This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
streamline libgcc clz, ctz, parity
- From: Richard Henderson <rth at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Sat, 1 Feb 2003 13:05:03 -0800
- Subject: streamline libgcc clz, ctz, parity
Using the longlong.h implementations of count_leading_zeros
and count_trailing_zeros means that we can get arch-specific
code that you wouldn't necessarily want to inline, e.g. the
Alpha version pre-CIX extension, or the PA version.
Marking the tables ATTRIBUTE_HIDDEN means that the compiler
can often use a shorter PIC sequence to reference the data.
E.g. pc-relative instead of through the GOT.
Finally, on most targets getting the (pic) address of
__popcount_tab is more expensive than continuing to simplify
the parity with shifts.
r~
* libgcc2.c: Include auto-host.h.
(ATTRIBUTE_HIDDEN): New.
(__clz_tab): Don't declare here for clz and ctz.
(__clzsi2, __clzdi2): Use count_leading_zeros.
(__ctzsi2, __ctzdi2): Use count_trailing_zeros.
(__popcount_tab): Mark ATTRIBUTE_HIDDEN.
(__paritysi2, __paritydi2): Use shifts instead of __popcount_tab.
* longlong.h (__clz_tab): Mark ATTRIBUTE_HIDDEN.
Index: libgcc2.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/libgcc2.c,v
retrieving revision 1.158
diff -c -p -d -r1.158 libgcc2.c
*** libgcc2.c 1 Feb 2003 18:59:46 -0000 1.158
--- libgcc2.c 1 Feb 2003 20:58:12 -0000
*************** along with GCC; see the file COPYING. I
*** 29,38 ****
Software Foundation, 59 Temple Place - Suite 330, Boston, MA
02111-1307, USA. */
/* It is incorrect to include config.h here, because this file is being
compiled for the target, and hence definitions concerning only the host
do not apply. */
-
#include "tconfig.h"
#include "tsystem.h"
#include "coretypes.h"
--- 29,42 ----
Software Foundation, 59 Temple Place - Suite 330, Boston, MA
02111-1307, USA. */
+
+ /* We include auto-host.h here to get HAVE_GAS_HIDDEN. This is
+ supposedly valid even though this is a "target" file. */
+ #include "auto-host.h"
+
/* It is incorrect to include config.h here, because this file is being
compiled for the target, and hence definitions concerning only the host
do not apply. */
#include "tconfig.h"
#include "tsystem.h"
#include "coretypes.h"
*************** Software Foundation, 59 Temple Place - S
*** 43,48 ****
--- 47,58 ----
#undef abort
#endif
+ #ifdef HAVE_GAS_HIDDEN
+ #define ATTRIBUTE_HIDDEN __attribute__ ((__visibility__ ("hidden")))
+ #else
+ #define ATTRIBUTE_HIDDEN
+ #endif
+
#include "libgcc2.h"
#ifdef DECLARE_LIBRARY_RENAMES
*************** __udiv_w_sdiv (UWtype *rp __attribute__
*** 495,505 ****
#define L_udivmoddi4
#endif
- #if (defined (L_clzsi2) || defined (L_clzdi2) || \
- defined (L_ctzsi2) || defined (L_ctzdi2))
- extern const UQItype __clz_tab[];
- #endif
-
#ifdef L_clz
const UQItype __clz_tab[] =
{
--- 505,510 ----
*************** const UQItype __clz_tab[] =
*** 518,539 ****
Wtype
__clzsi2 (USItype x)
{
! Wtype a;
! /* Note that we've already verified that BITS_PER_UNIT == 8, and
! thus SItype is 32 bits wide. */
! if (x < (1 << 2 * 8))
! if (x < (1 << 1 * 8))
! a = 0 * 8;
! else
! a = 1 * 8;
! else
! if (x < (1 << 3 * 8))
! a = 2 * 8;
! else
! a = 3 * 8;
! return 32 - (__clz_tab[x >> a] + a);
}
#endif
--- 523,535 ----
Wtype
__clzsi2 (USItype x)
{
! UWtype w = x;
! Wtype ret;
! count_leading_zeros (ret, w);
! ret -= (sizeof(w) - sizeof(x)) * BITS_PER_UNIT;
! return ret;
}
#endif
*************** __clzsi2 (USItype x)
*** 541,555 ****
Wtype
__clzdi2 (UDItype x)
{
! Wtype a;
! /* Note that we've already verified that BITS_PER_UNIT == 8, and
! thus DItype is 64 bits wide. */
! for (a = 64 - 8; a > 0; a -= 8)
! if (((x >> a) & 0xff) != 0)
! break;
! return 64 - (__clz_tab[x >> a] + a);
}
#endif
--- 537,560 ----
Wtype
__clzdi2 (UDItype x)
{
! UWtype word;
! Wtype ret, add;
! if (sizeof(x) > sizeof(word))
! {
! DWunion uu;
! uu.ll = x;
! if (uu.s.high)
! word = uu.s.high, add = 0;
! else
! word = uu.s.low, add = W_TYPE_SIZE;
! }
! else
! word = x, add = (Wtype)(sizeof(x) - sizeof(word)) * BITS_PER_UNIT;
!
