This is a follow up patch for pr83488 to fix an error in setting...

This is a follow up patch for pr83488 to fix an error in setting
OPTION_MASK_ISA_AVX512VNNI_SET and OPTION_MASK_ISA_AVX512F_SET bits.
There were both set in ix86_isa_flags2 while being defined in
different ISA sets. Additionally move OPTION_MASK_ISA_AVX512VNNI_SET
to ix86_isa_flags as it can be used with OPTION_MASK_ISA_AVX512VL_SET.

gcc/
* common/config/i386/i386-common.c (OPTION_MASK_ISA_AVX512VNNI_SET):
Or in OPTION_MASK_ISA_AVX512F_SET.
(OPTION_MASK_ISA_AVX512F_UNSET): Or in
OPTION_MASK_ISA_AVX512VNNI_UNSET.
(ix86_handle_option): Adjust for
OPTION_MASK_ISA_AVX512VNNI_*SET being in ix86_isa_flags.
* config/i386/i386-builtin.def: Move VNNI builtins from ARGS2
section to ARGS.
* config/i386/i386-c.c: Check for OPTION_MASK_ISA_AVX512VNNI in
isa_flag instead of isa_flag2.
* config/i386/i386.c (ix86_target_string): Move -mavx512vnni from
isa_opts2 to isa_opts.
* config/i386/i386.opt (mavx512vnni): Move from ix86_isa_flags2
to ix86_isa_flags.

From-SVN: r255974

extend.texi (Loop-Specific Pragmas): Document pragma GCC unroll.

* doc/extend.texi (Loop-Specific Pragmas): Document pragma GCC unroll.
c-family/
* c-pragma.c (init_pragma): Register pragma GCC unroll.
* c-pragma.h (enum pragma_kind): Add PRAGMA_UNROLL.
c/
* c-parser.c (c_parser_while_statement): Add unroll parameter and
build ANNOTATE_EXPR if present.  Add 3rd operand to ANNOTATE_EXPR.
(c_parser_do_statement): Likewise.
(c_parser_for_statement): Likewise.
(c_parser_statement_after_labels): Adjust calls to above.
(c_parse_pragma_ivdep): New static function.
(c_parser_pragma_unroll): Likewise.
(c_parser_pragma) <PRAGMA_IVDEP>: Add support for pragma Unroll.
<PRAGMA_UNROLL>: New case.
cp/
* constexpr.c (cxx_eval_constant_expression) <ANNOTATE_EXPR>: Remove
assertion on 2nd operand.
(potential_constant_expression_1): Likewise.
* cp-tree.def (RANGE_FOR_STMT): Take a 5th operand.
* cp-tree.h (RANGE_FOR_UNROLL): New macro.
(cp_convert_range_for): Adjust prototype.
(finish_while_stmt_cond): Likewise.
(finish_do_stmt): Likewise.
(finish_for_cond): Likewise.
* init.c (build_vec_init): Adjut call to finish_for_cond.
* parser.c (cp_parser_statement): Adjust call to
cp_parser_iteration_statement.
(cp_parser_for): Add unroll parameter and pass it in calls to
cp_parser_range_for and cp_parser_c_for.
(cp_parser_c_for): Add unroll parameter and pass it in call to
finish_for_cond.
(cp_parser_range_for): Add unroll parameter, set in on RANGE_FOR_STMT
and pass it in call to cp_convert_range_for.
(cp_convert_range_for): Add unroll parameter and pass it in call to
finish_for_cond.
(cp_parser_iteration_statement): Add unroll parameter and pass it in
calls to finish_while_stmt_cond, finish_do_stmt and cp_parser_for.
(cp_parser_pragma_ivdep): New static function.
(cp_parser_pragma_unroll): Likewise.
(cp_parser_pragma) <PRAGMA_IVDEP>: Add support for pragma Unroll.
<PRAGMA_UNROLL>: New case.
* pt.c (tsubst_expr) <FOR_STMT>: Adjust call to finish_for_cond.
<RANGE_FOR_STMT>: Pass unrolling factor to cp_convert_range_for.
<WHILE_STMT>: Adjust call to finish_while_stmt_cond.
<DO_STMT>: Adjust call to finish_do_stmt.
* semantics.c (finish_while_stmt_cond): Add unroll parameter and
build ANNOTATE_EXPR if present.
(finish_do_stmt): Likewise.
(finish_for_cond): Likewise.
(begin_range_for_stmt): Build RANGE_FOR_STMT with 5th operand.
fortran/
* array.c (gfc_copy_iterator): Copy unroll field.
* decl.c (directive_unroll): New global variable.
(gfc_match_gcc_unroll): New function.
* gfortran.h (gfc_iterator]): Add unroll field.
(directive_unroll): Declare:
* match.c (gfc_match_do): Use memset to initialize the iterator.
* match.h (gfc_match_gcc_unroll): New prototype.
* parse.c (decode_gcc_attribute): Match "unroll".
(parse_do_block): Set iterator's unroll.
(parse_executable): Diagnose misplaced unroll directive.
* trans-stmt.c (gfc_trans_simple_do) Annotate loop condition with
annot_expr_unroll_kind.
(gfc_trans_do): Likewise.
* gfortran.texi (GNU Fortran Compiler Directives): Split section into
subections 'ATTRIBUTES directive' and 'UNROLL directive'.

From-SVN: r255973

compiler: improve escape analysis diagnostics

    This CL brings escape analysis diagnostics closer to the gc
    compiler's. This makes porting and debugging escape analysis
    code easier. A few changes:

    - In the gc compiler, the variable expression is represented
      with the variable node itself (ONAME), the location of which
      is the location of definition. We add a definition_location
      method to Node, and make use of it when the gc compiler emits
      diagnostics at the definition locations.

    - In the gc compiler, methods are named T.M or (*T).M. Add the
      type to the method name when possible.

    - Print "moved to heap" messages only for variables.

    - Reduce some duplicated diagnostics.

    - Print "does not escape" messages in more situations which the
      gc compiler does.

    - Remove the special handling for closure numbers. In gofrontend,
      closures are named "$nested#" where # is a global counter
      starting from 0, whereas in the gc compiler they are named
      "outer.func#" where # is a per-function counter starting from
      1. We tried to adjust the closure name to better matching the
      ones in the gc compiler, however, it cannot match exactly
      because of the difference of the counter. Instead, just print
      "outer.$nested#".

    Reviewed-on: https://go-review.googlesource.com/83875

From-SVN: r255967

[SFN] sync up debug-only stmt list's side effects with empty stmts too

for  gcc/c-family/ChangeLog

PR debug/83527
PR debug/83419
* c-semantics.c (only_debug_stmts_after_p): New.
(pop_stmt_list): Clear side effects in debug-only stmt list.
Check for single nondebug stmt followed by debug stmts only.

for  gcc/testsuite/ChangeLog

PR debug/83527
PR debug/83419
* gcc.dg/pr83527.c: New.

From-SVN: r255966

Daily bump.

From-SVN: r255965

PR testsuite/83462 - c-c++-common/Warray-bounds-3.c fails

gcc/testsuite/ChangeLog

* c-c++-common/Warray-bounds-3.c: Adjust dg-warning grep pattern.

From-SVN: r255962

re PR middle-end/83487 (ICE in expand_call, at calls.c:4098)

PR middle-end/83487
* config/i386/i386.c (ix86_function_arg_boundary): Return
PARM_BOUNDARY for TYPE_EMPTY_P types.

* gcc.c-torture/compile/pr83487.c: New test.
* gcc.dg/compat/pr83487-1.h: New file.
* gcc.dg/compat/pr83487-1_main.c: New test.
* gcc.dg/compat/pr83487-1_x.c: New file.
* gcc.dg/compat/pr83487-1_y.c: New file.
* gcc.dg/compat/pr83487-2_main.c: New test.
* gcc.dg/compat/pr83487-2_x.c: New file.
* gcc.dg/compat/pr83487-2_y.c: New file.
* g++.dg/abi/pr83487.C: New test.
* g++.dg/compat/abi/pr83487-1_main.C: New test.
* g++.dg/compat/abi/pr83487-1_x.C: New file.
* g++.dg/compat/abi/pr83487-1_y.C: New file.
* g++.dg/compat/abi/pr83487-2_main.C: New test.
* g++.dg/compat/abi/pr83487-2_x.C: New file.
* g++.dg/compat/abi/pr83487-2_y.C: New file.

From-SVN: r255961

re PR c/83448 (ice in get_source_location_for_substring, at input.c:1507)

PR c/83448
* gimple-ssa-sprintf.c (maybe_warn): Don't call set_caret_index
if navail is >= dir.len.

* gcc.c-torture/compile/pr83448.c: New test.
* gcc.dg/tree-ssa/builtin-snprintf-warn-4.c: New test.

From-SVN: r255960

Add missing PR number

From-SVN: r255959

decl.c (gnat_to_gnu_entity): Always take into account the Esize if it is known.

* gcc-interface/decl.c (gnat_to_gnu_entity) <E_Variable>: Always take
into account the Esize if it is known.

From-SVN: r255958

t-aarch64-linux (MULTILIB_OSDIRNAMES): Fix triplet for ilp32.

2017-12-21  Steve Ellcey  <sellcey@cavium.com>

* config/aarch64/t-aarch64-linux (MULTILIB_OSDIRNAMES): Fix
triplet for ilp32.

From-SVN: r255957

move ChangeLog entry to the right file

From-SVN: r255955

re PR rtl-optimization/80747 (gcc.dg/tree-ssa/tailrecursion-4.c fails with ICE when compiled with options "-fprofile-use -freorder-blocks-and-partition")

PR rtl-optimization/80747
PR rtl-optimization/83512
* cfgrtl.c (force_nonfallthru_and_redirect): When splitting
succ edge from ENTRY, copy partition from e->dest to the newly
created bb.
* bb-reorder.c (reorder_basic_blocks_simple): If last_tail is
ENTRY, use BB_PARTITION of its successor block as current_partition.
Don't copy partition when splitting succ edge from ENTRY.

* gcc.dg/pr80747.c: New test.
* gcc.dg/pr83512.c: New test.

From-SVN: r255954

re PR tree-optimization/83523 (ICE: verify_gimple failed (error: statement marked for throw, but doesn't))

PR tree-optimization/83523
* tree-ssa-math-opts.c (is_widening_mult_p): Return false if
for INTEGER_TYPE TYPE_OVERFLOW_TRAPS.
(convert_mult_to_fma): Likewise.

* g++.dg/tree-ssa/pr83523.C: New test.

From-SVN: r255953

re PR tree-optimization/83521 (ICE: verify_gimple failed (error: invalid operand in unary operation))

PR tree-optimization/83521
* tree-ssa-phiopt.c (factor_out_conditional_conversion): Use
gimple_build_assign without code on result of
fold_build1 (VIEW_CONVERT_EXPR, ...), as it might not create
a VIEW_CONVERT_EXPR.

* gcc.dg/pr83521.c: New test.

From-SVN: r255952

t-aarch64-linux (MULTILIB_OSDIRNAMES): Handle multi-arch for ilp32.

2017-12-21  Andrew Pinski  <apinski@cavium.com>
    Steve Ellcey  <sellcey@cavium.com>

* config/aarch64/t-aarch64-linux (MULTILIB_OSDIRNAMES): Handle
multi-arch for ilp32.

