http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50168
Bug #: 50168
Summary: __builtin_ctz() and intrinsics __bsr(), __bsf()
generate suboptimal code on x86_64
Classification: Unclassified
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: gp...@web.de
Testcase:
--------------------
#include <x86intrin.h>
static inline long my_bsfq(long x) __attribute__((__always_inline__));
static inline long my_bsfq(long x) {
long result;
asm(" bsfq %1, %0 \n"
: "=r"(result)
: "r"(x)
);
return result;
}
long c[64];
long f(long i) {
return c[ __bsfq(i) ];
}
long g(long i) {
return c[ __builtin_ctzll(i) ];
}
long h(long i) {
return c[ my_bsfq(i) ];
}
----------------------
When I compile this with 'gcc -O3 -g testcase.c -c -o testcase.o
&& objdump -d testcase', I get
----------------------
0000000000000000 <f>:
0: 48 0f bc ff bsf %rdi,%rdi
4: 48 63 ff movslq %edi,%rdi
7: 48 8b 04 fd 00 00 00 mov 0x0(,%rdi,8),%rax
e: 00
f: c3 retq
0000000000000010 <g>:
10: 48 0f bc ff bsf %rdi,%rdi
14: 48 63 ff movslq %edi,%rdi
17: 48 8b 04 fd 00 00 00 mov 0x0(,%rdi,8),%rax
1e: 00
1f: c3 retq
0000000000000020 <h>:
20: 48 0f bc ff bsf %rdi,%rdi
24: 48 8b 04 fd 00 00 00 mov 0x0(,%rdi,8),%rax
2b: 00
2c: c3 retq
-----------------------
Please note the unneeded 32 to 64 bit conversion 'movslq ...' inserted by the
compiler in functions f() and g(). It should look like h() instead.
I suspect the source is the prototype of the builtin, whose return type 'int'
does not match the "natural" return type on x86_64, which is 64 bit, the same
register size as the input register.
If I replace the builtin/intrinsic with the selfmade asm one, I get a nice
speedup of 2% in my chessengine.
