Global constant propagation

[Jeffrey A Law]

  In message < 19990405203851.A5911@daimi.au.dk >you write:
  > On Sun, Apr 04, 1999 at 05:47:47AM -0600, Jeffrey A Law wrote:
  > > 
  > >   In message < 19990404133100.A3351@daimi.au.dk >you write:
  > >   > 
  > >   > Here is a small C fn:
  > >   > 
  > >   > int testfn(int a, int b)
  > >   > {
  > >   >    a = 1;
  > >   >    if(b == 1)
  > >   >    {
  > >   >      b = 2;
  > >   >    }
  > >   >    else
  > >   >    {
  > >   >      b = 3;
  > >   >    }
  > >   >    return a*b;
  > >   > }
  > > 
  > > If the if-then-else is confusing cprop, then that would be a bug.
  > 
  > It turns out it isn't.  The problem is that on MIPS the
  > multiplication is inside a PARALLEL rtl (due to the clobbers
  > of the HI and LO registers as far as I can see), which means
  > the code in gcse.c can't recognise it and doesn't propagate
  > the a=1 into it (see cprop_insn).  Looks like this:
Quite possible.  gcse doesn't try to do any simplifications.  It merely
does a straightforward replacement, then tries to recognize the resulting
insn.


  > (insn 34 51 36 (parallel[ 
  >             (set (reg:SI 85)
  >                 (mult:SI (reg/v:SI 81)
  >                     (reg/v:SI 82)))
  >             (clobber (scratch:SI))
  >             (clobber (scratch:SI))
  >             (clobber (scratch:SI))
  >         ] ) 57 {mulsi3_r4000} (nil)
  >     (nil))
  > 
  > If the a=1 and the multiplication are in the same basic block
  > then it is spotted by the code in cse, which understands
  > parralels better
Local cse does a lot of simplifications on resulting rtx expressions.  That
probably explains why it's the slowest part of the compiler :(


  > (it seems to be doing a rather complete
  > constant propagation of its own, though not much across basic
  > blocks).
Correct.  local cse has only minimal support for cse or cprop across
basic blocks.

  > A solution might be to get gcse.c to recognise parallels that
  > consist only of a 'set' and some clobbers and treat them  like
  > 'set's.  I'm not really into this enough to do that...
Probably still not enough for the mips since I believe the mips requires
that the operands to the multiply to be in registers.  ie, a constant
is not a valid operand to a multiply.  So not only does it have to handle
parallels better, it has to perform simplifications, then recognize the
simplified result.

jeff