Optimize n_revbin for Arm64

And remove byte_swap from longlong.h
flintlib · May 14, 2024 · 4602f9a · 4602f9a
1 parent e8ed578
commit 4602f9a
Show file tree

Hide file tree

Showing 7 changed files with 54 additions and 122 deletions.
diff --git a/doc/source/longlong.rst b/doc/source/longlong.rst
@@ -79,11 +79,3 @@ Division
 
     Works like ``udiv_qrnnd``, but takes a precomputed inverse ``di`` as 
     computed by ::func::`n_preinvert_limb`.
-
-Miscellaneous
--------------------------------------------------------------------------------
-
-.. macro:: byte_swap(x)
-
-    Swap the order of the bytes in the word `x`, i.e. most significant byte
-    becomes least significant byte, etc.
diff --git a/src/longlong.h b/src/longlong.h
@@ -34,11 +34,6 @@ extern "C" {
 #  define flint_ctz __builtin_ctzl
 # endif
 
-/* Byte swap */
-# define _FLINT_CAT_(X,Y) X##Y
-# define _FLINT_CAT(X,Y) _FLINT_CAT_(X,Y)
-# define byte_swap(x) do { (x) = _FLINT_CAT(__builtin_bswap, FLINT_BITS)(x); } while (0)
-
 /* Addition, subtraction and multiplication */
 # if defined(__clang__)
 #  include "longlong_asm_clang.h"
@@ -97,29 +92,6 @@ static inline int flint_ctz(ulong x)
 }
 #endif
 
-/* Byte swap */
-#if !defined(byte_swap)
-# if FLINT_BITS == 32
-#  define byte_swap(n) \
-  do { \
-      /* swap adjacent bytes */ \
-      (n) = ((((n) & 0xff00ff00) >> 8) | (((n) & 0x00ff00ff) << 8)); \
-      /* swap adjacent words */ \
-      (n) = (((n) >> 16) | ((n) << 16)); \
-  } while (0)
-# else
-#  define byte_swap(n) \
-  do { \
-      /* swap adjacent bytes */ \
-      (n) = ((((n) & 0xff00ff00ff00ff00) >> 8) | (((n) & 0x00ff00ff00ff00ff) << 8)); \
-      /* swap adjacent words */ \
-      (n) = ((((n) & 0xffff0000ffff0000) >> 16) | (((n) & 0x0000ffff0000ffff) << 16)); \
-      /* swap adjacent double words */ \
-      (n) = (((n) >> 32) | ((n) << 32)); \
-  } while (0)
-# endif
-#endif
-
 /* Addition and subtraction */
 #if !defined(add_ssaaaa)
 # define add_ssaaaa(s1, s0, a1, a0, b1, b0) \

diff --git a/src/longlong_msc_arm64.h b/src/longlong_msc_arm64.h
@@ -26,9 +26,6 @@ static inline int flint_ctz(ulong x)
     return index;
 }
 
-/* Byte swap */
-# define byte_swap(x) do { (x) = _byteswap_uint64(x); } while (0)
-
 /* Multiplication */
 #define umul_ppmm(r1, r0, u, v) \
 do \

diff --git a/src/longlong_msc_x86.h b/src/longlong_msc_x86.h
@@ -22,9 +22,6 @@
 # define flint_clz _lzcnt_u32
 # define flint_ctz _tzcnt_u32
 
-/* Byte swap */
-# define byte_swap(x) do { (x) = _byteswap_ulong(x); } while (0)
-
 /* Addition and subtraction */
 # define _FLINT_ADC _addcarry_u32
 # define _FLINT_SBB _subborrow_u32
@@ -56,9 +53,6 @@ do \
 # define flint_clz _lzcnt_u64
 # define flint_ctz _tzcnt_u64
 
