tile: optimize memcmp

Customize memcmp.c for tile, using similar tricks from memcpy:

- replace MERGE macro with dblalign.
- replace memcmp_bytes function with revbytes.
- use __glibc_likely.
- use post-increment addressing.

The schedule is still not perfect: the compiler is not hoisting
code above the comparison branch, which could save a bundle or two.
memcmp speeds up by 30-40% on shorter aligned tests in benchtest,
with some tests with unaligned lengths taking a small performance hit.

1 files changed, 367 insertions, 0 deletions

diff --git a/sysdeps/tile/memcmp.c b/sysdeps/tile/memcmp.c
new file mode 100644
index 0000000..8d39921
--- /dev/null
+++ b/sysdeps/tile/memcmp.c
@@ -0,0 +1,367 @@
+/* Copyright (C) 1991-2014 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Torbjorn Granlund (tege@sics.se).
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#undef	__ptr_t
+#define __ptr_t	void *
+
+#if defined HAVE_STRING_H || defined _LIBC
+# include <string.h>
+#endif
+
+#undef memcmp
+
+#ifndef MEMCMP
+# define MEMCMP memcmp
+#endif
+
+#ifdef _LIBC
+
+# include <memcopy.h>
+# include <endian.h>
+
+# if __BYTE_ORDER == __BIG_ENDIAN
+#  define WORDS_BIGENDIAN
+# endif
+
+#else	/* Not in the GNU C library.  */
+
+# include <sys/types.h>
+
+/* Type to use for aligned memory operations.
+   This should normally be the biggest type supported by a single load
+   and store.  Must be an unsigned type.  */
+# define op_t	unsigned long int
+# define OPSIZ	(sizeof(op_t))
+
+/* Threshold value for when to enter the unrolled loops.  */
+# define OP_T_THRES	16
+
+/* Type to use for unaligned operations.  */
+typedef unsigned char byte;
+
+#endif	/* In the GNU C library.  */
+
+/* Provide the appropriate builtins to shift two registers based on
+   the alignment of a pointer held in a third register, and to reverse
+   the bytes in a word.  */
+#ifdef __tilegx__
+#define DBLALIGN __insn_dblalign
+#define REVBYTES __insn_revbytes
+#else
+#define DBLALIGN __insn_dword_align
+#define REVBYTES __insn_bytex
+#endif
+
+#ifdef WORDS_BIGENDIAN
+# define CMP_LT_OR_GT(a, b) ((a) > (b) ? 1 : -1)
+#else
+# define CMP_LT_OR_GT(a, b) (REVBYTES(a) > REVBYTES(b) ? 1 : -1)
+#endif
+
+/* BE VERY CAREFUL IF YOU CHANGE THIS CODE!  */
+
+/* The strategy of this memcmp is:
+
+   1. Compare bytes until one of the block pointers is aligned.
+
+   2. Compare using memcmp_common_alignment or
+      memcmp_not_common_alignment, regarding the alignment of the other
+      block after the initial byte operations.  The maximum number of
+      full words (of type op_t) are compared in this way.
+
+   3. Compare the few remaining bytes.  */
+
+static int memcmp_common_alignment (long, long, size_t) __THROW;
+
+/* memcmp_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN `op_t'
+   objects (not LEN bytes!).  Both SRCP1 and SRCP2 should be aligned for
+   memory operations on `op_t's.  */
+static int
+memcmp_common_alignment (srcp1, srcp2, len)
+     long int srcp1;
+     long int srcp2;
+     size_t len;
+{
+  op_t a0, a1;
+  op_t b0, b1;