Co-Authored-By: Steve Ellcey <sellcey@cavium.com>
From-SVN: r255951

[PR c++/83406] deducing lambda type

https://gcc.gnu.org/ml/gcc-patches/2017-12/msg01432.html
PR c++/83406
* parser.c (cp_parser_lambda_body): Remove obsolete
single-return-statement handling.

PR c++/83406
* g++.dg/cpp0x/lambda/lambda-ice15.C: Adjust error.
* g++.dg/cpp1y/pr83406.C: New.

From-SVN: r255950

re PR target/83467 (ICE: in assign_by_spills, at lra-assigns.c:1476: unable to find a register to spill with -flive-range-shrinkage -m8bit-idiv)

PR target/83467
* config/i386/i386.md (*ashl<mode>3_mask): Add operand
constraints to operand 2.
(*ashl<mode>3_mask_1): Ditto.
(*<shift_insn><mode>3_mask): Ditto.
(*<shift_insn><mode>3_mask_1): Ditto.
(*<rotate_insn><mode>3_mask): Ditto.
(*<rotate_insn><mode>3_mask_1): Ditto.

testsuite/ChangeLog:

PR target/83467
* gcc.target/i386/pr83467-1.c: New test.
* gcc.target/i386/pr83467-2.c: Ditto.

From-SVN: r255949

[-fcompare-debug] retain insn locations when turning dbr seq into return

A number of -fcompare-debug errors on sparc arise as we split a dbr
SEQUENCE back into separate insns to turn the branch into a return.
If we just take the location from the PREV_INSN, it might be a debug
insn without INSN_LOCATION, or an insn with an unrelated location.
But that's silly: each of the SEQUENCEd insns is still an insn with
its own INSN_LOCATION, so use that instead, even though some may have
been adjusted while constructing the SEQUENCE.

for  gcc/ChangeLog

* reorg.c (make_return_insns): Reemit each insn with its own
location.

From-SVN: r255948

[SFN] propagate single-nondebug-stmt's side effects to enclosing list

Statements without side effects, preceded by debug begin stmt markers,
would become a statement list with side effects, although the stmt on
its own would be extracted from the list and remain not having side
effects.  This causes debug info and possibly codegen differences.
This patch fixes it, identifying the situation in which the stmt would
have been extracted from the stmt list, and propagating the side
effects flag from the stmt to the list.

for  gcc/ChangeLog

PR debug/83419
* c-family/c-semantics.c (pop_stmt_list): Propagate side
effects from single nondebug stmt to container list.

for  gcc/testsuite/ChangeLog

PR debug/83419
* gcc.dg/pr83419.c: New.

From-SVN: r255947

[patch AArch64] Do not perform a vector splat for vector initialisation if it is not useful

Our current vector initialisation code will first duplicate
the first element to both lanes, then overwrite the top lane with a new
value.

This duplication can be clunky and wasteful.

Better would be to simply use the fact that we will always be overwriting
the remaining bits, and simply move the first element to the corrcet place
(implicitly zeroing all other bits).

We also need a new pattern in simplify-rtx.c:simplify_ternary_operation ,
to ensure we can still simplify:

  (vec_merge:OUTER
     (vec_duplicate:OUTER x:INNER)
     (subreg:OUTER y:INNER 0)
     (const_int N))

To:

  (vec_concat:OUTER x:INNER y:INNER) or (vec_concat y x)

---
gcc/

* config/aarch64/aarch64.c (aarch64_expand_vector_init): Modify code
generation for cases where splatting a value is not useful.
* simplify-rtx.c (simplify_ternary_operation): Simplify vec_merge
across a vec_duplicate and a paradoxical subreg forming a vector
mode to a vec_concat.

gcc/testsuite/

* gcc.target/aarch64/vect-slp-dup.c: New.

From-SVN: r255946

[Patch combine] Don't create ZERO_EXTEND from subregs unless we have a scalar int mode

gcc/

* combine.c (simplify_set): Do not transform subregs to zero_extends
if the destination is not a scalar int mode.

From-SVN: r255945

re PR c++/82872 (ICE in ignore_overflows on __PTRDIFF_MAX__ index)

PR c++/82872
* convert.c (convert_to_integer_1) <POINTER_TYPE>: Do not return the
shared zero if the input has overflowed.

From-SVN: r255944

[arm] Specify +dotprod support for Cortex-A55 and Cortex-A75 in native system detection

Since support for -mcpu=cortex-a55 and -mcpu=cortex-a75
was added we added support for the +dotprod extension
which these CPUs support.

We already specify as such in the arm-cpus.in entries for
these processors. However the table in driver-arm.c was
not adding +dotproct to the -march string that it generates.

This patch fixes that oversight.
In the future I'd like to get the arm_cpu_table in driver-arm.c
be auto-generated somehow from the arm-cpus.in data so
that we don't have to keep track of discrepancies explicitly...

Bootstrapped and tested on arm-none-linux-gnueabihf.

* config/arm/driver-arm.c (arm_cpu_table): Specify dotprod
support for Cortex-A55 and Cortex-A75.

From-SVN: r255943

[arm] Fix assembler option rewrite alphabetical comparison

* common/config/arm/arm-common.c (compare_opt_names): Add function
comment.  Use strcmp instead of manual loop.

From-SVN: r255942

Fix gcov-dump tool for GCDA files (PR gcov-profile/83509).

2017-12-21  Martin Liska  <mliska@suse.cz>

PR gcov-profile/83509
* gcov-dump.c (dump_gcov_file): Do not read info about
support_unexecuted_blocks for gcda files.

From-SVN: r255941

re PR rtl-optimization/82973 (ICE in output_constant_pool_2, at varasm.c:3896 on aarch64)

PR rtl-optimization/82973
* emit-rtl.h (valid_for_const_vec_duplicate_p): Rename to ...
(valid_for_const_vector_p): ... this.
* emit-rtl.c (valid_for_const_vec_duplicate_p): Rename to ...
(valid_for_const_vector_p): ... this.  Adjust function comment.
(gen_vec_duplicate): Adjust caller.
* optabs.c (expand_vector_broadcast): Likewise.
* simplify-rtx.c (simplify_const_unary_operation): Don't optimize into
CONST_VECTOR if some element isn't simplified valid_for_const_vector_p
constant.
(simplify_const_binary_operation): Likewise.  Use CONST_FIXED_P macro
instead of GET_CODE == CONST_FIXED.
(simplify_subreg): Use CONST_FIXED_P macro instead of
GET_CODE == CONST_FIXED.

* gfortran.dg/pr82973.f90: New test.

From-SVN: r255939

re PR rtl-optimization/82973 (ICE in output_constant_pool_2, at varasm.c:3896 on aarch64)

PR rtl-optimization/82973
* emit-rtl.h (valid_for_const_vec_duplicate_p): Rename to ...
(valid_for_const_vector_p): ... this.
* emit-rtl.c (valid_for_const_vec_duplicate_p): Rename to ...
(valid_for_const_vector_p): ... this.  Adjust function comment.
(gen_vec_duplicate): Adjust caller.
* optabs.c (expand_vector_broadcast): Likewise.
* simplify-rtx.c (simplify_const_unary_operation): Don't optimize into
CONST_VECTOR if some element isn't simplified valid_for_const_vector_p
constant.
(simplify_const_binary_operation): Likewise.  Use CONST_FIXED_P macro
instead of GET_CODE == CONST_FIXED.
(simplify_subreg): Use CONST_FIXED_P macro instead of
GET_CODE == CONST_FIXED.

* gfortran.dg/pr82973.f90: New test.

From-SVN: r255938

re PR target/83488 (ICE on a CET test-case)

PR target/83488
* config/i386/i386.c (ix86_target_string): Move -mavx512vbmi2 and
-mshstk entries from isa_opts2 to isa_opts and -mhle, -mmovbe,
-mclzero and -mmwaitx entries from isa_opts to isa_opts2.
(ix86_option_override_internal): Adjust for
OPTION_MASK_ISA_{HLE,MOVBE,CLZERO,MWAITX} moving to ix86_isa_flags2
and OPTION_MASK_ISA_SHSTK moving to ix86_isa_flags.
(BDESC_VERIFYS): Remove SPECIAL_ARGS2 related checks.
(ix86_init_mmx_sse_builtins): Remove bdesc_special_args2 handling.
Use def_builtin2 instead of def_builtin for OPTION_MASK_ISA_MWAITX
and OPTION_MASK_ISA_CLZERO builtins.  Use def_builtin instead of
def_builtin2 for CET builtins.
(ix86_expand_builtin): Remove bdesc_special_args2 handling.  Fix
up formatting in IX86_BUILTIN_RDPID code.
* config/i386/i386-builtin.def: Move VBMI2 builtins from SPECIAL_ARGS2
section to SPECIAL_ARGS and from ARGS2 section to ARGS.
* config/i386/i386.opt (mavx512vbmi2, mshstk): Move from
ix86_isa_flags2 to ix86_isa_flags.
(mhle, mmovbe, mclzero, mmwaitx): Move from ix86_isa_flags to
ix86_isa_flags2.
* config/i386/i386-c.c (ix86_target_macros_internal): Check for
OPTION_MASK_ISA_{CLZERO,MWAITX} in isa_flag2 instead of isa_flag.
Check for OPTION_MASK_ISA_{SHSTK,AVX512VBMI2} in isa_flag instead
of isa_flag2.
* common/config/i386/i386-common.c (OPTION_MASK_ISA_AVX512VBMI2_SET):
Or in OPTION_MASK_ISA_AVX512F_SET.
(OPTION_MASK_ISA_AVX512F_UNSET): Or in
OPTION_MASK_ISA_AVX512VBMI2_UNSET.
(ix86_handle_option): Adjust for
OPTION_MASK_ISA_{SHSTK,AVX512VBMI2}_*SET being in ix86_isa_flags
and OPTION_MASK_ISA_{MOVBE,MWAITX,CLZERO}_*SET in ix86_isa_flags2.

* gcc.target/i386/pr83488.c: New test.