-/* Byte swap */
-# define byte_swap(x) do { (x) = _byteswap_uint64(x); } while (0)
-
 /* Addition and subtraction */
 # define _FLINT_ADC _addcarry_u64
 # define _FLINT_SBB _subborrow_u64

diff --git a/src/test/main.c b/src/test/main.c
@@ -14,7 +14,6 @@
 #include "t-add_ssaaaa.c"
 #include "t-add_sssaaaaaa.c"
 #include "t-add_ssssaaaaaaaa.c"
-#include "t-byte_swap.c"
 #include "t-flint_clz.c"
 #include "t-flint_ctz.c"
 #include "t-io.c"
@@ -34,7 +33,6 @@ test_struct tests[] =
     TEST_FUNCTION(add_ssaaaa),
     TEST_FUNCTION(add_sssaaaaaa),
     TEST_FUNCTION(add_ssssaaaaaaaa),
-    TEST_FUNCTION(byte_swap),
     TEST_FUNCTION(flint_clz),
     TEST_FUNCTION(flint_ctz),
     TEST_FUNCTION(flint_fprintf),

diff --git a/src/test/t-byte_swap.c b/src/test/t-byte_swap.c
diff --git a/src/ulong_extras/revbin.c b/src/ulong_extras/revbin.c
@@ -1,5 +1,6 @@
 /*
     Copyright (C) 2009, 2015 William Hart
+    Copyright (C) 2024 Albin Ahlbäck
 
     This file is part of FLINT.
 
@@ -9,9 +10,23 @@
     (at your option) any later version.  See <https://www.gnu.org/licenses/>.
 */
 
-#include "flint.h"
 #include "ulong_extras.h"
 
+#if defined(__GNUC__) && FLINT64 && defined(__aarch64__)
+# include <arm_acle.h>
+ulong
+n_revbin(ulong n, ulong b)
+{
+    FLINT_ASSERT(b <= FLINT_BITS);
+
+    n = __rbitll(n);
+
+    if (b == 0)
+        return 0;
+    else
+        return n >> (FLINT_BITS - b);
+}
+#else
 static const unsigned char flint_revtab[] = {
     0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0,
     0x30, 0xB0, 0x70, 0xF0,
@@ -47,9 +62,42 @@ static const unsigned char flint_revtab[] = {
     0x3F, 0xBF, 0x7F, 0xFF
 };
 
-/*
-   Computes the reverse binary representation of a number of b bits.
-*/
+#if defined(__GNUC__)
+# if FLINT64
+#  define byte_swap __builtin_bswap64
+# else
+#  define byte_swap __builtin_bswap32
+# endif
+#elif defined(_MSC_VER)
+# include <stdlib.h>
+# if FLINT64
+#  define byte_swap _byteswap_uint64
+# else
+#  define byte_swap _byteswap_ulong
+# endif
+#else
+# if FLINT64
+FLINT_FORCE_INLINE byte_swap(ulong n)
+{
+    /* swap adjacent bytes */
+    n = ((n & 0xff00ff00ff00ff00) >> 8) | ((n & 0x00ff00ff00ff00ff) << 8);
+    /* swap adjacent words */
+    n = ((n & 0xffff0000ffff0000) >> 16) | ((n & 0x0000ffff0000ffff) << 16);
+    /* swap adjacent double words */
+    n = (n >> 32) | (n << 32);
+    return n;
+}
+# else
+FLINT_FORCE_INLINE byte_swap(ulong n)
+{
+    /* swap adjacent bytes */
+    n = ((n & 0xff00ff00) >> 8) | ((n & 0x00ff00ff) << 8);
+    /* swap adjacent words */
+    n = (n >> 16) | (n << 16);
+    return n;
+}
+# endif
+#endif
 
 ulong
 n_revbin(ulong n, ulong b)
@@ -83,8 +131,7 @@ n_revbin(ulong n, ulong b)
         n = (((n & 0xf0f0f0f0) >> 4) | ((n & 0x0f0f0f0f) << 4));
 #endif
 
-        byte_swap(n);
-
-        return n >> (FLINT_BITS - b);
+        return byte_swap(n) >> (FLINT_BITS - b);
     }
 }
+#endif