+
+  switch (len % 4)
+    {
+    default: /* Avoid warning about uninitialized local variables.  */
+    case 2:
+      a0 = ((op_t *) srcp1)[0];
+      b0 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      len += 2;
+      goto do1;
+    case 3:
+      a1 = ((op_t *) srcp1)[0];
+      b1 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      len += 1;
+      goto do2;
+    case 0:
+      if (OP_T_THRES <= 3 * OPSIZ && len == 0)
+	return 0;
+      a0 = ((op_t *) srcp1)[0];
+      b0 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      goto do3;
+    case 1:
+      a1 = ((op_t *) srcp1)[0];
+      b1 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      len -= 1;
+      if (OP_T_THRES <= 3 * OPSIZ && len == 0)
+	goto do0;
+      /* Fall through.  */
+    }
+
+  do
+    {
+      a0 = ((op_t *) srcp1)[0];
+      b0 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (a1 != b1))
+	return CMP_LT_OR_GT (a1, b1);
+
+    do3:
+      a1 = ((op_t *) srcp1)[0];
+      b1 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (a0 != b0))
+	return CMP_LT_OR_GT (a0, b0);
+
+    do2:
+      a0 = ((op_t *) srcp1)[0];
+      b0 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (a1 != b1))
+	return CMP_LT_OR_GT (a1, b1);
+
+    do1:
+      a1 = ((op_t *) srcp1)[0];
+      b1 = ((op_t *) srcp2)[0];
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (a0 != b0))
+	return CMP_LT_OR_GT (a0, b0);
+
+      len -= 4;
+    }
+  while (len != 0);
+
+  /* This is the right position for do0.  Please don't move
+     it into the loop.  */
+ do0:
+  if (__glibc_likely (a1 != b1))
+    return CMP_LT_OR_GT (a1, b1);
+  return 0;
+}
+
+static int memcmp_not_common_alignment (long, long, size_t) __THROW;
+
+/* memcmp_not_common_alignment -- Compare blocks at SRCP1 and SRCP2 with LEN
+   `op_t' objects (not LEN bytes!).  SRCP2 should be aligned for memory
+   operations on `op_t', but SRCP1 *should be unaligned*.  */
+static int
+memcmp_not_common_alignment (srcp1, srcp2, len)
+     long int srcp1;
+     long int srcp2;
+     size_t len;
+{
+  void * srcp1i;
+  op_t a0, a1, a2, a3;
+  op_t b0, b1, b2, b3;
+  op_t x;
+
+  /* Calculate how to shift a word read at the memory operation
+     aligned srcp1 to make it aligned for comparison.  */
+
+  srcp1i = (void *) srcp1;
+
+  /* Make SRCP1 aligned by rounding it down to the beginning of the `op_t'
+     it points in the middle of.  */
+  srcp1 &= -OPSIZ;
+
+  switch (len % 4)
+    {
+    default: /* Avoid warning about uninitialized local variables.  */
+    case 2:
+      a1 = ((op_t *) srcp1)[0];
+      a2 = ((op_t *) srcp1)[1];
+      b2 = ((op_t *) srcp2)[0];
+      srcp1 += 2 * OPSIZ;
+      srcp2 += 1 * OPSIZ;
+      len += 2;
+      goto do1;
+    case 3:
+      a0 = ((op_t *) srcp1)[0];
+      a1 = ((op_t *) srcp1)[1];
+      b1 = ((op_t *) srcp2)[0];
+      srcp1 += 2 * OPSIZ;
+      srcp2 += 1 * OPSIZ;
+      len += 1;
+      goto do2;
+    case 0:
+      if (OP_T_THRES <= 3 * OPSIZ && len == 0)
+	return 0;
+      a3 = ((op_t *) srcp1)[0];
