[Bug target/114605] New: [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[Bug target/114605] New: [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

            Bug ID: 114605
           Summary: [14 Regression] wrong code with -march=z13 -O0 since
                    r14-5831
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

Since r14-5831-gaae723d360ca26cd9fd0b039fb0a616bd0eae363 we miscompile the
following testcase on s390x-linux with -O0 -march=z13:
typedef struct { const float *a; int b, c; float *d; } S;

__attribute__((noipa)) void
bar (void)
{
}

__attribute__((noinline, optimize (2))) static void
foo (S *e)
{
  const float *f;
  float *g;
  float h[4] = { 0.0, 0.0, 1.0, 1.0 };
  if (!e->b)
    f = h;
  else
    f = e->a;
  g = &e->d[0];
  __builtin_memcpy (g, f, sizeof (float) * 4);
  bar ();
  if (!e->b)
    if (g[0] != 0.0 || g[1] != 0.0 || g[2] != 1.0 || g[3] != 1.0)
      __builtin_abort ();
}

int
main ()
{
  float d[4];
  S e = { .d = d };
  foo (&e);
  return 0;
}

The optimized dump difference from the previous commit is
@@ -23,7 +23,6 @@ Removing basic block 17
 __attribute__((optimize (2), noinline))
 void foo (struct S * e)
 {
-  float h[4];
   float * g;
   const float * f;
   int _1;
@@ -32,10 +31,10 @@ void foo (struct S * e)
   float _4;
   float _5;
   float _6;
-  uint128_t _13;
+  uint128_t pretmp_18;
+  uint128_t prephitmp_19;

   <bb 2> [local count: 1073741824]:
-  h = *.LC0;
   _1 = e_10(D)->b;
   if (_1 == 0)
     goto <bb 4>; [50.00%]
@@ -44,12 +43,12 @@ void foo (struct S * e)

   <bb 3> [local count: 536870912]:
   f_11 = e_10(D)->a;
+  pretmp_18 = MEM <uint128_t> [(char * {ref-all})f_11];

   <bb 4> [local count: 1073741824]:
-  # f_7 = PHI <&h(2), f_11(3)>
+  # prephitmp_19 = PHI <4575657222473777152(2), pretmp_18(3)>
   g_12 = e_10(D)->d;
-  _13 = MEM <uint128_t> [(char * {ref-all})f_7];
-  MEM <uint128_t> [(char * {ref-all})g_12] = _13;
+  MEM <uint128_t> [(char * {ref-all})g_12] = prephitmp_19;
   bar ();
   _2 = e_10(D)->b;
   if (_2 == 0)
@@ -89,7 +88,6 @@ void foo (struct S * e)
   __builtin_abort ();

   <bb 10> [local count: 1073741824]:
-  h ={v} {CLOBBER(eol)};
   return;

 }
4575657222473777152 is 0x3f8000003f800000 so on big endian that looks
reasonable.

[Bug target/114605] [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0
           Priority|P3                          |P1

[Bug target/114605] [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-04-05
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This is a clear bug in s390_const_int_pool_entry_p.  It just uses
get_pool_constant but ignores get_pool_mode, which on big-endian is
significant.  The function doesn't allow offsets like e.g. simplify-rtx.cc
(avoid_constant_pool_reference), but still needs to deal with mode mismatches.

[Bug target/114605] [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 57889
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57889&action=edit
gcc14-pr114605.patch

Untested fix.

[Bug target/114605] [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:d5d84487dec06186fd9246b505f44ef68a66d6a2

commit r14-9834-gd5d84487dec06186fd9246b505f44ef68a66d6a2
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Mon Apr 8 16:22:13 2024 +0200

    s390: Fix s390_const_int_pool_entry_p and movdi peephole2 [PR114605]

    The following testcase is miscompiled, because we have initially
    a movti which loads the 0x3f8000003f800000ULL TImode constant
    from constant pool.  Later on we split it into a pair of DImode
    loads.  Now, for the first load (why just that?, though not stage4
    material) we trigger the peephole2 which uses s390_const_int_pool_entry_p.
    That function doesn't check at all the constant pool mode though, sees
    the constant pool at that address has a CONST_INT value and just assumes
    that is the value to return, which is especially wrong for big-endian,
    if it is a DImode load from offset 0, it should be loading 0 rather than
    0x3f8000003f800000ULL.
    The following patch adds checks if we are extracing a MODE_INT mode,
    if the constant pool has MODE_INT mode as well, punts if constant pool
    has smaller mode size than the extraction one (then it would be UB),
    if it has the same mode as before keeps using what it did before,
    if constant pool has a larger mode than the one being extracted, uses
    simplify_subreg.  I'd have used avoid_constant_pool_reference
    instead which can handle also offsets into the constant pool constants,
    but it can't handle UNSPEC_LTREF.

    Another thing is that once that is fixed, we ICE when we extract constant
    like 0, ior insn predicate require non-0 constant.  So, the patch also
    fixes the peephole2 so that if either 32-bit half is zero, it uses a mere
    load of the constant into register rather than a pair of such load and ior.

    2024-04-08  Jakub Jelinek  <jakub@redhat.com>

            PR target/114605
            * config/s390/s390.cc (s390_const_int_pool_entry_p): Punt
            if mem doesn't have MODE_INT mode, or pool constant doesn't
            have MODE_INT mode, or if pool constant mode is smaller than
            mem mode.  If mem mode is different from pool constant mode,
            try to simplify subreg.  If that doesn't work, punt, if it
            does, use the simplified constant instead of the constant pool
            constant.
            * config/s390/s390.md (movdi from const pool peephole): If
            either low or high 32-bit part is zero, just emit move insn
            instead of move + ior.

            * gcc.dg/pr114605.c: New test.

[Bug target/114605] [14 Regression] wrong code with -march=z13 -O0 since r14-5831

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114605

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed.