Keeping the rest of variables using VARINT/NT/ANY as ICM_INT as
they are related to shapes and array allocation - will be changed later

reading array elements that have the same first-dimensional index as
array elements that we write in the same with-loop iteration, we know at
compile time that the reads will be local.

(should fix nested with-loops bug)
Optimized write accesses to distributed arrays
Cache can be allocated outside of GASNet segment
Added build scripts

that (X - c) > X is always false [-Wstrict-overflow]" when compiling
stdlib with type checks enabled.
Distributed memory backend:
 - Introduced DSM distribution variant
 - Got broadcasts working.
 - Added option to enable trace output for a specific rank only. 
 - Improved tracing.
 - Extended heap manager tracing.

Added cross compilation target distmemcheck
Fixed various bugs

 - Implemented runtime checks.
 - Added support for primitive modarray functions 
 - Implemented profiling.
 - Improved tracing.
 - Arrays of all types are supported now.
 - Fixed bug in tracing reads from distributed arrays.
 - Fixed bug when calculating the distribution of an array with dim0
size = 0.
 - Updated options for distmem backend.
 - Added compiler phase to identify functions with side effects.
 - Distmem backend: added phase to identify distributable arrays.
 - Introduced execution modes for distmem backend to prevent nodes from
getting out of sync and as a first step to support for nested
with-loops.
 - Fixed bugs related to mirror/descriptor variables for DIS-variables

Status:
 - All basic tests are working, embar.sac is not yet working.

Enabling DBUG_ASSERT expansion in the production compiler as well as
in the development one.

We introduce DBUG_UNREACHABLE macro to mark the places in the code
that must not happen.  We produce an error message stating that we
hit such a place in the code.

    * debug.[ch]:    DBUG_UNREACHABLE macro defined for both DBUG_OFF
                     and !DBUG_OFF cases.

                     exit_function was moved outside of !DBUG_OFF
                     region.

    * ctinfo.[ch]:   Removed CTIterminateCompilationSilient in favour
                     of existing CTIexit and replaced occurences.

    * main.c         Adjusted comments to match doxygen style, set
                     exit_function unconditionally to DBUG_OFF.

Artem Shinkarov authored 12 years ago and Artem Shinkarov committed 11 years ago

Currently sel operation returns a scalar value, which is useless in case
of vector operations.  So _simd_sel_VxA_ (vec_len, idx, array)
acts similarly to sel, and accepts the same arguments, but it then
selects VEC_LEN elements starting from ARRAY[IDX].

Simd multiplication is very similar to the addition:
        _mul_SMxSM_ (vec_len, a, b)

The goal is to make C icm expansion more simple and readable.

        * Replace fprintf(gobal.outfile with `out` and `indout`
          mainly for aesthetic purposes.  Function `indout` is
          indented out, and is expanded to {INDENT; out (...)}.

        * Reorganize ICM templates like COND1, FOR_LOOP, etc
          to avoid statement passing as macro-parameters.  Instead
          every template consists of _BEGIN and _END part, like
          for example:
                FOR_LOOP_BEGIN ("int SAC_i = 0; SAC_i < 3; SAC_i++")
                        // Do something here
                FOR_LOOP_END ();

has two child nodes: Vardecs (instead of Vardec) and Assigns
(instead of Instr).

Mostly, changes for AWLF:

1. Inserted wlbscnf call before SAACYC

2. Implemented cube-slicing in AWLF.
   (Currently disabled, but code is there. The slicing is not
   performed properly in cases where the index vector axes
   are permuted or duplicated, such as in transpose (j,i) or major
   axis selection (i,i).

3. Killed FLATG before SAACYC.

4. Introduced PRFUNR extensions for extrema propagation, post-Loch Ness.
   We now unroll vector_based guards into scalar-based guards.