From-SVN: r255937

poly_int: prune_runtime_alias_test_list

This patch makes prune_runtime_alias_test_list take the iteration
factor as a poly_int and tracks polynomial offsets internally
as well.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-data-ref.h (prune_runtime_alias_test_list): Take the
factor as a poly_uint64 rather than an unsigned HOST_WIDE_INT.
* tree-data-ref.c (prune_runtime_alias_test_list): Likewise.
Track polynomial offsets.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255936

poly_int: compute_data_ref_alignment

This patch makes vect_compute_data_ref_alignment treat DR_INIT as a
poly_int and handles cases in which the calculated misalignment might
not be constant.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-vect-data-refs.c (vect_compute_data_ref_alignment):
Treat drb->init as a poly_int.  Fail if its misalignment wrt
vector_alignment isn't known.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255935

poly_int: loop versioning threshold

This patch splits the loop versioning threshold out from the
cost model threshold so that the former can become a poly_uint64.
We still use a single test to enforce both limits where possible.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-vectorizer.h (_loop_vec_info): Add a versioning_threshold
field.
(LOOP_VINFO_VERSIONING_THRESHOLD): New macro
(vect_loop_versioning): Take the loop versioning threshold as a
separate parameter.
* tree-vect-loop-manip.c (vect_loop_versioning): Likewise.
* tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Initialize
versioning_threshold.
(vect_analyze_loop_2): Compute the loop versioning threshold
whenever loop versioning is needed, and store it in the new
field rather than combining it with the cost model threshold.
(vect_transform_loop): Update call to vect_loop_versioning.
Try to combine the loop versioning and cost thresholds here.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255934

poly_int: tree-ssa-loop-ivopts.c:iv_use

This patch makes ivopts handle polynomial address offsets
when recording potential IV uses.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssa-loop-ivopts.h (strip_offset): Return the offset as
poly_uint64_pod rather than an unsigned HOST_WIDE_INT.
* tree-loop-distribution.c (classify_builtin_st): Update accordingly.
* tree-ssa-loop-ivopts.c (iv_use::addr_offset): Change from
an unsigned HOST_WIDE_INT to a poly_uint64_pod.
(group_compare_offset): Update accordingly.
(split_small_address_groups_p): Likewise.
(record_use): Take addr_offset as a poly_uint64 rather than
an unsigned HOST_WIDE_INT.
(strip_offset): Return the offset as a poly_uint64 rather than
an unsigned HOST_WIDE_INT.
(record_group_use, split_address_groups): Track polynomial offsets.
(add_iv_candidate_for_use): Likewise.
(addr_offset_valid_p): Take the offset as a poly_int64 rather
than a HOST_WIDE_INT.
(strip_offset_1): Return the offset as a poly_int64 rather than
a HOST_WIDE_INT.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255933

poly_int: get_binfo_at_offset

This patch changes the offset parameter to get_binfo_at_offset
from HOST_WIDE_INT to poly_int64.  This function probably doesn't
need to handle polynomial offsets in practice, but it's easy
to do and avoids forcing the caller to check first.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree.h (get_binfo_at_offset): Take the offset as a poly_int64
rather than a HOST_WIDE_INT.
* tree.c (get_binfo_at_offset): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255932

poly_int: build_ref_for_offset

This patch changes the offset parameter to build_ref_for_offset
from HOST_WIDE_INT to poly_int64.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* ipa-prop.h (build_ref_for_offset): Take the offset as a poly_int64
rather than a HOST_WIDE_INT.
* tree-sra.c (build_ref_for_offset): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255931

poly_int: MEM_REF offsets

This patch allows MEM_REF offsets to be polynomial, with mem_ref_offset
now returning a poly_offset_int instead of an offset_int.  The
non-mechanical changes to callers of mem_ref_offset were handled by
previous patches.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* fold-const.h (mem_ref_offset): Return a poly_offset_int rather
than an offset_int.
* tree.c (mem_ref_offset): Likewise.
(build_simple_mem_ref_loc): Treat MEM_REF offsets as poly_ints.
* builtins.c (get_object_alignment_2): Likewise.
* expr.c (get_inner_reference, expand_expr_real_1): Likewise.
* gimple-fold.c (get_base_constructor): Likewise.
* gimple-ssa-strength-reduction.c (restructure_reference): Likewise.
* gimple-ssa-warn-restrict.c (builtin_memref::builtin_memref):
Likewise.
* ipa-polymorphic-call.c
(ipa_polymorphic_call_context::ipa_polymorphic_call_context): Likewise.
* ipa-prop.c (compute_complex_assign_jump_func): Likewise.
(get_ancestor_addr_info): Likewise.
* ipa-param-manipulation.c (ipa_get_adjustment_candidate): Likewise.
* match.pd: Likewise.
* tree-data-ref.c (dr_analyze_innermost): Likewise.
* tree-dfa.c (get_addr_base_and_unit_offset_1): Likewise.
* tree-eh.c (tree_could_trap_p): Likewise.
* tree-object-size.c (addr_object_size): Likewise.
* tree-ssa-address.c (copy_ref_info): Likewise.
* tree-ssa-alias.c (indirect_ref_may_alias_decl_p): Likewise.
(indirect_refs_may_alias_p): Likewise.
* tree-ssa-sccvn.c (copy_reference_ops_from_ref): Likewise.
* tree-ssa.c (maybe_rewrite_mem_ref_base): Likewise.
(non_rewritable_mem_ref_base): Likewise.
* tree-vect-data-refs.c (vect_check_gather_scatter): Likewise.
* tree-vrp.c (vrp_prop::check_array_ref): Likewise.
* varasm.c (decode_addr_const): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255930

poly_int: find_bswap_or_nop_load

This patch handles polynomial offsets in find_bswap_or_nop_load,
which could be useful for constant-sized data at a variable offset.
It is needed for a later patch to compile.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* gimple-ssa-stor-merging.c (find_bswap_or_nop_load): Track polynomial
offsets for MEM_REFs.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255929

poly_int: adjust_ptr_info_misalignment

This patch makes adjust_ptr_info_misalignment take the adjustment
as a poly_uint64 rather than an unsigned int.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssanames.h (adjust_ptr_info_misalignment): Take the increment
as a poly_uint64 rather than an unsigned int.
* tree-ssanames.c (adjust_ptr_info_misalignment): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255928

poly_int: decode_addr_const

This patch makes the varasm-local addr_const track polynomial offsets.
I'm not sure how useful this is, but it was easier to convert than not.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* varasm.c (addr_const::offset): Change from HOST_WIDE_INT
to poly_int64.
(decode_addr_const): Update accordingly.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255927

poly_int: bit_field_size/offset

verify_expr ensured that the size and offset in gimple BIT_FIELD_REFs
satisfied tree_fits_uhwi_p.  This patch extends that so that they can
be poly_uint64s, and adds helper routines for accessing them when the
verify_expr requirements apply.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree.h (bit_field_size, bit_field_offset): New functions.
* hsa-gen.c (gen_hsa_addr): Use them.
* tree-ssa-forwprop.c (simplify_bitfield_ref): Likewise.
(simplify_vector_constructor): Likewise.
* tree-ssa-sccvn.c (copy_reference_ops_from_ref): Likewise.
* tree-cfg.c (verify_expr): Require the sizes and offsets of a
BIT_FIELD_REF to be poly_uint64s rather than uhwis.
* fold-const.c (fold_ternary_loc): Protect tree_to_uhwi with
tree_fits_uhwi_p.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255926

poly_int: emit_group_load/store

This patch changes the sizes passed to emit_group_load and
emit_group_store from int to poly_int64.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expr.h (emit_group_load, emit_group_load_into_temps)
(emit_group_store): Take the size as a poly_int64 rather than an int.
* expr.c (emit_group_load_1, emit_group_load): Likewise.
(emit_group_load_into_temp, emit_group_store): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255925

poly_int: reload<->ira interface

This patch uses poly_int64 for:

- ira_reuse_stack_slot
- ira_mark_new_stack_slot
- ira_spilled_reg_stack_slot::width

all of which are part of the IRA/reload interface.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* ira-int.h (ira_spilled_reg_stack_slot::width): Change from
an unsigned int to a poly_uint64.
* ira.h (ira_reuse_stack_slot, ira_mark_new_stack_slot): Take the
sizes as poly_uint64s rather than unsigned ints.
* ira-color.c (ira_reuse_stack_slot, ira_mark_new_stack_slot):
Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255924

poly_int: emit_inc

This patch changes the LRA emit_inc routine so that it takes
a poly_int64 rather than an int.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* lra-constraints.c (emit_inc): Change inc_amount from an int
to a poly_int64.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255923

poly_int: cfgexpand stack variables

This patch changes the type of stack_var::size from HOST_WIDE_INT
to poly_uint64.  The difference in signedness is because the
field was set by:

  v->size = tree_to_uhwi (size);

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* cfgexpand.c (stack_var::size): Change from a HOST_WIDE_INT
to a poly_uint64.
(add_stack_var, stack_var_cmp, partition_stack_vars)
(dump_stack_var_partition): Update accordingly.
(alloc_stack_frame_space): Take the size as a poly_int64 rather
than a HOST_WIDE_INT.
(expand_stack_vars, expand_one_stack_var_1): Handle polynomial sizes.
(defer_stack_allocation, estimated_stack_frame_size): Likewise.
(account_stack_vars, expand_one_var): Likewise.  Return a poly_uint64
rather than a HOST_WIDE_INT.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255922

poly_int: argument sizes

This patch changes various bits of state related to argument sizes so
that they have type poly_int64 rather than HOST_WIDE_INT.  This includes:

- incoming_args::pops_args and incoming_args::size
- rtl_data::outgoing_args_size
- pending_stack_adjust
- stack_pointer_delta
- stack_usage::pushed_stack_size
- args_size::constant

It also changes TARGET_RETURN_POPS_ARGS so that the size of the
arguments passed in and the size returned by the hook are both
poly_int64s.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* target.def (return_pops_args): Treat both the input and output
sizes as poly_int64s rather than HOST_WIDE_INTS.
* targhooks.h (default_return_pops_args): Update accordingly.
* targhooks.c (default_return_pops_args): Likewise.
* doc/tm.texi: Regenerate.
* emit-rtl.h (incoming_args): Change pops_args, size and
outgoing_args_size from int to poly_int64_pod.
* function.h (expr_status): Change x_pending_stack_adjust and
x_stack_pointer_delta from int to poly_int64.
(args_size::constant): Change from HOST_WIDE_INT to poly_int64.
(ARGS_SIZE_RTX): Update accordingly.
* calls.c (highest_outgoing_arg_in_use): Change from int to
unsigned int.
(stack_usage_watermark, stored_args_watermark): New variables.
(stack_region_maybe_used_p, mark_stack_region_used): New functions.
(emit_call_1): Change the stack_size and rounded_stack_size
parameters from HOST_WIDE_INT to poly_int64.  Track n_popped
as a poly_int64.
(save_fixed_argument_area): Check stack_usage_watermark.
(initialize_argument_information): Change old_pending_adj from
a HOST_WIDE_INT * to a poly_int64_pod *.
(compute_argument_block_size): Return the size as a poly_int64
rather than an int.
(finalize_must_preallocate): Track polynomial argument sizes.
(compute_argument_addresses): Likewise.
(internal_arg_pointer_based_exp): Track polynomial offsets.
(mem_overlaps_already_clobbered_arg_p): Rename to...
(mem_might_overlap_already_clobbered_arg_p): ...this and take the
size as a poly_uint64 rather than an unsigned HOST_WIDE_INT.
Check stored_args_used_watermark.
(load_register_parameters): Update accordingly.
(check_sibcall_argument_overlap_1): Likewise.
(combine_pending_stack_adjustment_and_call): Take the unadjusted
args size as a poly_int64 rather than an int.  Return a bool
indicating whether the optimization was possible and return
the new adjustment by reference.
(check_sibcall_argument_overlap): Track polynomail argument sizes.
Update stored_args_watermark.
(can_implement_as_sibling_call_p): Handle polynomial argument sizes.
(expand_call): Likewise.  Maintain stack_usage_watermark and
stored_args_watermark.  Update calls to
combine_pending_stack_adjustment_and_call.
(emit_library_call_value_1): Handle polynomial argument sizes.
Call stack_region_maybe_used_p and mark_stack_region_used.
Maintain stack_usage_watermark.
(store_one_arg): Likewise.  Update call to
mem_overlaps_already_clobbered_arg_p.
* config/arm/arm.c (arm_output_function_prologue): Add a cast to
HOST_WIDE_INT.
* config/avr/avr.c (avr_outgoing_args_size): Likewise.
* config/microblaze/microblaze.c (microblaze_function_prologue):
Likewise.
* config/cr16/cr16.c (cr16_return_pops_args): Update for new
TARGET_RETURN_POPS_ARGS interface.
(cr16_compute_frame, cr16_initial_elimination_offset): Add casts
to HOST_WIDE_INT.
* config/ft32/ft32.c (ft32_compute_frame): Likewise.
* config/i386/i386.c (ix86_return_pops_args): Update for new
TARGET_RETURN_POPS_ARGS interface.
(ix86_expand_split_stack_prologue): Add a cast to HOST_WIDE_INT.
* config/moxie/moxie.c (moxie_compute_frame): Likewise.
* config/m68k/m68k.c (m68k_return_pops_args): Update for new
TARGET_RETURN_POPS_ARGS interface.
* config/vax/vax.c (vax_return_pops_args): Likewise.
* config/pa/pa.h (STACK_POINTER_OFFSET): Add a cast to poly_int64.
(EXIT_IGNORE_STACK): Update reference to crtl->outgoing_args_size.
* config/arm/arm.h (CALLER_INTERWORKING_SLOT_SIZE): Likewise.
* config/powerpcspe/aix.h (STACK_DYNAMIC_OFFSET): Likewise.
* config/powerpcspe/darwin.h (STACK_DYNAMIC_OFFSET): Likewise.
* config/powerpcspe/powerpcspe.h (STACK_DYNAMIC_OFFSET): Likewise.
* config/rs6000/aix.h (STACK_DYNAMIC_OFFSET): Likewise.
* config/rs6000/darwin.h (STACK_DYNAMIC_OFFSET): Likewise.
* config/rs6000/rs6000.h (STACK_DYNAMIC_OFFSET): Likewise.
* dojump.h (saved_pending_stack_adjust): Change x_pending_stack_adjust
and x_stack_pointer_delta from int to poly_int64.
* dojump.c (do_pending_stack_adjust): Update accordingly.
* explow.c (allocate_dynamic_stack_space): Handle polynomial
stack_pointer_deltas.
* function.c (STACK_DYNAMIC_OFFSET): Add a cast to poly_int64.
(pad_to_arg_alignment): Track polynomial offsets.
(assign_parm_find_stack_rtl): Likewise.
(assign_parms, locate_and_pad_parm): Handle polynomial argument sizes.
* toplev.c (output_stack_usage): Update reference to
current_function_pushed_stack_size.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255921