! count_leading_zeros (ret, word);
! return ret + add;
}
#endif
*************** __clzdi2 (UDItype x)
*** 557,580 ****
Wtype
__ctzsi2 (USItype x)
{
! Wtype a;
!
! x = x & -x;
! /* Note that we've already verified that BITS_PER_UNIT == 8, and
! thus SItype is 32 bits wide. */
! if (x < (1 << 2 * 8))
! if (x < (1 << 1 * 8))
! a = 0 * 8;
! else
! a = 1 * 8;
! else
! if (x < (1 << 3 * 8))
! a = 2 * 8;
! else
! a = 3 * 8;
! return __clz_tab[x >> a] + a - 1;
}
#endif
--- 562,572 ----
Wtype
__ctzsi2 (USItype x)
{
! Wtype ret;
! count_trailing_zeros (ret, x);
! return ret;
}
#endif
*************** __ctzsi2 (USItype x)
*** 582,601 ****
Wtype
__ctzdi2 (UDItype x)
{
! Wtype a;
! x = x & -x;
! for (a = 64 - 8; a > 0; a -= 8)
! if (((x >> a) & 0xff) != 0)
! break;
! return __clz_tab[x >> a] + a - 1;
}
#endif
! #if (defined (L_popcountsi2) || defined (L_popcountdi2) || \
! defined (L_paritysi2) || defined (L_paritydi2))
! extern const UQItype __popcount_tab[];
#endif
#ifdef L_popcount_tab
--- 574,603 ----
Wtype
__ctzdi2 (UDItype x)
{
! UWtype word;
! Wtype ret, add;
! if (sizeof(x) > sizeof(word))
! {
! DWunion uu;
! uu.ll = x;
! if (uu.s.low)
! word = uu.s.low, add = 0;
! else
! word = uu.s.high, add = W_TYPE_SIZE;
! }
! else
! word = x, add = 0;
!
! count_trailing_zeros (ret, word);
! return ret + add;
}
#endif
! #if (defined (L_popcountsi2) || defined (L_popcountdi2) \
! || defined (L_popcount_tab))
! extern const UQItype __popcount_tab[] ATTRIBUTE_HIDDEN;
#endif
#ifdef L_popcount_tab
*************** __popcountdi2 (UDItype x)
*** 642,650 ****
Wtype
__paritysi2 (USItype x)
{
! x ^= x >> 16;
! x ^= x >> 8;
! return __popcount_tab[x & 0xff] & 1;
}
#endif
--- 644,656 ----
Wtype
__paritysi2 (USItype x)
{
! UWtype nx = x;
! nx ^= nx >> 16;
! nx ^= nx >> 8;
! nx ^= nx >> 4;
! nx ^= nx >> 2;
! nx ^= nx >> 1;
! return nx & 1;
}
#endif
*************** __paritysi2 (USItype x)
*** 652,661 ****
Wtype
__paritydi2 (UDItype x)
{
! Wtype nx = x ^ (x >> 32);
nx ^= nx >> 16;
nx ^= nx >> 8;
! return __popcount_tab[nx & 0xff] & 1;
}
#endif
--- 658,670 ----
Wtype
__paritydi2 (UDItype x)
{
! UWtype nx = x ^ (x >> 32);
nx ^= nx >> 16;
nx ^= nx >> 8;
! nx ^= nx >> 4;
! nx ^= nx >> 2;
! nx ^= nx >> 1;
! return nx & 1;
}
#endif
Index: longlong.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/longlong.h,v
retrieving revision 1.33
diff -c -p -d -r1.33 longlong.h
*** longlong.h 3 Oct 2002 20:39:08 -0000 1.33
--- longlong.h 1 Feb 2003 20:58:12 -0000
*************** extern UDItype __udiv_qrnnd (UDItype *,
*** 134,140 ****
__asm__("cttz %1,%0" : "=r"(COUNT) : "r"(X))
#define COUNT_LEADING_ZEROS_0 64
#else
! extern const UQItype __clz_tab[];
#define count_leading_zeros(COUNT,X) \
do { \
UDItype __xr = (X), __t, __a; \
--- 134,140 ----
__asm__("cttz %1,%0" : "=r"(COUNT) : "r"(X))
#define COUNT_LEADING_ZEROS_0 64
#else
! extern const UQItype __clz_tab[] ATTRIBUTE_HIDDEN;
#define count_leading_zeros(COUNT,X) \
do { \
UDItype __xr = (X), __t, __a; \
*************** UDItype __umulsidi3 (USItype, USItype);
*** 1287,1293 ****
#endif
#if !defined (count_leading_zeros)
! extern const UQItype __clz_tab[];
#define count_leading_zeros(count, x) \
do { \
UWtype __xr = (x); \
--- 1287,1293 ----
#endif
#if !defined (count_leading_zeros)
! extern const UQItype __clz_tab[] ATTRIBUTE_HIDDEN;
#define count_leading_zeros(count, x) \
do { \
UWtype __xr = (x); \