This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Another GLIBC build error with GCC6


Roland McGrath wrote:
the left shift doesn't even lose any bits, because the high
bit is already zero (and it's a constant, so the compiler knows that).

But it does lose a bit. It loses a zero bit, which can't be recovered the way that a non-overflowing signed left shift is recovered by shifting right the same amount.

that's the whole point of the thing: to sign-extend the 31-bit value
to 32 bits.  I don't think it should complain about 0x7fffffff << 1.

Such complaints are useful for compilers that take advantage of C's overflow rules to generate efficient-but-weird code when you violate the rules (something that describes GCC more and more nowadays...).

Anyway, if I understand this one correctly it should be easy enough to fix portably. Something like the attached patch, say. (I haven't tested it.)
>From 64586c5448f6f5537bd36bb5f271324790055c52 Mon Sep 17 00:00:00 2001
From: Paul Eggert <eggert@cs.ucla.edu>
Date: Tue, 21 Jul 2015 22:50:29 -0700
Subject: [PATCH] Port the 0x7efe...feff pattern to GCC 6.

See Steve Ellcey's bug report in:
https://sourceware.org/ml/libc-alpha/2015-07/msg00673.html
* string/memrchr.c (MEMRCHR):
* string/rawmemchr.c (RAWMEMCHR):
* string/strchr.c (strchr):
* string/strchrnul.c (STRCHRNUL):
Rewrite code to avoid issues with signed shift overflow.
---
 ChangeLog          | 11 +++++++++++
 string/memrchr.c   | 11 ++---------
 string/rawmemchr.c | 11 ++---------
 string/strchr.c    |  9 ++-------
 string/strchrnul.c |  9 ++-------
 5 files changed, 19 insertions(+), 32 deletions(-)

diff --git a/ChangeLog b/ChangeLog
index 652d968..9f87cd9 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,14 @@
+2015-07-21  Paul Eggert  <eggert@cs.ucla.edu>
+
+	Port the 0x7efe...feff pattern to GCC 6.
+	See Steve Ellcey's bug report in:
+	https://sourceware.org/ml/libc-alpha/2015-07/msg00673.html
+	* string/memrchr.c (MEMRCHR):
+	* string/rawmemchr.c (RAWMEMCHR):
+	* string/strchr.c (strchr):
+	* string/strchrnul.c (STRCHRNUL):
+	Rewrite code to avoid issues with signed shift overflow.
+
 2015-07-22  Mike Frysinger  <vapier@gentoo.org>
 
 	* sysdeps/unix/sysv/linux/ia64/bits/msq.h: Change sys/types.h include
diff --git a/string/memrchr.c b/string/memrchr.c
index 0c8fd84..86cd5b9 100644
--- a/string/memrchr.c
+++ b/string/memrchr.c
@@ -96,15 +96,8 @@ MEMRCHR
 
      The 1-bits make sure that carries propagate to the next 0-bit.
      The 0-bits provide holes for carries to fall into.  */
-
-  if (sizeof (longword) != 4 && sizeof (longword) != 8)
-    abort ();
-
-#if LONG_MAX <= LONG_MAX_32_BITS
-  magic_bits = 0x7efefeff;
-#else
-  magic_bits = ((unsigned long int) 0x7efefefe << 32) | 0xfefefeff;
-#endif
+  magic_bits = -1;
+  magic_bits = magic_bits / 0xff * 0xfe << 1 >> 1 | 1;
 
   /* Set up a longword, each of whose bytes is C.  */
   charmask = c | (c << 8);
diff --git a/string/rawmemchr.c b/string/rawmemchr.c
index 05b22be..228ca9d 100644
--- a/string/rawmemchr.c
+++ b/string/rawmemchr.c
@@ -86,15 +86,8 @@ RAWMEMCHR (s, c_in)
 
      The 1-bits make sure that carries propagate to the next 0-bit.
      The 0-bits provide holes for carries to fall into.  */
-
-  if (sizeof (longword) != 4 && sizeof (longword) != 8)
-    abort ();
-
-#if LONG_MAX <= LONG_MAX_32_BITS
-  magic_bits = 0x7efefeff;
-#else
-  magic_bits = ((unsigned long int) 0x7efefefe << 32) | 0xfefefeff;
-#endif
+  magic_bits = -1;
+  magic_bits = magic_bits / 0xff * 0xfe << 1 >> 1 | 1;
 
   /* Set up a longword, each of whose bytes is C.  */
   charmask = c | (c << 8);
diff --git a/string/strchr.c b/string/strchr.c
index 5f90075..f13b2b3 100644
--- a/string/strchr.c
+++ b/string/strchr.c
@@ -60,13 +60,8 @@ strchr (const char *s, int c_in)
 
      The 1-bits make sure that carries propagate to the next 0-bit.
      The 0-bits provide holes for carries to fall into.  */
-  switch (sizeof (longword))
-    {
-    case 4: magic_bits = 0x7efefeffL; break;
-    case 8: magic_bits = ((0x7efefefeL << 16) << 16) | 0xfefefeffL; break;
-    default:
-      abort ();
-    }
+  magic_bits = -1;
+  magic_bits = magic_bits / 0xff * 0xfe << 1 >> 1 | 1;
 
   /* Set up a longword, each of whose bytes is C.  */
   charmask = c | (c << 8);
diff --git a/string/strchrnul.c b/string/strchrnul.c
index 2678f1d..daf0b3f 100644
--- a/string/strchrnul.c
+++ b/string/strchrnul.c
@@ -66,13 +66,8 @@ STRCHRNUL (s, c_in)
 
      The 1-bits make sure that carries propagate to the next 0-bit.
      The 0-bits provide holes for carries to fall into.  */
-  switch (sizeof (longword))
-    {
-    case 4: magic_bits = 0x7efefeffL; break;
-    case 8: magic_bits = ((0x7efefefeL << 16) << 16) | 0xfefefeffL; break;
-    default:
-      abort ();
-    }
+  magic_bits = -1;
+  magic_bits = magic_bits / 0xff * 0xfe << 1 >> 1 | 1;
 
   /* Set up a longword, each of whose bytes is C.  */
   charmask = c | (c << 8);
-- 
2.1.0


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]