poly_int: instantiate_virtual_regs

This patch makes the instantiate virtual regs pass track offsets
as poly_ints.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* function.c (in_arg_offset, var_offset, dynamic_offset)
(out_arg_offset, cfa_offset): Change from int to poly_int64.
(instantiate_new_reg): Return the new offset as a poly_int64_pod
rather than a HOST_WIDE_INT.
(instantiate_virtual_regs_in_rtx): Track polynomial offsets.
(instantiate_virtual_regs_in_insn): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255920

poly_int: REG_ARGS_SIZE

This patch adds new utility functions for manipulating REG_ARGS_SIZE
notes and allows the notes to carry polynomial as well as constant sizes.

The code was inconsistent about whether INT_MIN or HOST_WIDE_INT_MIN
should be used to represent an unknown size.  The patch uses
HOST_WIDE_INT_MIN throughout.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtl.h (get_args_size, add_args_size_note): New functions.
(find_args_size_adjust): Return a poly_int64 rather than a
HOST_WIDE_INT.
(fixup_args_size_notes): Likewise.  Make the same change to the
end_args_size parameter.
* rtlanal.c (get_args_size, add_args_size_note): New functions.
* builtins.c (expand_builtin_trap): Use add_args_size_note.
* calls.c (emit_call_1): Likewise.
* explow.c (adjust_stack_1): Likewise.
* cfgcleanup.c (old_insns_match_p): Update use of
find_args_size_adjust.
* combine.c (distribute_notes): Track polynomial arg sizes.
* dwarf2cfi.c (dw_trace_info): Change beg_true_args_size,
end_true_args_size, beg_delay_args_size and end_delay_args_size
from HOST_WIDE_INT to poly_int64.
(add_cfi_args_size): Take the args_size as a poly_int64 rather
than a HOST_WIDE_INT.
(notice_args_size, notice_eh_throw, maybe_record_trace_start)
(maybe_record_trace_start_abnormal, scan_trace, connect_traces): Track
polynomial arg sizes.
* emit-rtl.c (try_split): Use get_args_size.
* recog.c (peep2_attempt): Likewise.
* reload1.c (reload_as_needed): Likewise.
* expr.c (find_args_size_adjust): Return the adjustment as a
poly_int64 rather than a HOST_WIDE_INT.
(fixup_args_size_notes): Change end_args_size from a HOST_WIDE_INT
to a poly_int64 and change the return type in the same way.
(emit_single_push_insn): Track polynomial arg sizes.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255919

poly_int: push_block/emit_push_insn

This patch changes the "extra" parameters to push_block and
emit_push_insn from int to poly_int64.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expr.h (push_block, emit_push_insn): Change the "extra" parameter
from HOST_WIDE_INT to poly_int64.
* expr.c (push_block, emit_push_insn): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255918

poly_int: frame allocations

This patch converts the frame allocation code (mostly in function.c)
to use poly_int64 rather than HOST_WIDE_INT for frame offsets and
sizes.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* function.h (frame_space): Change start and length from HOST_WIDE_INT
to poly_int64.
(get_frame_size): Return the size as a poly_int64 rather than a
HOST_WIDE_INT.
(frame_offset_overflow): Take the offset as a poly_int64 rather
than a HOST_WIDE_INT.
(assign_stack_local_1, assign_stack_local, assign_stack_temp_for_type)
(assign_stack_temp): Likewise for the size.
* function.c (get_frame_size): Return a poly_int64 rather than
a HOST_WIDE_INT.
(frame_offset_overflow): Take the offset as a poly_int64 rather
than a HOST_WIDE_INT.
(try_fit_stack_local): Take the start, length and size as poly_int64s
rather than HOST_WIDE_INTs.  Return the offset as a poly_int64_pod
rather than a HOST_WIDE_INT.
(add_frame_space): Take the start and end as poly_int64s rather than
HOST_WIDE_INTs.
(assign_stack_local_1, assign_stack_local, assign_stack_temp_for_type)
(assign_stack_temp): Likewise for the size.
(temp_slot): Change size, base_offset and full_size from HOST_WIDE_INT
to poly_int64.
(find_temp_slot_from_address): Handle polynomial offsets.
(combine_temp_slots): Likewise.
* emit-rtl.h (rtl_data::x_frame_offset): Change from HOST_WIDE_INT
to poly_int64.
* cfgexpand.c (alloc_stack_frame_space): Return the offset as a
poly_int64 rather than a HOST_WIDE_INT.
(expand_one_stack_var_at): Take the offset as a poly_int64 rather
than a HOST_WIDE_INT.
(expand_stack_vars, expand_one_stack_var_1, expand_used_vars): Handle
polynomial frame offsets.
* config/m32r/m32r-protos.h (m32r_compute_frame_size): Take the size
as a poly_int64 rather than an int.
* config/m32r/m32r.c (m32r_compute_frame_size): Likewise.
* config/v850/v850-protos.h (compute_frame_size): Likewise.
* config/v850/v850.c (compute_frame_size): Likewise.
* config/xtensa/xtensa-protos.h (compute_frame_size): Likewise.
* config/xtensa/xtensa.c (compute_frame_size): Likewise.
* config/pa/pa-protos.h (pa_compute_frame_size): Likewise.
* config/pa/pa.c (pa_compute_frame_size): Likewise.
* explow.h (get_dynamic_stack_base): Take the offset as a poly_int64
rather than a HOST_WIDE_INT.
* explow.c (get_dynamic_stack_base): Likewise.
* final.c (final_start_function): Use the constant lower bound
of the frame size for -Wframe-larger-than.
* ira.c (do_reload): Adjust for new get_frame_size return type.
* lra.c (lra): Likewise.
* reload1.c (reload): Likewise.
* config/avr/avr.c (avr_asm_function_end_prologue): Likewise.
* config/pa/pa.h (EXIT_IGNORE_STACK): Likewise.
* rtlanal.c (get_initial_register_offset): Return the offset as
a poly_int64 rather than a HOST_WIDE_INT.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255917

poly_int: reload1.c

This patch makes a few small poly_int64 changes to reload1.c,
mostly related to eliminations.  Again, there's no real expectation
that reload will be used for targets that have polynomial-sized modes,
but it seemed easier to convert it anyway.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* reload1.c (elim_table): Change initial_offset, offset and
previous_offset from HOST_WIDE_INT to poly_int64_pod.
(offsets_at): Change the target array's element type from
HOST_WIDE_INT to poly_int64_pod.
(set_label_offsets, eliminate_regs_1, eliminate_regs_in_insn)
(elimination_costs_in_insn, update_eliminable_offsets)
(verify_initial_elim_offsets, set_offsets_for_label)
(init_eliminable_invariants): Update after above changes.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255916

poly_int: reload.c

This patch makes a few small poly_int64 changes to reload.c,
such as in the "decomposition" structure.  In practice, any
port with polynomial-sized modes should be using LRA rather
than reload, but it's easier to convert reload anyway than
to sprinkle to_constants everywhere.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* reload.h (reload::inc): Change from an int to a poly_int64_pod.
* reload.c (combine_reloads, debug_reload_to_stream): Likewise.
(decomposition): Change start and end from HOST_WIDE_INT
to poly_int64_pod.
(decompose, immune_p): Update accordingly.
(find_inc_amount): Return a poly_int64 rather than an int.
* reload1.c (inc_for_reload): Take the inc_amount as a poly_int64
rather than an int.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255915

poly_int: get_inner_reference & co.

This patch makes get_inner_reference and ptr_difference_const return the
bit size and bit position as poly_int64s rather than HOST_WIDE_INTS.
The non-mechanical changes were handled by previous patches.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree.h (get_inner_reference): Return the bitsize and bitpos
as poly_int64_pods rather than HOST_WIDE_INT.
* fold-const.h (ptr_difference_const): Return the pointer difference
as a poly_int64_pod rather than a HOST_WIDE_INT.
* expr.c (get_inner_reference): Return the bitsize and bitpos
as poly_int64_pods rather than HOST_WIDE_INT.
(expand_expr_addr_expr_1, expand_expr_real_1): Track polynomial
offsets and sizes.
* fold-const.c (make_bit_field_ref): Take the bitpos as a poly_int64
rather than a HOST_WIDE_INT.  Update call to get_inner_reference.
(optimize_bit_field_compare): Update call to get_inner_reference.
(decode_field_reference): Likewise.
(fold_unary_loc): Track polynomial offsets and sizes.
(split_address_to_core_and_offset): Return the bitpos as a
poly_int64_pod rather than a HOST_WIDE_INT.
(ptr_difference_const): Likewise for the pointer difference.
* asan.c (instrument_derefs): Track polynomial offsets and sizes.
* config/mips/mips.c (r10k_safe_mem_expr_p): Likewise.
* dbxout.c (dbxout_expand_expr): Likewise.
* dwarf2out.c (loc_list_for_address_of_addr_expr_of_indirect_ref)
(loc_list_from_tree_1, fortran_common): Likewise.
* gimple-laddress.c (pass_laddress::execute): Likewise.
* gimple-ssa-store-merging.c (find_bswap_or_nop_load): Likewise.
* gimplify.c (gimplify_scan_omp_clauses): Likewise.
* simplify-rtx.c (delegitimize_mem_from_attrs): Likewise.
* tree-affine.c (tree_to_aff_combination): Likewise.
(get_inner_reference_aff): Likewise.
* tree-data-ref.c (split_constant_offset_1): Likewise.
(dr_analyze_innermost): Likewise.
* tree-scalar-evolution.c (interpret_rhs_expr): Likewise.
* tree-sra.c (ipa_sra_check_caller): Likewise.
* tree-vect-data-refs.c (vect_check_gather_scatter): Likewise.
* ubsan.c (maybe_instrument_pointer_overflow): Likewise.
(instrument_bool_enum_load, instrument_object_size): Likewise.
* gimple-ssa-strength-reduction.c (slsr_process_ref): Update call
to get_inner_reference.
* hsa-gen.c (gen_hsa_addr): Likewise.
* sanopt.c (maybe_optimize_ubsan_ptr_ifn): Likewise.
* tsan.c (instrument_expr): Likewise.
* match.pd: Update call to ptr_difference_const.

