gdb/410: -fstrict-aliasing and sim badness

[ac131313 at redhat dot com]

>Number:         410
>Category:       gdb
>Synopsis:       -fstrict-aliasing and sim badness
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Fri Mar 08 14:38:01 PST 2002
>Closed-Date:
>Last-Modified:
>Originator:     ac131313@redhat.com
>Release:        unknown-1.0
>Organization:
>Environment:

>Description:
Below is a reply from an e-mail I sent to gcc@.  This was apparently discovered ~two years ago.  I'm not sure any current src/sim simulators suffer this problem.  I'm mainly adding it for reference.

GDB, is also very much exposed to the same problem.
>How-To-Repeat:
To: Andrew Cagney
CC: gcc@
Subject: Re: -fstrict-aliasing and naughty code?
From: Geoff Keating

Hello,
> 
> I'm trying to understand how to write ``bad'' (host dependant) code
> that doesn't get screwed by strict aliasing.
> 
> For instance, the code snipit:
> 

> > unsigned i;
> > unsigned64 tmp_reg, tmp_reg1;
> > for (i = 0; i < 4; i++)
> > *( (i < 2 ? (unsigned32 *) &tmp_reg
> > : (unsigned32 *) &tmp_reg1)
> > + (1 - i % 2) ) = ...;
> > cpu->registers[...] = tmp_reg;
> >

> 
> I'm told, is bad because:
> 

> > apparently, when -fstrict-aliasing is in effect, gcc is
> > allowed to assume that the expression inside the for loop
> > has no effect on the value of tmp_reg and tmp_reg1, since
> > the assignment is to an object of dissimilar type.

> 
> Provided I make (wild?) assumptions about the host and compiler, can I
> instead write the above to use something like:
> 
> union {
> unsigned64 u64;
> unsigned32 u32[2];
> } tmp_reg, tmp_reg1;
> 
> for (i = 0; i < 4; i++)
> if (i < 2)
> tmp_reg.u32[1 - i % 2] = ...
> else
> tmp_reg1.u32[1 - i %2] = ...;
> cpu->registers[...] = tmp_reg.u64;


Yes, this is documented to work:

     The practice of reading from a different union member than the one
     most recently written to (called "type-punning") is common.  Even
     with `-fstrict-aliasing', type-punning is allowed, provided the
     memory is accessed through the union type.

However, it will be no more efficient than the more portable

  unsigned32 tmp_reg[2], tmp_reg1[2];

  for (i = 0; i < 4; i++)
  if (i < 2)
    tmp_reg[1 - i % 2] = ...
  else
    tmp_reg1[1 - i %2] = ...;
cpu->registers[...] = (unsigned64)tmp_reg[0] << 32 | tmp_reg[1];

in fact it will usually be less efficient because GCC will allocate
registers better for the second example.



>Fix:

>Release-Note:
>Audit-Trail:
>Unformatted: