On 09/03/2015 10:18 AM, Segher Boessenkool wrote:
But there are more efficient ways to emit single bit and masks tests that apply
to most CPUs rather than doing something specific that works for just one target
only. For example single bit test is a simple shift into carry flag or into the
sign bit, and for mask tests, you shift out all the non-mask bits.
Most of those are quite target-specific. Some others are already done,
and/or done by other passes.
But what combine does here is even more target-specific.
Combine puts everything (well, most things) through
make_compound_operation, on all targets.
Note huge parts of combine are structured around the needs of processors
from the late 80s and early 90s. The canonical forms selected for those
processors may not be optimal for today's processors.
The problem (of course) is changing the canonical forms can be a ton of
work in both the backends as well as combine itself to ensure quality of
code doesn't regress.
Combine converts the merged instructions to what it thinks is the
canonical or cheapest form, and uses that. It does not try multiple
options (the zero_ext* -> and+shift rewriting is not changing the
semantics of the pattern at all).
But the change from AND to zero_extract is already changing semantics...
Oh? It is not supposed to!
Combine should never change semantics. It can change form and may
change what happens to "don't care" bits. But it should never change
visible semantics.
Jeff
