[PATCH] Modulo-scheduling improvements. Patch 1 of 2.

[Revital1 Eres <ERES at il dot ibm dot com>]

Hello,

This patch includes the first patch Vladimir had posted to improve
modulo-schedualing; without allowing reg-moves by limiting the scheduling
window.  (http://gcc.gnu.org/ml/gcc-patches/2007-01/msg01468.html).
I intend to add the enablement of reg-moves once I'll finish it's testing.

This patch also avoids performing SMS when the candidate loop contains
auto-increment instruction.

A new reduced Fortran testcase for forall_10.f90 test (sms1.f90) was
added; this testcase failed with -fmodulo-sched flag when the current
patch is not applied.

This patch was bootstrapped and tested on ppc64 with -fmodulo-sched
flag. (all languages except Ada)

OK for mainline?

Thanks,
Revital

(See attached file: sms-antideps.txt)(See attached file: sms-1.f90.txt)(See
attached file: patch_sms_15_7.txt)

 2007-07-17  Vladimir Yanovsky  <yanov@il.ibm.com>

        * ddg.c (create_ddg_dependence): Don't ignore ANTI_DEPs.
        * modulo-sched.c (print_node_sched_params): Add parameter and
        verbosity.
        (calculate_maxii): Remove function.
        (undo_generate_reg_moves): Likewise.
        (undo_permute_partial_schedule): Likewise.
        (kernel_number_of_cycles): Likewise.
        (MAXII_FACTOR): New definition to calculate the upper bound of II.
        (sms_schedule): Use it.  Remove profitability checks and avoid
loops
        which includes auto-increment instructions.
        (get_sched_window): Prevent necessitating regmoves by limiting
        the scheduling window.
        (sms_schedule_by_order): Use must_follow & must_precede bitmaps
        to determine order of nodes within the cycle.

        * testsuite/gcc.dg/sms-antideps.c: New test.
        * testsuite/gfortran.dg/sms1.f90: New test.

Attachment: sms-antideps.txt
Description: Text document

Attachment: sms-1.f90.txt
Description: Text document

Attachment: patch_sms_15_7.txt
Description: Text document