Bug 106513 - [10 Regression] bswap pass misses that >>56 for signed types can be replicate the sign bit
Summary: [10 Regression] bswap pass misses that >>56 for signed types can be replicate...
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: tree-optimization (show other bugs)
Version: 13.0
: P2 normal
Target Milestone: 10.5
Assignee: Richard Biener
URL:
Keywords: wrong-code
Depends on:
Blocks:
 
Reported: 2022-08-03 11:56 UTC by Krister Walfridsson
Modified: 2022-10-14 10:48 UTC (History)
1 user (show)

See Also:
Host:
Target:
Build:
Known to work: 10.4.1, 11.3.1, 12.1.1, 13.0, 4.4.7
Known to fail: 10.4.0, 11.3.0, 12.1.0, 4.5.3, 4.6.4, 4.9.0
Last reconfirmed: 2022-08-05 00:00:00


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Krister Walfridsson 2022-08-03 11:56:49 UTC
GCC may incorrectly generate bswap instructions for code not doing a correct swap. This can be seen by running the function from testsuite/gcc.dg/pr40501.c as

typedef long int int64_t;

__attribute__((noinline)) int64_t
swap64 (int64_t n)
{
  return (((n & (((int64_t) 0xff) )) << 56) |
          ((n & (((int64_t) 0xff) << 8)) << 40) |
          ((n & (((int64_t) 0xff) << 16)) << 24) |
          ((n & (((int64_t) 0xff) << 24)) << 8) |
          ((n & (((int64_t) 0xff) << 32)) >> 8) |
          ((n & (((int64_t) 0xff) << 40)) >> 24) |
          ((n & (((int64_t) 0xff) << 48)) >> 40) |
          ((n & (((int64_t) 0xff) << 56)) >> 56));
}

int main (void)
{
  volatile int64_t n = 0x8000000000000000l;

  if (swap64(n) != 0xffffffffffffff80l)
    __builtin_abort ();

  return 0;
}

This fails at -Os and higher optimization levels.
Comment 1 Andreas Schwab 2022-08-03 12:12:20 UTC
This subexpression has undefined behaviour: (((int64_t) 0xff) << 56).
Comment 2 Krister Walfridsson 2022-08-03 12:41:31 UTC
(In reply to Andreas Schwab from comment #1)
> This subexpression has undefined behaviour: (((int64_t) 0xff) << 56).

I thought that was allowed in GCC as the manual says (https://gcc.gnu.org/onlinedocs/gcc-12.1.0/gcc/Integers-implementation.html#Integers-implementation) "As an extension to the C language, GCC does not use the latitude given in C99 and C11 only to treat certain aspects of signed ‘<<’ as undefined."

If not, what behavior does the manual refer to?
Comment 3 Andrew Pinski 2022-08-05 22:15:25 UTC
Confirmed.

Better testcase (without the questionable undefined behavior):

typedef long long int int64_t;

__attribute__((noinline)) int64_t
swap64 (int64_t n)
{
  return (((n & (((int64_t) 0xff) )) << 56) |
          ((n & (((int64_t) 0xff) << 8)) << 40) |
          ((n & (((int64_t) 0xff) << 16)) << 24) |
          ((n & (((int64_t) 0xff) << 24)) << 8) |
          ((n & (((int64_t) 0xff) << 32)) >> 8) |
          ((n & (((int64_t) 0xff) << 40)) >> 24) |
          ((n & (((int64_t) 0xff) << 48)) >> 40) |
          ((n & ((int64_t)(0xffull << 56))) >> 56));
}

int main (void)
{
  volatile int64_t n = 0x8000000000000000l;

  if (swap64(n) != 0xffffffffffffff80l)
    __builtin_abort ();

  return 0;
}
Comment 4 Richard Biener 2022-08-10 13:43:30 UTC
Mine.

diff --git a/gcc/gimple-ssa-store-merging.cc b/gcc/gimple-ssa-store-merging.cc
index 0640168bcc4..b80b8eac444 100644
--- a/gcc/gimple-ssa-store-merging.cc
+++ b/gcc/gimple-ssa-store-merging.cc
@@ -263,7 +263,7 @@ do_shift_rotate (enum tree_code code,
                 int count)
 {
   int i, size = TYPE_PRECISION (n->type) / BITS_PER_UNIT;
-  unsigned head_marker;
+  uint64_t head_marker;
 
   if (count < 0
       || count >= TYPE_PRECISION (n->type)
Comment 5 GCC Commits 2022-08-10 14:39:25 UTC
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:f675afa4eeac9910a2c085a95aa04d6d9f2fd8d6

commit r13-2013-gf675afa4eeac9910a2c085a95aa04d6d9f2fd8d6
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Aug 10 15:45:22 2022 +0200

    tree-optimization/106513 - fix mistake in bswap symbolic number shifts
    
    This fixes a mistake in typing a local variable in the symbolic
    shift routine.
    
            PR tree-optimization/106513
            * gimple-ssa-store-merging.cc (do_shift_rotate): Use uint64_t
            for head_marker.
    
            * gcc.dg/torture/pr106513.c: New testcase.
Comment 6 GCC Commits 2022-08-10 14:40:09 UTC
The releases/gcc-12 branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:ab2ca2dbd528f0564b80fa0e6eda96e0237742bc

commit r12-8677-gab2ca2dbd528f0564b80fa0e6eda96e0237742bc
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Aug 10 15:45:22 2022 +0200

    tree-optimization/106513 - fix mistake in bswap symbolic number shifts
    
    This fixes a mistake in typing a local variable in the symbolic
    shift routine.
    
            PR tree-optimization/106513
            * gimple-ssa-store-merging.cc (do_shift_rotate): Use uint64_t
            for head_marker.
    
            * gcc.dg/torture/pr106513.c: New testcase.
    
    (cherry picked from commit f675afa4eeac9910a2c085a95aa04d6d9f2fd8d6)
Comment 7 GCC Commits 2022-10-11 13:04:40 UTC
The releases/gcc-11 branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:ec718fb7a3a58955017cd80f6a8927afbe340fc4

commit r11-10306-gec718fb7a3a58955017cd80f6a8927afbe340fc4
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Aug 10 15:45:22 2022 +0200

    tree-optimization/106513 - fix mistake in bswap symbolic number shifts
    
    This fixes a mistake in typing a local variable in the symbolic
    shift routine.
    
            PR tree-optimization/106513
            * gimple-ssa-store-merging.c (do_shift_rotate): Use uint64_t
            for head_marker.
    
            * gcc.dg/torture/pr106513.c: New testcase.
    
    (cherry picked from commit f675afa4eeac9910a2c085a95aa04d6d9f2fd8d6)
Comment 8 GCC Commits 2022-10-14 10:47:45 UTC
The releases/gcc-10 branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:29ad829c93d530e75dee88d2550350c14fcaf8f6

commit r10-11038-g29ad829c93d530e75dee88d2550350c14fcaf8f6
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Aug 10 15:45:22 2022 +0200

    tree-optimization/106513 - fix mistake in bswap symbolic number shifts
    
    This fixes a mistake in typing a local variable in the symbolic
    shift routine.
    
            PR tree-optimization/106513
            * gimple-ssa-store-merging.c (do_shift_rotate): Use uint64_t
            for head_marker.
    
            * gcc.dg/torture/pr106513.c: New testcase.
    
    (cherry picked from commit f675afa4eeac9910a2c085a95aa04d6d9f2fd8d6)
Comment 9 Richard Biener 2022-10-14 10:48:55 UTC
Fixed.