[FORTRAN PATCH] Optimize FORALLs with constant mask expressions

["Roger Sayle" <roger at eyesopen dot com>]

The following simple patch tweaks the way that we expand conditional
FORALL statements in gfortran, to improve the case where the conditional
expression (mask) is a compile-time constant, such as .true. or .false. 
Although unlikely to occur commonly in real code, these are obvious
optimizations that are easily implemented before more complex
dependency-based transformations.

I'm wondering what the gfortran maintainers think about implementing these
sort of source-to-source transformations earlier, perhaps during parsing
or a simplification pass?  It'd would simplify translation if "where
(.true.) a(:) = b(:)" were previously lowered to "a(:) = b(:)" or if
"elsewhere (.false.)" were eliminated, etc...  I'm not sure where best it
would be appropriate to canonicalize "forall (..., .true.)" into "forall
(...)", but prior to this  patch we'd generate significantly different
code for each.

The following code has been tested on x86_64-unknown-linux-gnu with a full
"make bootstrap", including gfortran, and regression tested with a
top-level "make -k check" with no new failures.

Ok for mainline?


2007-02-07  Roger Sayle  <roger@eyesopen.com>

        * trans-stmt.c (gfc_trans_forall_1): Optimize the cases where the
        mask expression is a compile-time constant (".true." or ".false.").

        * gfortran.dg/forall_8.f90: New test case.
        * gfortran.dg/forall_9.f90: Likewise.

Roger
--

Attachment: patchf.txt
Description: Text document

Attachment: forall_8.f90
Description: Binary data

Attachment: forall_9.f90
Description: Binary data