gcc/ada/
* gcc-interface/trans.c (Attribute_to_gnu): Track polynomial
offsets and sizes.
* gcc-interface/utils2.c (build_unary_op): Likewise.

gcc/cp/
* constexpr.c (check_automatic_or_tls): Track polynomial
offsets and sizes.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255914

poly_int: fold_comparison

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* fold-const.c (fold_comparison): Track sizes and offsets as
poly_int64s rather than HOST_WIDE_INTs when folding address
comparisons.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255913

poly_int: get_bit_range

This patch makes get_bit_range return the range and position as poly_ints.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expr.h (get_bit_range): Return the bitstart and bitend as
poly_uint64s rather than unsigned HOST_WIDE_INTs.  Return the bitpos
as a poly_int64 rather than a HOST_WIDE_INT.
* expr.c (get_bit_range): Likewise.
(expand_assignment): Update call accordingly.
* fold-const.c (optimize_bit_field_compare): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255912

[AArch64] Tweak aarch64_classify_address interface

Previously aarch64_classify_address used an rtx code to distinguish
LDP/STP addresses from normal addresses; the code was PARALLEL
to select LDP/STP and anything else to select normal addresses.
This patch replaces that parameter with a dedicated enum.

The SVE port will add another enum value that didn't map naturally
to an rtx code.

2017-12-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* config/aarch64/aarch64-protos.h (aarch64_addr_query_type): New enum.
(aarch64_legitimate_address_p): Use it instead of an rtx code,
as an optional final parameter.
* config/aarch64/aarch64.c (aarch64_classify_address): Likewise.
(aarch64_legitimate_address_p): Likewise.
(aarch64_print_address_internal): Take an aarch64_addr_query_type
instead of an rtx code.
(aarch64_address_valid_for_prefetch_p): Update calls accordingly.
(aarch64_legitimate_address_hook_p): Likewise.
(aarch64_print_ldpstp_address): Likewise.
(aarch64_print_operand_address): Likewise.
(aarch64_address_cost): Likewise.
* config/aarch64/constraints.md (Uml, Umq, Ump, Utq): Likewise.
* config/aarch64/predicates.md (aarch64_mem_pair_operand): Likewise.
(aarch64_mem_pair_lanes_operand): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255911

tree-ssa-dom.c (dom_opt_dom_walker::optimize_stmt): Call update_stmt_if_modified.

* tree-ssa-dom.c (dom_opt_dom_walker::optimize_stmt): Call
update_stmt_if_modified.

From-SVN: r255910

Daily bump.

From-SVN: r255909

re PR tree-optimization/83491 (ICE in execute_cse_reciprocals_1 at gcc/tree-ssa-math-opts.c:585)

PR tree-optimization/83491
* tree-ssa-math-opts.c (execute_cse_reciprocals_1): Check for SSA_NAME
before walking uses.  Improve coding style and comments.

PR tree-optimization/83491
* gcc.dg/pr83491.c: Add new test.

From-SVN: r255906

Simplify fold_internal_goacc_dim

2017-12-20  Tom de Vries  <tom@codesourcery.com>

* gimple-fold.c (fold_internal_goacc_dim): Simplify.

From-SVN: r255905

re PR libstdc++/82522 (std::map::insert(value_type &&) not selected)

2017-12-20  François Dumont  <fdumont@gcc.gnu.org>

PR libstdc++/82522
* include/debug/map.h (map::insert(value_type&&))
(map::insert(const_iterator, value_type&&)): Add overload for rvalues.
* include/debug/multimap.h (multimap::insert(value_type&&))
(multimap::insert(const_iterator, value_type&&)): Likewise.
* include/debug/unordered_map (unordered_map::insert(value_type&&))
(unordered_map::insert(const_iterator, value_type&&))
(unordered_multimap::insert(value_type&&))
(unordered_multimap::insert(const_iterator, value_type&&)): Likewise.
* testsuite/23_containers/map/modifiers/insert/dr2354.cc (test02): New.
* testsuite/23_containers/multimap/modifiers/insert/dr2354.cc (test02):
New.
* testsuite/23_containers/unordered_map/insert/dr2354.cc (test02): New.
* testsuite/23_containers/unordered_multimap/insert/dr2354.cc (test02):
New.

From-SVN: r255904

re PR fortran/83436 (Internal file cannot be accessed by UNFORMATTED data transfer when reading from /dev/urandom)

2017-12-20  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/83436
* gfortran.dg/internal_readwrite_4.f90: New test.

From-SVN: r255902

re PR ipa/83506 (ICE: Segmentation fault in force_nonfallthru_and_redirect)

PR ipa/83506
* ipa-fnsummary.c (pass_data_ipa_free_fn_summary): Use 0 for
todo_flags_finish.
(pass_ipa_free_fn_summary): Add small_p private data member,
initialize to false in the ctor.
(pass_ipa_free_fn_summary::clone,
pass_ipa_free_fn_summary::set_pass_param,
pass_ipa_free_fn_summary::gate): New methods.
(pass_ipa_free_fn_summary::execute): Return TODO_remove_functions
| TODO_dump_symtab if small_p.
* passes.def: Add true parm for the existing pass_ipa_free_fn_summary
entry and add another instance of the pass with false parm after
ipa-pure-const.
* ipa-pure-const.c (pass_ipa_pure_const): Don't call
ipa_free_fn_summary here.

* gcc.dg/pr83506.c: New test.
* gcc.dg/ipa/ctor-empty-1.c: Use -fdump-ipa-free-fnsummary1 instead
of -fdump-ipa-free-fnsummary and scan in free-fnsummary1 instead of
free-fnsummary dump.

From-SVN: r255901

gimplify.c (gimplify_return_expr): Remove dead error_mark_node check.

2017-12-20  Paolo Carlini  <paolo.carlini@oracle.com>

* gimplify.c (gimplify_return_expr): Remove dead error_mark_node check.

From-SVN: r255900

PR testsuite/83483 - [8 Regression] gcc.dg/memcpy-6.c FAILs

gcc/testsuite/ChangeLog:

PR testsuite/83483
* gcc.dg/memcpy-6.c: Skip for targets other than x86.

From-SVN: r255899

PR testsuite/83131 - c-c++/common/attr-nonstring-3 failure for strcmp tests on PowerPC

gcc/ChangeLog:

PR testsuite/83131
* builtins.c (expand_builtin_strlen): Use get_callee_fndecl.
(expand_builtin_strcmp): Call maybe_warn_nonstring_arg.
(expand_builtin_strncmp): Same.

gcc/testsuite/ChangeLog:

PR testsuite/83131
* c-c++-common/attr-nonstring-4.c: New test.

From-SVN: r255898

[PATCH, PR83492] Fix selection of aarch64 big-endian shift parameters based on __AARCH64EB__

2017-12-20  Michael Weiser  <michael.weiser@gmx.de>

PR preprocessor/83492
* lex.c (search_line_fast) [__ARM_NEON && __ARM_64BIT_STATE]:
Fix selection of big-endian shift parameters by using
__ARM_BIG_ENDIAN.

From-SVN: r255896

[SFN] debug markers before labels no more

Make sure that gimple and RTL IRs don't have debug markers before
labels.  When we build the CFG, we move labels before any markers
appearing before them.  Then, make sure we don't mistakenly
reintroduce them.

This reverts some of the complexity that had been brought about by the
initial SFN patches.

for  gcc/ChangeLog

PR bootstrap/83396
* cfgexpand.c (label_rtx_for_bb): Revert SFN changes that
allowed debug stmts before labels.
(expand_gimple_basic_block): Likewise.
* gimple-iterator.c (gimple_find_edge_insert_loc): Likewise.
* gimple-iterator.h (gsi_after_labels): Likewise.
* tree-cfgcleanup (remove_forwarder_block): Likewise, but
rename reused variable, and simplify using gsi_move_before.
* tree-ssa-tail-merge.c (find_duplicate): Likewise.
* tree-cfg.c (make_edges, cleanup_dead_labels): Likewise.
(gimple_can_merge_blocks_p, verify_gimple_in_cfg): Likewise.
(gimple_verify_flow_info, gimple_block_label): Likewise.
(make_blocks): Move debug markers after adjacent labels.
* cfgrtl.c (skip_insns_after_block): Revert SFN changes that
allowed debug insns outside blocks.
* df-scan.c (df_insn_delete): Likewise.
* lra-constraints.c (update_ebb_live_info): Likewise.
* var-tracking.c (get_first_insn, vt_emit_notes): Likewise.
(vt_initialize, delete_vta_debug_insns): Likewise.
(reemit_marker_as_note): Drop BB parm.  Adjust callers.

From-SVN: r255895

poly_int: store merging

This patch makes pass_store_merging track polynomial sizes
and offsets.  store_immediate_info remains restricted to stores
with a constant offset and size.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* poly-int-types.h (round_down_to_byte_boundary): New macro.
(round_up_to_byte_boundary): Likewise.
* expr.h (get_bit_range): Add temporary shim.
* gimple-ssa-store-merging.c (store_operand_info): Change the
bitsize, bitpos, bitregion_start and bitregion_end fields from
unsigned HOST_WIDE_INT to poly_uint64.
(merged_store_group): Likewise load_align_base.
(compatible_load_p, compatible_load_p): Update accordingly.
(imm_store_chain_info::coalesce_immediate_stores): Likewise.
(split_group, imm_store_chain_info::output_merged_store): Likewise.
(mem_valid_for_store_merging): Return the bitsize, bitpos,
bitregion_start and bitregion_end as poly_uint64s rather than
unsigned HOST_WIDE_INTs.  Track polynomial offsets internally.
(handled_load): Take the bitsize, bitpos,
bitregion_start and bitregion_end as poly_uint64s rather than
unsigned HOST_WIDE_INTs.
(pass_store_merging::process_store): Update call to
mem_valid_for_store_merging.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255894

poly_int: get_object_alignment_2

This patch makes get_object_alignment_2 track polynomial offsets
and sizes.  The real work is done by get_inner_reference, but we
then need to handle the alignment correctly.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* builtins.c (get_object_alignment_2): Track polynomial offsets
and sizes.  Update the alignment handling.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255893

poly_int: expand_debug_expr

This patch makes expand_debug_expr track polynomial memory offsets.
It simplifies the handling of the case in which the reference is not
to the first byte of the base, which seemed non-trivial enough to
make it worth splitting out as a separate patch.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree.h (get_inner_reference): Add a version that returns the
offset and size as poly_int64_pods rather than HOST_WIDE_INTs.
* cfgexpand.c (expand_debug_expr): Track polynomial offsets.  Simply
the case in which bitpos is not associated with the first byte.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255892

poly_int: get_inner_reference_aff

This patch makes get_inner_reference_aff return the size as a
poly_widest_int rather than a widest_int.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-affine.h (get_inner_reference_aff): Return the size as a
poly_widest_int.
* tree-affine.c (get_inner_reference_aff): Likewise.
* tree-data-ref.c (dr_may_alias_p): Update accordingly.
* tree-ssa-loop-im.c (mem_refs_may_alias_p): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255891

poly_int: pointer_may_wrap_p

This patch changes the bitpos argument to pointer_may_wrap_p from
HOST_WIDE_INT to poly_int64.  A later patch makes the callers track
polynomial offsets.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* fold-const.c (pointer_may_wrap_p): Take the offset as a
HOST_WIDE_INT rather than a poly_int64.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255890

poly_int: symbolic_number

This patch changes symbol_number::bytepos from a HOST_WIDE_INT
to a poly_int64.  perform_symbolic_merge can cope with symbolic
offsets as long as the difference between the two offsets is
constant.  (This could happen for a constant-sized field that
occurs at a variable offset, for example.)

