Hi all,

     This is the 2/2 patch that contains the functional changes needed
     for MVE Tail Predicated Low Overhead Loops.  See my previous email
     for a general introduction of MVE LOLs.

     This support is added through the already existing loop-doloop
     mechanisms that are used for non-MVE dls/le looping.

     Mid-end changes are:

     1) Relax the loop-doloop mechanism in the mid-end to allow for
        decrement numbers other that -1 and for `count` to be an
        rtx containing a simple REG (which in this case will contain
        the number of elements to be processed), rather
        than an expression for calculating the number of iterations.
     2) Added a new df utility function: `df_bb_regno_only_def_find` that
        will return the DEF of a REG only if it is DEF-ed once within the
        basic block.

     And many things in the backend to implement the above optimisation:

     3)  Implement the `arm_predict_doloop_p` target hook to instruct the
         mid-end about Low Overhead Loops (MVE or not), as well as
         `arm_loop_unroll_adjust` which will prevent unrolling of any loops
         that are valid for becoming MVE Tail_Predicated Low Overhead Loops
         (unrolling can transform a loop in ways that invalidate the dlstp/
         letp tranformation logic and the benefit of the dlstp/letp loop
         would be considerably higher than that of unrolling)
     4)  Appropriate changes to the define_expand of doloop_end, new
         patterns for dlstp and letp, new iterators,  unspecs, etc.
     5) `arm_mve_loop_valid_for_dlstp` and a number of checking functions:
        * `arm_mve_dlstp_check_dec_counter`
        * `arm_mve_dlstp_check_inc_counter`
        * `arm_mve_check_reg_origin_is_num_elems`
        * `arm_mve_check_df_chain_back_for_implic_predic`
        * `arm_mve_check_df_chain_fwd_for_implic_predic_impact`
        This all, in smoe way or another, are running checks on the loop
        structure in order to determine if the loop is valid for dlstp/letp
        transformation.
     6) `arm_attempt_dlstp_transform`: (called from the define_expand of
         doloop_end) this function re-checks for the loop's suitability for
         dlstp/letp transformation and then implements it, if possible.
     7) Various utility functions:
        *`arm_mve_get_vctp_lanes` to map
        from vctp unspecs to number of lanes, and `arm_get_required_vpr_reg`
        to check an insn to see if it requires the VPR or not.
        * `arm_mve_get_loop_vctp`
        * `arm_mve_get_vctp_lanes`
        * `arm_emit_mve_unpredicated_insn_to_seq`
        * `arm_get_required_vpr_reg`
        * `arm_get_required_vpr_reg_param`
        * `arm_get_required_vpr_reg_ret_val`
        * `arm_mve_vec_insn_is_predicated_with_this_predicate`
        * `arm_mve_vec_insn_is_unpredicated_or_uses_other_predicate`

     No regressions on arm-none-eabi with various targets and on
     aarch64-none-elf. Thoughts on getting this into trunk?

     Thank you,
     Stam Markianos-Wright

     gcc/ChangeLog:

             * config/arm/arm-protos.h (arm_target_insn_ok_for_lob): 
Rename to...
             (arm_target_bb_ok_for_lob): ...this
             (arm_attempt_dlstp_transform): New.
             * config/arm/arm.cc (TARGET_LOOP_UNROLL_ADJUST): New.
             (TARGET_PREDICT_DOLOOP_P): New.
             (arm_block_set_vect):
             (arm_target_insn_ok_for_lob): Rename from 
arm_target_insn_ok_for_lob.
             (arm_target_bb_ok_for_lob): New.
             (arm_mve_get_vctp_lanes): New.
             (arm_get_required_vpr_reg): New.
             (arm_get_required_vpr_reg_param): New.
             (arm_get_required_vpr_reg_ret_val): New.
             (arm_mve_get_loop_vctp): New.
(arm_mve_vec_insn_is_unpredicated_or_uses_other_predicate): New.
             (arm_mve_vec_insn_is_predicated_with_this_predicate): New.
             (arm_mve_check_df_chain_back_for_implic_predic): New.
             (arm_mve_check_df_chain_fwd_for_implic_predic_impact): New.
             (arm_mve_check_reg_origin_is_num_elems): New.
             (arm_mve_dlstp_check_inc_counter): New.
             (arm_mve_dlstp_check_dec_counter): New.
             (arm_mve_loop_valid_for_dlstp): New.
             (arm_predict_doloop_p): New.
             (arm_loop_unroll_adjust): New.
             (arm_emit_mve_unpredicated_insn_to_seq): New.
             (arm_attempt_dlstp_transform): New.
             * config/arm/iterators.md (DLSTP): New.
             (mode1): Add DLSTP mappings.
             * config/arm/mve.md (*predicated_doloop_end_internal): New.
             (dlstp<mode1>_insn): New.
             * config/arm/thumb2.md (doloop_end): Update for MVE LOLs.
             * config/arm/unspecs.md: New unspecs.
             * df-core.cc (df_bb_regno_only_def_find): New.
             * df.h (df_bb_regno_only_def_find): New.
             * loop-doloop.cc (doloop_condition_get): Relax conditions.
             (doloop_optimize): Add support for elementwise LoLs.

     gcc/testsuite/ChangeLog:

             * gcc.target/arm/lob.h: Update framework.
             * gcc.target/arm/lob1.c: Likewise.
             * gcc.target/arm/lob6.c: Likewise.
             * gcc.target/arm/mve/dlstp-compile-asm.c: New test.
             * gcc.target/arm/mve/dlstp-int16x8.c: New test.
             * gcc.target/arm/mve/dlstp-int32x4.c: New test.
             * gcc.target/arm/mve/dlstp-int64x2.c: New test.
             * gcc.target/arm/mve/dlstp-int8x16.c: New test.
             * gcc.target/arm/mve/dlstp-invalid-asm.c: New test.