Hi Eric,
Your ffs, ffsl, ffsll modules make use of __builtin_ffs for GCC >= 3.4.
But the same compiler versions also have __builtin_ffsl and __builtin_ffsll.
Their use simplifies (and most certainly speeds up) the code generated
by GCC. For example, on MacOS X in 64-bit mode, currently the code is
.globl _ffsll
_ffsll:
LFB2:
pushq %rbp
LCFI0:
xorl %eax, %eax
testq %rdi, %rdi
movq %rsp, %rbp
LCFI1:
je L6
xorl %edx, %edx
testl %edi, %edi
movl %edi, %eax
jne L5
.align 4,0x90
L8:
shrq $32, %rdi
addl $32, %edx
testl %edi, %edi
movl %edi, %eax
je L8
L5:
bsfl %eax, %eax
movl $-1, %ecx
cmove %ecx, %eax
incl %eax
addl %edx, %eax
L6:
leave
ret
LFE2:
whereas with the attached patch it becomes:
.globl _ffsll
_ffsll:
LFB2:
pushq %rbp
LCFI0:
movq $-1, %rax
bsfq %rdi, %rdi
movq %rsp, %rbp
LCFI1:
leave
cmove %rax, %rdi
incq %rdi
movl %edi, %eax
ret
LFE2:
Of course, I've verified that the test suite still passes.
2011-10-13 Bruno Haible <br...@clisp.org>
ffsl, ffsll: Optimize for GCC.
* lib/ffsl.h (FUNC): Use GCC_BUILTIN if defined.
* lib/ffsl.c (GCC_BUILTIN): New macro.
* lib/ffsll.c (GCC_BUILTIN): Likewise.
--- lib/ffsl.h.orig Fri Oct 14 00:47:36 2011
+++ lib/ffsl.h Fri Oct 14 00:40:34 2011
@@ -31,6 +31,9 @@
int
FUNC (TYPE i)
{
+#if (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4)) && defined
GCC_BUILTIN
+ return GCC_BUILTIN (i);
+#else
int result = 0;
unsigned TYPE j = i;
@@ -44,4 +47,5 @@
j >>= CHAR_BIT * sizeof (unsigned int);
result += CHAR_BIT * sizeof (unsigned int);
}
+#endif
}
--- lib/ffsl.c.orig Fri Oct 14 00:47:36 2011
+++ lib/ffsl.c Fri Oct 14 00:40:35 2011
@@ -1,3 +1,4 @@
#define FUNC ffsl
#define TYPE long int
+#define GCC_BUILTIN __builtin_ffsl
#include "ffsl.h"
--- lib/ffsll.c.orig Fri Oct 14 00:47:36 2011
+++ lib/ffsll.c Fri Oct 14 00:40:33 2011
@@ -1,3 +1,4 @@
#define FUNC ffsll
#define TYPE long long int
+#define GCC_BUILTIN __builtin_ffsll
#include "ffsl.h"
--
In memoriam Bekir Çoban-zade <http://en.wikipedia.org/wiki/Bekir_Çoban-zade>