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* gimple-ssa-store-merging.c (symbolic_number::bytepos): Change from
HOST_WIDE_INT to poly_int64_pod.
(perform_symbolic_merge): Update accordingly.
(bswap_replace): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255889

poly_int: aff_tree

This patch changes the type of aff_tree::offset from widest_int to
poly_widest_int and adjusts the function interfaces in the same way.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-affine.h (aff_tree::offset): Change from widest_int
to poly_widest_int.
(wide_int_ext_for_comb): Delete.
(aff_combination_const, aff_comb_cannot_overlap_p): Take the
constants as poly_widest_int rather than widest_int.
(aff_combination_constant_multiple_p): Return the multiplier
as a poly_widest_int.
(aff_combination_zero_p, aff_combination_singleton_var_p): Handle
polynomial offsets.
* tree-affine.c (wide_int_ext_for_comb): Make original widest_int
version static and add an overload for poly_widest_int.
(aff_combination_const, aff_combination_add_cst)
(wide_int_constant_multiple_p, aff_comb_cannot_overlap_p): Take
the constants as poly_widest_int rather than widest_int.
(tree_to_aff_combination): Generalize INTEGER_CST case to
poly_int_tree_p.
(aff_combination_to_tree): Track offsets as poly_widest_ints.
(aff_combination_add_product, aff_combination_mult): Handle
polynomial offsets.
(aff_combination_constant_multiple_p): Return the multiplier
as a poly_widest_int.
* tree-predcom.c (determine_offset): Return the offset as a
poly_widest_int.
(split_data_refs_to_components, suitable_component_p): Update
accordingly.
(valid_initializer_p): Update call to
aff_combination_constant_multiple_p.
* tree-ssa-address.c (addr_to_parts): Handle polynomial offsets.
* tree-ssa-loop-ivopts.c (get_address_cost_ainc): Take the step
as a poly_int64 rather than a HOST_WIDE_INT.
(get_address_cost): Handle polynomial offsets.
(iv_elimination_compare_lt): Likewise.
(rewrite_use_nonlinear_expr): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255888

poly_int: get_addr_base_and_unit_offset

This patch changes the values returned by
get_addr_base_and_unit_offset from HOST_WIDE_INT to poly_int64.

maxsize in gimple_fold_builtin_memory_op goes from HOST_WIDE_INT
to poly_uint64 (rather than poly_int) to match the previous use
of tree_fits_uhwi_p.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-dfa.h (get_addr_base_and_unit_offset_1): Return the offset
as a poly_int64_pod rather than a HOST_WIDE_INT.
(get_addr_base_and_unit_offset): Likewise.
* tree-dfa.c (get_addr_base_and_unit_offset_1): Likewise.
(get_addr_base_and_unit_offset): Likewise.
* doc/match-and-simplify.texi: Change off from HOST_WIDE_INT
to poly_int64 in example.
* fold-const.c (fold_binary_loc): Update call to
get_addr_base_and_unit_offset.
* gimple-fold.c (gimple_fold_builtin_memory_op): Likewise.
(maybe_canonicalize_mem_ref_addr): Likewise.
(gimple_fold_stmt_to_constant_1): Likewise.
* gimple-ssa-warn-restrict.c (builtin_memref::builtin_memref):
Likewise.
* ipa-param-manipulation.c (ipa_modify_call_arguments): Likewise.
* match.pd: Likewise.
* omp-low.c (lower_omp_target): Likewise.
* tree-sra.c (build_ref_for_offset): Likewise.
(build_debug_ref_for_model): Likewise.
* tree-ssa-address.c (maybe_fold_tmr): Likewise.
* tree-ssa-alias.c (ao_ref_init_from_ptr_and_size): Likewise.
* tree-ssa-ccp.c (optimize_memcpy): Likewise.
* tree-ssa-forwprop.c (forward_propagate_addr_expr_1): Likewise.
(constant_pointer_difference): Likewise.
* tree-ssa-loop-niter.c (expand_simple_operations): Likewise.
* tree-ssa-phiopt.c (jump_function_from_stmt): Likewise.
* tree-ssa-pre.c (create_component_ref_by_pieces_1): Likewise.
* tree-ssa-sccvn.c (vn_reference_fold_indirect): Likewise.
(vn_reference_maybe_forwprop_address, vn_reference_lookup_3): Likewise.
(set_ssa_val_to): Likewise.
* tree-ssa-strlen.c (get_addr_stridx, addr_stridxptr)
(maybe_diag_stxncpy_trunc): Likewise.
* tree-vrp.c (vrp_prop::check_array_ref): Likewise.
* tree.c (build_simple_mem_ref_loc): Likewise.
(array_at_struct_end_p): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255887

poly_int: get_ref_base_and_extent

This patch changes the types of the bit offsets and sizes returned
by get_ref_base_and_extent to poly_int64.

There are some callers that can't sensibly operate on polynomial
offsets or handle cases where the offset and size aren't known
exactly.  This includes the IPA devirtualisation code (since
there's no defined way of having vtables at variable offsets)
and some parts of the DWARF code.  The patch therefore adds
a helper function get_ref_base_and_extent_hwi that either returns
exact HOST_WIDE_INT bit positions and sizes or returns a null
base to indicate failure.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-dfa.h (get_ref_base_and_extent): Return the base, size and
max_size as poly_int64_pods rather than HOST_WIDE_INTs.
(get_ref_base_and_extent_hwi): Declare.
* tree-dfa.c (get_ref_base_and_extent): Return the base, size and
max_size as poly_int64_pods rather than HOST_WIDE_INTs.
(get_ref_base_and_extent_hwi): New function.
* cfgexpand.c (expand_debug_expr): Update call to
get_ref_base_and_extent.
* dwarf2out.c (add_var_loc_to_decl): Likewise.
* gimple-fold.c (get_base_constructor): Return the offset as a
poly_int64_pod rather than a HOST_WIDE_INT.
(fold_const_aggregate_ref_1): Track polynomial sizes and offsets.
* ipa-polymorphic-call.c
(ipa_polymorphic_call_context::set_by_invariant)
(extr_type_from_vtbl_ptr_store): Track polynomial offsets.
(ipa_polymorphic_call_context::ipa_polymorphic_call_context)
(check_stmt_for_type_change): Use get_ref_base_and_extent_hwi
rather than get_ref_base_and_extent.
(ipa_polymorphic_call_context::get_dynamic_type): Likewise.
* ipa-prop.c (ipa_load_from_parm_agg, compute_complex_assign_jump_func)
(get_ancestor_addr_info, determine_locally_known_aggregate_parts):
Likewise.
* ipa-param-manipulation.c (ipa_get_adjustment_candidate): Update
call to get_ref_base_and_extent.
* tree-sra.c (create_access, get_access_for_expr): Likewise.
* tree-ssa-alias.c (ao_ref_base, aliasing_component_refs_p)
(stmt_kills_ref_p): Likewise.
* tree-ssa-dce.c (mark_aliased_reaching_defs_necessary_1): Likewise.
* tree-ssa-scopedtables.c (avail_expr_hash, equal_mem_array_ref_p):
Likewise.
* tree-ssa-sccvn.c (vn_reference_lookup_3): Likewise.
Use get_ref_base_and_extent_hwi rather than get_ref_base_and_extent
when calling native_encode_expr.
* tree-ssa-structalias.c (get_constraint_for_component_ref): Update
call to get_ref_base_and_extent.
(do_structure_copy): Use get_ref_base_and_extent_hwi rather than
get_ref_base_and_extent.
* var-tracking.c (track_expr_p): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255886

poly_int: ipa_parm_adjustment

This patch changes the type of ipa_parm_adjustment::offset from
HOST_WIDE_INT to poly_int64 and updates uses accordingly.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* ipa-param-manipulation.h (ipa_parm_adjustment::offset): Change from
HOST_WIDE_INT to poly_int64_pod.
* ipa-param-manipulation.c (ipa_modify_call_arguments): Track
polynomail parameter offsets.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255885

poly_int: DWARF CFA offsets

This patch makes the DWARF code use poly_int64 rather than
HOST_WIDE_INT for CFA offsets.  The main changes are:

- to make reg_save use a DW_CFA_expression representation when
  the offset isn't constant and

- to record the CFA information alongside a def_cfa_expression
  if either offset is polynomial, since it's quite difficult
  to reconstruct the CFA information otherwise.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* gengtype.c (main): Handle poly_int64_pod.
* dwarf2out.h (dw_cfi_oprnd_cfa_loc): New dw_cfi_oprnd_type.
(dw_cfi_oprnd::dw_cfi_cfa_loc): New field.
(dw_cfa_location::offset, dw_cfa_location::base_offset): Change
from HOST_WIDE_INT to poly_int64_pod.
* dwarf2cfi.c (queued_reg_save::cfa_offset): Likewise.
(copy_cfa): New function.
(lookup_cfa_1): Use the cached dw_cfi_cfa_loc, if it exists.
(cfi_oprnd_equal_p): Handle dw_cfi_oprnd_cfa_loc.
(cfa_equal_p, dwarf2out_frame_debug_adjust_cfa)
(dwarf2out_frame_debug_cfa_offset, dwarf2out_frame_debug_expr)
(initial_return_save): Treat offsets as poly_ints.
(def_cfa_0): Likewise.  Cache the CFA in dw_cfi_cfa_loc if either
offset is nonconstant.
(reg_save): Take the offset as a poly_int64.  Fall back to
DW_CFA_expression for nonconstant offsets.
(queue_reg_save): Take the offset as a poly_int64.
* dwarf2out.c (dw_cfi_oprnd2_desc): Handle DW_CFA_def_cfa_expression.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255884

poly_int: operand_subword

This patch makes operand_subword and operand_subword_force take
polynomial offsets.  This is a fairly old-school interface and
these days should only be used when splitting multiword operations
into word operations.  It still doesn't hurt to support polynomial
offsets and it helps make callers easier to write.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtl.h (operand_subword, operand_subword_force): Take the offset
as a poly_uint64 an unsigned int.
* emit-rtl.c (operand_subword, operand_subword_force): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255883

poly_int: SUBREG_BYTE

This patch changes SUBREG_BYTE from an int to a poly_int.
Since valid SUBREG_BYTEs must be contained within the mode of the
SUBREG_REG, the required range is the same as for GET_MODE_SIZE,
i.e. unsigned short.  The patch therefore uses poly_uint16(_pod)
for the SUBREG_BYTE.

