Improve VRP, partial fix pr26344

[Jeffrey A Law <law at redhat dot com>]

I initially thought that the best solution for these problems
was going to be to move the forwprop pass.  While that approach
looked appealing (and still has some overall appeal), it's
no longer my preferred approach.  Basically it's unclear at this
time if it's really going to be better or if it's just going to
trade one set of missed micro optimizations for another set of
missed micro optimizations.

So I went back and revisited VRP in the hope of finding a way
to safely and efficiently implement the backwards propagation of
non-null ranges through pointer typecasts.

For those not familiar with VRP and the propagation engine it
uses, it is *critical* that we not go backwards through the 
lattice.  ie, once something is VR_VARYING, it does not go
back to VR_RANGE or VR_ANTI_RANGE.

It is also important to not generate lots of useless ASSERT_EXPRs
from a compile-time performance standpoint.  For example, if
a pointer is used once and only once in a dereference, we do not
want to generate an ASSERT_EXPR for it claiming it has a non-null
range.  Note this need makes it virtually impossible to perform
the backwards propagation after the propagation engine has run
(we won't have ASSERT_EXPRs for the key pointers and thus no ranges
for them).

The good news is initial discovery of non-null pointers occurs
before the VRP propagation engine runs.  If we could do the
backwards propagation during that initial non-null discovery,
then we would have a fighting chance.  And that's exactly what
this patch does.

With this patch we always attempt to infer a value range, even
for single use operands -- note carefully that we still suppress
creating ASSERT_EXPRs for such operands to avoid compile time
regressions.  When we discover a non-null pointer, we then walk
up its use-def chain to see if the pointer was set via a NOP_EXPR
type conversion from another pointer type.  If we find such a
type conversion and the source pointer for that type conversion is
used more than once, then we register an ASSERT_EXPR for the
source pointer in the NOP_EXPR.

This fixes two of the three gcc.dg/tree-ssa missed optimization
regressions as well as the one reported for C++.

Bootstrapped and regression tested on i686-pc-linux-gnu.

Attachment: PPP
Description: Text document