+      a0 = ((op_t *) srcp1)[1];
+      b0 = ((op_t *) srcp2)[0];
+      srcp1 += 2 * OPSIZ;
+      srcp2 += 1 * OPSIZ;
+      goto do3;
+    case 1:
+      a2 = ((op_t *) srcp1)[0];
+      a3 = ((op_t *) srcp1)[1];
+      b3 = ((op_t *) srcp2)[0];
+      srcp1 += 2 * OPSIZ;
+      srcp2 += 1 * OPSIZ;
+      len -= 1;
+      if (OP_T_THRES <= 3 * OPSIZ && len == 0)
+	goto do0;
+      /* Fall through.  */
+    }
+
+  do
+    {
+      a0 = ((op_t *) srcp1)[0];
+      b0 = ((op_t *) srcp2)[0];
+      x = DBLALIGN (a2, a3, srcp1i);
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (x != b3))
+	return CMP_LT_OR_GT (x, b3);
+
+    do3:
+      a1 = ((op_t *) srcp1)[0];
+      b1 = ((op_t *) srcp2)[0];
+      x = DBLALIGN (a3, a0, srcp1i);
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (x != b0))
+	return CMP_LT_OR_GT (x, b0);
+
+    do2:
+      a2 = ((op_t *) srcp1)[0];
+      b2 = ((op_t *) srcp2)[0];
+      x = DBLALIGN (a0, a1, srcp1i);
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (x != b1))
+	return CMP_LT_OR_GT (x, b1);
+
+    do1:
+      a3 = ((op_t *) srcp1)[0];
+      b3 = ((op_t *) srcp2)[0];
+      x = DBLALIGN (a1, a2, srcp1i);
+      srcp1 += OPSIZ;
+      srcp2 += OPSIZ;
+      if (__glibc_likely (x != b2))
+	return CMP_LT_OR_GT (x, b2);
+
+      len -= 4;
+    }
+  while (len != 0);
+
+  /* This is the right position for do0.  Please don't move
+     it into the loop.  */
+ do0:
+  x = DBLALIGN (a2, a3, srcp1i);
+  if (__glibc_likely (x != b3))
+    return CMP_LT_OR_GT (x, b3);
+  return 0;
+}
+
+int
+MEMCMP (s1, s2, len)
+     const __ptr_t s1;
+     const __ptr_t s2;
+     size_t len;
+{
+  op_t a0;
+  op_t b0;
+  long int srcp1 = (long int) s1;
+  long int srcp2 = (long int) s2;
+  int res;
+
+  if (len >= OP_T_THRES)
+    {
+      /* There are at least some bytes to compare.  No need to test
+	 for LEN == 0 in this alignment loop.  */
+      while (srcp2 % OPSIZ != 0)
+	{
+	  a0 = ((byte *) srcp1)[0];
+	  b0 = ((byte *) srcp2)[0];
+	  srcp1 += 1;
+	  srcp2 += 1;
+	  res = a0 - b0;
+	  if (__glibc_likely (res != 0))
+	    return res;
+	  len -= 1;
+	}
+
+      /* SRCP2 is now aligned for memory operations on `op_t'.
+	 SRCP1 alignment determines if we can do a simple,
+	 aligned compare or need to shuffle bits.  */
+
+      if (srcp1 % OPSIZ == 0)
+	res = memcmp_common_alignment (srcp1, srcp2, len / OPSIZ);
+      else
+	res = memcmp_not_common_alignment (srcp1, srcp2, len / OPSIZ);
+      if (res != 0)
+	return res;
+
+      /* Number of bytes remaining in the interval [0..OPSIZ-1].  */
+      srcp1 += len & -OPSIZ;
+      srcp2 += len & -OPSIZ;
+      len %= OPSIZ;
+    }
+
+  /* There are just a few bytes to compare.  Use byte memory operations.  */
+  while (len != 0)
+    {
+      a0 = ((byte *) srcp1)[0];
+      b0 = ((byte *) srcp2)[0];
+      srcp1 += 1;
+      srcp2 += 1;
+      res = a0 - b0;
+      if (__glibc_likely (res != 0))
+	return res;
+      len -= 1;
+    }
+
+  return 0;
+}
+libc_hidden_builtin_def(memcmp)
+#ifdef weak_alias
+# undef bcmp
+weak_alias (memcmp, bcmp)
+#endif