Re: [PATCH][GCC][mid-end] Fix PR85123 incorrect copies

[Richard Biener <rguenther@suse.de>]

On Fri, 6 Apr 2018, Alan Modra wrote:

> On Thu, Apr 05, 2018 at 01:29:06PM +0100, Tamar Christina wrote:
> > diff --git a/gcc/expr.c b/gcc/expr.c
> > index 00660293f72e5441a6421a280b04c57fca2922b8..7daeb8c91d758edf0b3dc37f6927380b6f3df877 100644
> > --- a/gcc/expr.c
> > +++ b/gcc/expr.c
> > @@ -2749,7 +2749,7 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src)
> >  {
> >    int i, n_regs;
> >    unsigned HOST_WIDE_INT bitpos, xbitpos, padding_correction = 0, bytes;
> > -  unsigned int bitsize;
> > +   unsigned int bitsize = 0;
> >    rtx *dst_words, dst, x, src_word = NULL_RTX, dst_word = NULL_RTX;
> >    /* No current ABI uses variable-sized modes to pass a BLKmnode type.  */
> >    fixed_size_mode mode = as_a <fixed_size_mode> (mode_in);
> > @@ -2782,7 +2782,7 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src)
> >  
> >    n_regs = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
> >    dst_words = XALLOCAVEC (rtx, n_regs);
> > -  bitsize = BITS_PER_WORD;
> > +
> >    if (targetm.slow_unaligned_access (word_mode, TYPE_ALIGN (TREE_TYPE (src))))
> >      bitsize = MIN (TYPE_ALIGN (TREE_TYPE (src)), BITS_PER_WORD);
> >  
> 
> You calculate bitsize here, then override it in the loop?  Doesn't
> that mean strict align targets will use mis-aligned loads and stores?
> 
> > @@ -2791,6 +2791,17 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src)
> >         bitpos < bytes * BITS_PER_UNIT;
> >         bitpos += bitsize, xbitpos += bitsize)
> >      {
> > +      /* Find the largest integer mode that can be used to copy all or as
> > +	 many bits as possible of the structure.  */
> > +      opt_scalar_int_mode mode_iter;
> > +      FOR_EACH_MODE_IN_CLASS (mode_iter, MODE_INT)
> > +      if (GET_MODE_BITSIZE (mode_iter.require ())
> > +			    <= ((bytes * BITS_PER_UNIT) - bitpos)
> > +	  && GET_MODE_BITSIZE (mode_iter.require ()) <= BITS_PER_WORD)
> > +	bitsize = GET_MODE_BITSIZE (mode_iter.require ());
> > +      else
> > +	break;
> > +
> 
> This isn't correct.  Consider a 6 byte struct on a 4 byte word, 8 bit
> byte, big-endian target when targetm.calls.return_in_msb is false.
> 
> In this scenario, copy_blkmode_to_reg should return two registers, set
> as if they had been loaded from two words in memory laid out as
> follows (left to right increasing byte addresses):
>  _______________________         _______________________
> |  0  |  0  |  s0 |  s1 |       |  s2 |  s3 |  s4 |  s5 |
>  ~~~~~~~~~~~~~~~~~~~~~~~         ~~~~~~~~~~~~~~~~~~~~~~~
> 
> So we will have xbitpos=16 first time around the loop.  That means
> your new code will attempt to store 32 bits into a bit-field starting
> at bit 16 in the first 32-bit register, and of course fail.
> 
> This scenario used to be handled correctly, at least when the struct
> wasn't over-aligned.

I wonder if it's best to revert the original regression causing patch
and look for a proper solution in the GCC 9 timeframe?

Thanks,
Richard.