5. Made extrema N_avis sons, rather than attributes.
   This corrected a fair number of mystery crashes.

6. We now propagate extrema, minima and maxima, independently.

7. Insert SAA AVIS_SHAPE/DIM information within SAACYC now.
   Previously, this was performed only before SAACYC, and each
   optimization was responsible for maintaining SHAPE/DIM
   information itself.

8. Optimizer maintenance of SAA information (item 7) deleted with 
   #ifdef statements, for the nonce, on the assumption that
   item 7 will do the job now.

9.  Introduced CF optimizations for guards.

10. Introduced CF optimizations for extrema.

11. Introduced symbiotic expressions using sacprelude functions
    for AWLF intersect computations.

12. Moved stabilization cycle after SAACYC.
    Placement of this, and of traversals within SAACYC,
    should be vetted.

13. Introduced VP into SCYC to make IVESPLIT work again.

14. Introduced LS into SAACYC, to address Bug#697. See Bugs #709, #710.

15. Introduced scalar constraint/guard functions, for post-Loch Ness AWLF.

16. Fixed WLSIMP AVIS_SSAASSIGN bug.

17. IVE changed to flatten idx2offset and/or vect2offset nodes. This
    was required to make TC/CF, etc., at the end of phase 11 work
    properly. VP/CP at end of phase 11.

18. Added include for globals.h to prepare_inlining.c, to correct
    compilation error.

Known problems:

 - Many rbe-induced warnings during compilation, which I will fix up
   soon. 
 - LIR failure still causing severe performance losses. sah has this
   in hand.
 - AS failures cause AWLF to fail on relatively simple codes, e.g.,
   gauss.sac in awlf test suite.
   Since AS no longer exists, this is not surprising. I will
   attempt its revival, since symbolic CF can not do the job
   unless operands are shuffled.
 - I am of the opinion that stdlib take() and perhaps drop()
   generate bad code for negative take counts. That causes
   sac/apex/UTThornBoolean/UTThornBoolean.sac to crash when compiled
   with -ecc, with an error message that is correct, but which
   requires the user to be able to read IL in order to isolate.
 - I would appreciate code review/inspection of any/all of these
   documents. In particular, anything labeled FIXME should be
   checked.
 - -check c will fail for some newly introduced scalar guards. 
   There is something wrong memory/alloc.c for these guards,
   and the module does not offer any guidance on writing code
   for new primitives.
 - ct_prf.c: fails to completely typecheck some new guards,
   I think. Same problem as with -check c: no docn in code.

and their corresponding return type does not match. This is done to
ensure that the backend emits corresponding representation adjustments.

handling module str.h rather than the standard string.h.
Removed superfluous includes of string.h stroughout sac2c.

and ND_PRF_MODARRAY_AxVxS__DATA_arr.

This will be needed to fix bug 386 (WLUR).
Added new prfs idx_modarray_AxSxS and idx_modarray_AxSxA to
reflect changes in modarray and to have a clear separation between
scalar valued and array valued modarrays in the code generation.
Stripped dead prf switches out of compile, makedimexpr and makeshapeexr
after they have been replaced by proper function tables.

Mac file prf_info.mac rewritten.
Outdated table entries removed and functionality re-implemented.
Code generation for prfs re-implemented.
Some features added to string library.

and the associated ICMs renamed to fully qualified names:
dim_A, shape_A, sel_VxA, reshape_VxA, modarray_AxVxS.

However, internal_lib itself has not yet been removed from system.

compiler module support/str.
Extended refactoring mechanism by additional condition.

 - Compile prop_in and prop_out to C ICM's
 - Implement them for ST and MT

seperate copy prf.

with     ICMCompileND_SET__SHAPE_arr( to_NT, to_dim, shp);
in ND_PRF_SEL__SHAPE_id, ND_PRF_SEL__SHAPE_arr, ND_PRF_IDX_SEL__SHAPE
This should fix bug #62, but it would be interesting to know why there
was 1 instead of to_dim in the first place.