a->used < cn)
zmemcpy_range(a->chars, cc, n, m);
} else if (unlikely(a == c)) {
- zxor_impl_3(a->chars, bc, n);
+ ZMEM_OP(a->chars, b->chars, cc, n, ^);
if (a->used < bn)
zmemcpy_range(a->chars, bc, n, m);
} else if (m == bn) {
- zxor_impl_5(a->chars, cc, n, bc, m);
+ ZMEM_OP(a->chars, c->chars, b->chars, n, ^);
+ zmemcpy_range(a->chars, b->chars, n, m);
} else {
- zxor_impl_5(a->chars, bc, n, cc, m);
+ ZMEM_OP(a->chars, b->chars, c->chars, n, ^);
+ zmemcpy_range(a->chars, c->chars, n, m);
}
a->used = m;