Using poly_uint16_pod rtx fields requires a new field code ('p').
Since there are no other uses of 'p' besides SUBREG_BYTE, the patch
doesn't add an XPOLY or whatever; all uses should go via SUBREG_BYTE
instead.

The patch doesn't bother implementing 'p' support for legacy
define_peepholes, since none of the remaining ones have subregs
in their patterns.

As it happened, the rtl documentation used SUBREG as an example of a
code with mixed field types, accessed via XEXP (x, 0) and XINT (x, 1).
Since there's no direct replacement for XINT, and since people should
never use it even if there were, the patch changes the example to use
INT_LIST instead.

The patch also changes subreg-related helper functions so that they too
take and return polynomial offsets.  This makes the patch quite big, but
it's mostly mechanical.  The patch generally sticks to existing choices
wrt signedness.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/rtl.texi: Update documentation of SUBREG_BYTE.  Document the
'p' format code.  Use INT_LIST rather than SUBREG as the example of
a code with an XINT and an XEXP.  Remove the implication that
accessing an rtx field using XINT is expected to work.
* rtl.def (SUBREG): Change format from "ei" to "ep".
* rtl.h (rtunion::rt_subreg): New field.
(XCSUBREG): New macro.
(SUBREG_BYTE): Use it.
(subreg_shape): Change offset from an unsigned int to a poly_uint16.
Update constructor accordingly.
(subreg_shape::operator ==): Update accordingly.
(subreg_shape::unique_id): Return an unsigned HOST_WIDE_INT rather
than an unsigned int.
(subreg_lsb, subreg_lowpart_offset, subreg_highpart_offset): Return
a poly_uint64 rather than an unsigned int.
(subreg_lsb_1): Likewise.  Take the offset as a poly_uint64 rather
than an unsigned int.
(subreg_size_offset_from_lsb, subreg_size_lowpart_offset)
(subreg_size_highpart_offset): Return a poly_uint64 rather than
an unsigned int.  Take the sizes as poly_uint64s.
(subreg_offset_from_lsb): Return a poly_uint64 rather than
an unsigned int.  Take the shift as a poly_uint64 rather than
an unsigned int.
(subreg_regno_offset, subreg_offset_representable_p): Take the offset
as a poly_uint64 rather than an unsigned int.
(simplify_subreg_regno): Likewise.
(byte_lowpart_offset): Return the memory offset as a poly_int64
rather than an int.
(subreg_memory_offset): Likewise.  Take the subreg offset as a
poly_uint64 rather than an unsigned int.
(simplify_subreg, simplify_gen_subreg, subreg_get_info)
(gen_rtx_SUBREG, validate_subreg): Take the subreg offset as a
poly_uint64 rather than an unsigned int.
* rtl.c (rtx_format): Describe 'p' in comment.
(copy_rtx, rtx_equal_p_cb, rtx_equal_p): Handle 'p'.
* emit-rtl.c (validate_subreg, gen_rtx_SUBREG): Take the subreg
offset as a poly_uint64 rather than an unsigned int.
(byte_lowpart_offset): Return the memory offset as a poly_int64
rather than an int.
(subreg_memory_offset): Likewise.  Take the subreg offset as a
poly_uint64 rather than an unsigned int.
(subreg_size_lowpart_offset, subreg_size_highpart_offset): Take the
mode sizes as poly_uint64s rather than unsigned ints.  Return a
poly_uint64 rather than an unsigned int.
(subreg_lowpart_p): Treat subreg offsets as poly_ints.
(copy_insn_1): Handle 'p'.
* rtlanal.c (set_noop_p): Treat subregs offsets as poly_uint64s.
(subreg_lsb_1): Take the subreg offset as a poly_uint64 rather than
an unsigned int.  Return the shift in the same way.
(subreg_lsb): Return the shift as a poly_uint64 rather than an
unsigned int.
(subreg_size_offset_from_lsb): Take the sizes and shift as
poly_uint64s rather than unsigned ints.  Return the offset as
a poly_uint64.
(subreg_get_info, subreg_regno_offset, subreg_offset_representable_p)
(simplify_subreg_regno): Take the offset as a poly_uint64 rather than
an unsigned int.
* rtlhash.c (add_rtx): Handle 'p'.
* genemit.c (gen_exp): Likewise.
* gengenrtl.c (type_from_format, gendef): Likewise.
* gensupport.c (subst_pattern_match, get_alternatives_number)
(collect_insn_data, alter_predicate_for_insn, alter_constraints)
(subst_dup): Likewise.
* gengtype.c (adjust_field_rtx_def): Likewise.
* genrecog.c (find_operand, find_matching_operand, validate_pattern)
(match_pattern_2): Likewise.
(rtx_test::SUBREG_FIELD): New rtx_test::kind_enum.
(rtx_test::subreg_field): New function.
(operator ==, safe_to_hoist_p, transition_parameter_type)
(print_nonbool_test, print_test): Handle SUBREG_FIELD.
* genattrtab.c (attr_rtx_1): Say that 'p' is deliberately not handled.
* genpeep.c (match_rtx): Likewise.
* print-rtl.c (print_poly_int): Include if GENERATOR_FILE too.
(rtx_writer::print_rtx_operand): Handle 'p'.
(print_value): Handle SUBREG.
* read-rtl.c (apply_int_iterator): Likewise.
(rtx_reader::read_rtx_operand): Handle 'p'.
* alias.c (rtx_equal_for_memref_p): Likewise.
* cselib.c (rtx_equal_for_cselib_1, cselib_hash_rtx): Likewise.
* caller-save.c (replace_reg_with_saved_mem): Treat subreg offsets
as poly_ints.
* calls.c (expand_call): Likewise.
* combine.c (combine_simplify_rtx, expand_field_assignment): Likewise.
(make_extraction, gen_lowpart_for_combine): Likewise.
* loop-invariant.c (hash_invariant_expr_1, invariant_expr_equal_p):
Likewise.
* cse.c (remove_invalid_subreg_refs): Take the offset as a poly_uint64
rather than an unsigned int.  Treat subreg offsets as poly_ints.
(exp_equiv_p): Handle 'p'.
(hash_rtx_cb): Likewise.  Treat subreg offsets as poly_ints.
(equiv_constant, cse_insn): Treat subreg offsets as poly_ints.
* dse.c (find_shift_sequence): Likewise.
* dwarf2out.c (rtl_for_decl_location): Likewise.
* expmed.c (extract_low_bits): Likewise.
* expr.c (emit_group_store, undefined_operand_subword_p): Likewise.
(expand_expr_real_2): Likewise.
* final.c (alter_subreg): Likewise.
(leaf_renumber_regs_insn): Handle 'p'.
* function.c (assign_parm_find_stack_rtl, assign_parm_setup_stack):
Treat subreg offsets as poly_ints.
* fwprop.c (forward_propagate_and_simplify): Likewise.
* ifcvt.c (noce_emit_move_insn, noce_emit_cmove): Likewise.
* ira.c (get_subreg_tracking_sizes): Likewise.
* ira-conflicts.c (go_through_subreg): Likewise.
* ira-lives.c (process_single_reg_class_operands): Likewise.
* jump.c (rtx_renumbered_equal_p): Likewise.  Handle 'p'.
* lower-subreg.c (simplify_subreg_concatn): Take the subreg offset
as a poly_uint64 rather than an unsigned int.
(simplify_gen_subreg_concatn, resolve_simple_move): Treat
subreg offsets as poly_ints.
* lra-constraints.c (operands_match_p): Handle 'p'.
(match_reload, curr_insn_transform): Treat subreg offsets as poly_ints.
* lra-spills.c (assign_mem_slot): Likewise.
* postreload.c (move2add_valid_value_p): Likewise.
* recog.c (general_operand, indirect_operand): Likewise.
* regcprop.c (copy_value, maybe_mode_change): Likewise.
(copyprop_hardreg_forward_1): Likewise.
* reginfo.c (simplifiable_subregs_hasher::hash, simplifiable_subregs)
(record_subregs_of_mode): Likewise.
* rtlhooks.c (gen_lowpart_general, gen_lowpart_if_possible): Likewise.
* reload.c (operands_match_p): Handle 'p'.
(find_reloads_subreg_address): Treat subreg offsets as poly_ints.
* reload1.c (alter_reg, choose_reload_regs): Likewise.
(compute_reload_subreg_offset): Likewise, and return an poly_int64.
* simplify-rtx.c (simplify_truncation, simplify_binary_operation_1):
(test_vector_ops_duplicate): Treat subreg offsets as poly_ints.
(simplify_const_poly_int_tests<N>::run): Likewise.
(simplify_subreg, simplify_gen_subreg): Take the subreg offset as
a poly_uint64 rather than an unsigned int.
* valtrack.c (debug_lowpart_subreg): Likewise.
* var-tracking.c (var_lowpart): Likewise.
(loc_cmp): Handle 'p'.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255882

poly_int: ira subreg liveness tracking

Normmaly the IRA-reload interface tries to track the liveness of
individual bytes of an allocno if the allocno is sometimes written
to as a SUBREG.  This isn't possible for variable-sized allocnos,
but it doesn't matter because targets with variable-sized registers
should use LRA instead.

This patch adds a get_subreg_tracking_sizes function for deciding
whether it is possible to model a partial read or write.  Later
patches make it return false if anything is variable.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* ira.c (get_subreg_tracking_sizes): New function.
(init_live_subregs): Take an integer size rather than a register.
(build_insn_chain): Use get_subreg_tracking_sizes.  Update calls
to init_live_subregs.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255881

poly_int: store_field & co

This patch makes store_field and related routines use poly_ints
for bit positions and sizes.  It keeps the existing choices
between signed and unsigned types (there are a mixture of both).

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expr.c (store_constructor_field): Change bitsize from a
unsigned HOST_WIDE_INT to a poly_uint64 and bitpos from a
HOST_WIDE_INT to a poly_int64.
(store_constructor): Change size from a HOST_WIDE_INT to
a poly_int64.
(store_field): Likewise bitsize and bitpos.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255880

poly_int: C++ bitfield regions

This patch changes C++ bitregion_start/end values from constants to
poly_ints.  Although it's unlikely that the size needs to be polynomial
in practice, the offset could be with future language extensions.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expmed.h (store_bit_field): Change bitregion_start and
bitregion_end from unsigned HOST_WIDE_INT to poly_uint64.
* expmed.c (adjust_bit_field_mem_for_reg, strict_volatile_bitfield_p)
(store_bit_field_1, store_integral_bit_field, store_bit_field)
(store_fixed_bit_field, store_split_bit_field): Likewise.
* expr.c (store_constructor_field, store_field): Likewise.
(optimize_bitfield_assignment_op): Likewise.  Make the same change
to bitsize and bitpos.
* machmode.h (bit_field_mode_iterator): Change m_bitregion_start
and m_bitregion_end from HOST_WIDE_INT to poly_int64.  Make the
same change in the constructor arguments.
(get_best_mode): Change bitregion_start and bitregion_end from
unsigned HOST_WIDE_INT to poly_uint64.
* stor-layout.c (bit_field_mode_iterator::bit_field_mode_iterator):
Change bitregion_start and bitregion_end from HOST_WIDE_INT to
poly_int64.
(bit_field_mode_iterator::next_mode): Update for new types
of m_bitregion_start and m_bitregion_end.
(get_best_mode): Change bitregion_start and bitregion_end from
unsigned HOST_WIDE_INT to poly_uint64.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255879

poly_int: extract_bit_field bitrange

Similar to the previous store_bit_field patch, but for extractions
rather than insertions.  The patch splits out the extraction-as-subreg
handling into a new function (extract_bit_field_as_subreg), both for
ease of writing and because a later patch will add another caller.

