commit 84ad8821d456e8f7f40df43b7eb7245703004ce7
parent d6f4393542998276250bd3f3519bb824ca4b3d91
Author: Mattias Andrée <maandree@kth.se>
Date: Sat, 7 May 2016 18:15:59 +0200
Optimise libzahl_memcpy for clang
Signed-off-by: Mattias Andrée <maandree@kth.se>
Diffstat:
| M | INSTALL | | | 4 | +++- |
| M | STATUS | | | 40 | ++++++++++++++++++++-------------------- |
| M | zahl/memory.h | | | 45 | ++++++++++++++++++++++++++++++++++++++------- |
3 files changed, 61 insertions(+), 28 deletions(-)
diff --git a/INSTALL b/INSTALL
@@ -33,4 +33,6 @@ libzahl contains some (very little) assembly code. In the event
that the used instructions are not supported on your machine, please
report it, and in the meanwhile add -DZAHL_NO_ASM to CPPFLAGS. You
make also have to do this if you are compiling with a compiler that
-does not support extended inline assembly.
+does not support extended inline assembly. You may also have to add
+ #define ZAHL_NO_ASM
+to your program before includeing <zahl.h>
diff --git a/STATUS b/STATUS
@@ -18,7 +18,7 @@ processes are fixed to one CPU.
The following functions are probably implemented optimally:
-zset .................... always fastest (gcc); until ~1200 (clang [can be fixed with assembly])
+zset .................... always fastest
zseti(a, +) ............. tomsfastmath is faster
zseti(a, -) ............. tomsfastmath is faster
zsetu ................... tomsfastmath is faster
@@ -30,8 +30,20 @@ zodd .................... always fastest (shared with gmp)
zeven_nonzero ........... always fastest (shared with gmp)
zodd_nonzero ............ always fastest (shared with gmp)
zbtest .................. always fastest
-zsave ................... always fastest [clang needs zset fix]
-zload ................... always fastest [clang needs zset fix]
+zsave ................... always fastest
+zload ................... always fastest
+
+
+ The following functions are probably implemented optimally, but
+ depends on other functions or call-cases for better performance:
+
+zneg(a, b) .............. always fastest
+zabs(a, b) .............. always fastest
+ztrunc(a, b, c) ......... always fastest
+zbset(a, b, 1) .......... always fastest
+zbset(a, b, 0) .......... always fastest
+zbset(a, b, -1) ......... always fastest
+zsplit .................. alternating with gmp for fastest, but gmp is a bit faster on average
The following functions are probably implemented close to
@@ -40,26 +52,14 @@ zload ................... always fastest [clang needs zset fix]
zadd_unsigned ........... fastest after ~140 (depends on cc and libc) compared against zadd too
ztrunc(a, a, b) ......... fastest until ~100, then 77 % (gcc) or 68 % (clang) of tomsfastmath
zbset(a, a, 1) .......... always fastest
-zbset(a, a, 0) .......... always fastest (faster with clang than gcc)
+zbset(a, a, 0) .......... always fastest
zbset(a, a, -1) ......... always fastest (only marginally faster than gmp with clang)
zlsb .................... always fastest <<suspicious>>
zlsh .................... not too fast anymore
-zand .................... fastest after ~400, tomsfastmath before (gcc+glibc is slow)
-zor ..................... fastest after ~1150, tomsfastmath before (gcc+glibc is slow)
-zxor .................... fastest after ~400, tomsfastmath before (clang), gcc is slow
-znot .................... always fastest (faster with musl than glibc)
-
-
- The following functions are probably implemented optimally, but
- depends on other functions or call-cases for better performance:
-
-zneg(a, b) .............. always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-zabs(a, b) .............. always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-ztrunc(a, b, c) ......... always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-zbset(a, b, 1) .......... always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-zbset(a, b, 0) .......... always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-zbset(a, b, -1) ......... always fastest (gcc+musl); gcc is a bit slow [clang needs zset fix]
-zsplit .................. alternating with gmp for fastest (clang and glibc is slower)
+zand .................... fastest after ~400, tomsfastmath before
+zor ..................... fastest after ~1150, tomsfastmath before
+zxor .................... alternative with gmp after ~700, tomsfastmath before (musl), a bit slow with glibc
+znot .................... always fastest
The following functions require structural changes for
diff --git a/zahl/memory.h b/zahl/memory.h
@@ -34,16 +34,47 @@
ZAHL_INLINE void
-libzahl_memcpy(register zahl_char_t *restrict d, register const zahl_char_t *restrict s, size_t n)
+libzahl_memcpy(register zahl_char_t *restrict d, register const zahl_char_t *restrict s, register size_t n)
{
- size_t i;
#define LIBZAHL_X(I) case I: d[I - 1] = s[I - 1];
LIBZAHL_SMALL_INPUT_BEGIN(n);
- for (i = 0; i < n; i += 4) {
- d[i + 0] = s[i + 0];
- d[i + 1] = s[i + 1];
- d[i + 2] = s[i + 2];
- d[i + 3] = s[i + 3];
+ {
+#if defined(__x86_64__) && !defined(ZAHL_NO_ASM)
+ /* This crap is needed for clang. */
+ register zahl_char_t t;
+ __asm__ __volatile__ (
+# if defined(ZAHL_ISA_MISSING_INDIRECT_JUMP)
+ "\n testq %[e], %[e]"
+ "\n jz 2f"
+# endif
+ "\n shlq $3, %[e]"
+ "\n addq %[d], %[e]"
+ "\n 1:"
+ "\n movq 0(%[s]), %[t]"
+ "\n movq %[t], 0(%[d])"
+ "\n movq 8(%[s]), %[t]"
+ "\n movq %[t], 8(%[d])"
+ "\n movq 16(%[s]), %[t]"
+ "\n movq %[t], 16(%[d])"
+ "\n movq 24(%[s]), %[t]"
+ "\n movq %[t], 24(%[d])"
+ "\n addq $32, %[s]"
+ "\n addq $32, %[d]"
+ "\n cmpq %[e], %[d]"
+ "\n jl 1b"
+# if defined(ZAHL_ISA_MISSING_INDIRECT_JUMP)
+ "\n 2:"
+# endif
+ : [t]"=r"(t), [d]"+r"(d), [s]"+r"(s), [e]"+r"(n));
+#else
+ size_t i;
+ for (i = 0; i < n; i += 4) {
+ d[i + 0] = s[i + 0];
+ d[i + 1] = s[i + 1];
+ d[i + 2] = s[i + 2];
+ d[i + 3] = s[i + 3];
+ }
+#endif
}
LIBZAHL_SMALL_INPUT_END;
#undef LIBZAHL_X
