Re: Continue strict-volatile-bitfields fixes

[Thomas Schwinge <thomas@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 2284 bytes --]

Hi!

On Wed, 18 Apr 2012 18:16:32 +0200, Bernd Schmidt <bernds@codesourcery.com> wrote:
> On 04/18/2012 06:14 PM, Richard Earnshaw wrote:
> > On 18/04/12 16:37, Thomas Schwinge wrote:
> >> gcc/testsuite/
> >> 	* gcc.dg/tree-ssa/20030922-1.c: Compile with
> >> 	-fno-strict-volatile-bitfields.
> >> 	* gcc.dg/tree-ssa/foldconst-3.c: Likewise.
> >> 	* gcc.dg/tree-ssa/vrp15.c: Likewise.
> >>
> > 
> > None of these have any volatile bitfields, so the option should be a no-op.

That's what I thought, too.  :-)

> The problem is that we have to treat normal bitfields differently as
> well, since a variable may later be declared as volatile.

Here is my current test case, reduced from gcc.dg/tree-ssa/20030922-1.c:

    extern void abort (void);
    
    enum tree_code
    {
      BIND_EXPR,
    };
    struct tree_common
    {
      enum tree_code code:8;
    };
    void
    voidify_wrapper_expr (struct tree_common *common)
    {
      switch (common->code)
        {
        case BIND_EXPR:
          if (common->code != BIND_EXPR)
            abort ();
        }
    }

The diff for -fdump-tree-all between compiling with
-fno-strict-volatile-bitfields and -fstrict-volatile-bitfields begins
with the following, right in 20030922-1.c.003t.original:

diff -ru fnsvb/20030922-1.c.003t.original fsvb/20030922-1.c.003t.original
--- fnsvb/20030922-1.c.003t.original    2012-04-19 16:51:18.322150866 +0200
+++ fsvb/20030922-1.c.003t.original     2012-04-19 16:49:18.132088498 +0200
@@ -7,7 +7,7 @@
   switch ((int) common->code)
     {
       case 0:;
-      if (common->code != 0)
+      if ((BIT_FIELD_REF <*common, 32, 0> & 255) != 0)
         {
           abort ();
         }

That is, for -fno-strict-volatile-bitfields the second instance of
»common->code« it is a component_ref, whereas for
-fstrict-volatile-bitfields it is a bit_field_ref.  The former will be
optimized as intended, the latter will not.  So, what should be emitted
in the -fstrict-volatile-bitfields case?  A component_ref -- but this is
what Bernd's commit r182545 (for PR51200) changed, I think?  Or should
later optimization stages learn to better deal with a bit_field_ref, and
where would this have to happen?


Grüße,
 Thomas

[-- Attachment #2: Type: application/pgp-signature, Size: 489 bytes --]