The simplify_gen_subreg overload is temporary; it goes away
in a later patch.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtl.h (simplify_gen_subreg): Add a temporary overload that
accepts poly_uint64 offsets.
* expmed.h (extract_bit_field): Take bitsize and bitnum as
poly_uint64s rather than unsigned HOST_WIDE_INTs.
* expmed.c (lowpart_bit_field_p): Likewise.
(extract_bit_field_as_subreg): New function, split out from...
(extract_bit_field_1): ...here.  Take bitsize and bitnum as
poly_uint64s rather than unsigned HOST_WIDE_INTs.  For vector
extractions, check that BITSIZE matches the size of the extracted
value and that BITNUM is an exact multiple of that size.
If all else fails, try forcing the value into memory if
BITNUM is variable, and adjusting the address so that the
offset is constant.  Split the part that can only handle constant
bitsize and bitnum out into...
(extract_integral_bit_field): ...this new function.
(extract_bit_field): Take bitsize and bitnum as poly_uint64s
rather than unsigned HOST_WIDE_INTs.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255878

poly_int: store_bit_field bitrange

This patch changes the bitnum and bitsize arguments to
store_bit_field from unsigned HOST_WIDE_INTs to poly_uint64s.
The later part of store_bit_field_1 still needs to operate
on constant bit positions and sizes, so the patch splits
it out into a subfunction (store_integral_bit_field).

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* expmed.h (store_bit_field): Take bitsize and bitnum as
poly_uint64s rather than unsigned HOST_WIDE_INTs.
* expmed.c (simple_mem_bitfield_p): Likewise.  Add a parameter
that returns the byte size.
(store_bit_field_1): Take bitsize and bitnum as
poly_uint64s rather than unsigned HOST_WIDE_INTs.  Update call
to simple_mem_bitfield_p.  Split the part that can only handle
constant bitsize and bitnum out into...
(store_integral_bit_field): ...this new function.
(store_bit_field): Take bitsize and bitnum as poly_uint64s rather
than unsigned HOST_WIDE_INTs.
(extract_bit_field_1): Update call to simple_mem_bitfield_p.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255877

poly_int: lra frame offsets

This patch makes LRA use poly_int64s rather than HOST_WIDE_INTs
to store a frame offset (including in things like eliminations).

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* lra-int.h (lra_reg): Change offset from int to poly_int64.
(lra_insn_recog_data): Change sp_offset from HOST_WIDE_INT
to poly_int64.
(lra_eliminate_regs_1, eliminate_regs_in_insn): Change
update_sp_offset from a HOST_WIDE_INT to a poly_int64.
(lra_update_reg_val_offset, lra_reg_val_equal_p): Take the
offset as a poly_int64 rather than an int.
* lra-assigns.c (find_hard_regno_for_1): Handle poly_int64 offsets.
(setup_live_pseudos_and_spill_after_risky_transforms): Likewise.
* lra-constraints.c (equiv_address_substitution): Track offsets
as poly_int64s.
(emit_inc): Check poly_int_rtx_p instead of CONST_INT_P.
(curr_insn_transform): Handle the new form of sp_offset.
* lra-eliminations.c (lra_elim_table): Change previous_offset
and offset from HOST_WIDE_INT to poly_int64.
(print_elim_table, update_reg_eliminate): Update accordingly.
(self_elim_offsets): Change from HOST_WIDE_INT to poly_int64_pod.
(get_elimination): Update accordingly.
(form_sum): Check poly_int_rtx_p instead of CONST_INT_P.
(lra_eliminate_regs_1, eliminate_regs_in_insn): Change
update_sp_offset from a HOST_WIDE_INT to a poly_int64.  Handle
poly_int64 offsets generally.
(curr_sp_change): Change from HOST_WIDE_INT to poly_int64.
(mark_not_eliminable, init_elimination): Update accordingly.
(remove_reg_equal_offset_note): Return a bool and pass the new
offset back by pointer as a poly_int64.
* lra-remat.c (change_sp_offset): Take sp_offset as a poly_int64
rather than a HOST_WIDE_INT.
(do_remat): Track offsets poly_int64s.
* lra.c (lra_update_insn_recog_data, setup_sp_offset): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255876

poly_int: MEM_OFFSET and MEM_SIZE

This patch changes the MEM_OFFSET and MEM_SIZE memory attributes
from HOST_WIDE_INT to poly_int64.  Most of it is mechanical,
but there is one nonbovious change in widen_memory_access.
Previously the main while loop broke with:

      /* Similarly for the decl.  */
      else if (DECL_P (attrs.expr)
               && DECL_SIZE_UNIT (attrs.expr)
               && TREE_CODE (DECL_SIZE_UNIT (attrs.expr)) == INTEGER_CST
               && compare_tree_int (DECL_SIZE_UNIT (attrs.expr), size) >= 0
               && (! attrs.offset_known_p || attrs.offset >= 0))
        break;

but it seemed wrong to optimistically assume the best case
when the offset isn't known (and thus might be negative).
As it happens, the "! attrs.offset_known_p" condition was
always false, because we'd already nullified attrs.expr in
that case:

  /* If we don't know what offset we were at within the expression, then
     we can't know if we've overstepped the bounds.  */
  if (! attrs.offset_known_p)
    attrs.expr = NULL_TREE;

The patch therefore drops "! attrs.offset_known_p ||" when
converting the offset check to the may/must interface.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtl.h (mem_attrs): Add a default constructor.  Change size and
offset from HOST_WIDE_INT to poly_int64.
* emit-rtl.h (set_mem_offset, set_mem_size, adjust_address_1)
(adjust_automodify_address_1, set_mem_attributes_minus_bitpos)
(widen_memory_access): Take the sizes and offsets as poly_int64s
rather than HOST_WIDE_INTs.
* alias.c (ao_ref_from_mem): Handle the new form of MEM_OFFSET.
(offset_overlap_p): Take poly_int64s rather than HOST_WIDE_INTs
and ints.
(adjust_offset_for_component_ref): Change the offset from a
HOST_WIDE_INT to a poly_int64.
(nonoverlapping_memrefs_p): Track polynomial offsets and sizes.
* cfgcleanup.c (merge_memattrs): Update after mem_attrs changes.
* dce.c (find_call_stack_args): Likewise.
* dse.c (record_store): Likewise.
* dwarf2out.c (tls_mem_loc_descriptor, dw_sra_loc_expr): Likewise.
* print-rtl.c (rtx_writer::print_rtx): Likewise.
* read-rtl-function.c (test_loading_mem): Likewise.
* rtlanal.c (may_trap_p_1): Likewise.
* simplify-rtx.c (delegitimize_mem_from_attrs): Likewise.
* var-tracking.c (int_mem_offset, track_expr_p): Likewise.
* emit-rtl.c (mem_attrs_eq_p, get_mem_align_offset): Likewise.
(mem_attrs::mem_attrs): New function.
(set_mem_attributes_minus_bitpos): Change bitpos from a
HOST_WIDE_INT to poly_int64.
(set_mem_alias_set, set_mem_addr_space, set_mem_align, set_mem_expr)
(clear_mem_offset, clear_mem_size, change_address)
(get_spill_slot_decl, set_mem_attrs_for_spill): Directly
initialize mem_attrs.
(set_mem_offset, set_mem_size, adjust_address_1)
(adjust_automodify_address_1, offset_address, widen_memory_access):
Likewise.  Take poly_int64s rather than HOST_WIDE_INT.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255875

poly_int: rtx_addr_can_trap_p_1

This patch changes the offset and size arguments of
rtx_addr_can_trap_p_1 from HOST_WIDE_INT to poly_int64.  It also
uses a size of -1 rather than 0 to represent an unknown size and
BLKmode rather than VOIDmode to represent an unknown mode.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtlanal.c (rtx_addr_can_trap_p_1): Take the offset and size
as poly_int64s rather than HOST_WIDE_INTs.  Use a size of -1
rather than 0 to represent an unknown size.  Assert that the size
is known when the mode isn't BLKmode.
(may_trap_p_1): Use -1 for unknown sizes.
(rtx_addr_can_trap_p): Likewise.  Pass BLKmode rather than VOIDmode.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255874

poly_int: dse.c

This patch makes RTL DSE use poly_int for offsets and sizes.
The local phase can optimise them normally but the global phase
treats them as wild accesses.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* dse.c (store_info): Change offset and width from HOST_WIDE_INT
to poly_int64.  Update commentary for positions_needed.large.
(read_info_type): Change offset and width from HOST_WIDE_INT
to poly_int64.
(set_usage_bits): Likewise.
(canon_address): Return the offset as a poly_int64 rather than
a HOST_WIDE_INT.  Use strip_offset_and_add.
(set_all_positions_unneeded, any_positions_needed_p): Use
positions_needed.large to track stores with non-constant widths.
(all_positions_needed_p): Likewise.  Take the offset and width
as poly_int64s rather than ints.  Assert that rhs is nonnull.
(record_store): Cope with non-constant offsets and widths.
Nullify the rhs of an earlier store if we can't tell which bytes
of it are needed.
(find_shift_sequence): Take the access_size and shift as poly_int64s
rather than ints.
(get_stored_val): Take the read_offset and read_width as poly_int64s
rather than HOST_WIDE_INTs.
(check_mem_read_rtx, scan_stores, scan_reads, dse_step5): Handle
non-constant offsets and widths.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255873

poly_int: ao_ref and vn_reference_op_t

This patch changes the offset, size and max_size fields
of ao_ref from HOST_WIDE_INT to poly_int64 and propagates
the change through the code that references it.  This includes
changing the off field of vn_reference_op_struct in the same way.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* inchash.h (inchash::hash::add_poly_int): New function.
* tree-ssa-alias.h (ao_ref::offset, ao_ref::size, ao_ref::max_size):
Use poly_int64 rather than HOST_WIDE_INT.
(ao_ref::max_size_known_p): New function.
* tree-ssa-sccvn.h (vn_reference_op_struct::off): Use poly_int64_pod
rather than HOST_WIDE_INT.
* tree-ssa-alias.c (ao_ref_base): Apply get_ref_base_and_extent
to temporaries until its interface is adjusted to match.
(ao_ref_init_from_ptr_and_size): Handle polynomial offsets and sizes.
(aliasing_component_refs_p, decl_refs_may_alias_p)
(indirect_ref_may_alias_decl_p, indirect_refs_may_alias_p): Take
the offsets and max_sizes as poly_int64s instead of HOST_WIDE_INTs.
(refs_may_alias_p_1, stmt_kills_ref_p): Adjust for changes to
ao_ref fields.
* alias.c (ao_ref_from_mem): Likewise.
* tree-ssa-dce.c (mark_aliased_reaching_defs_necessary_1): Likewise.
* tree-ssa-dse.c (valid_ao_ref_for_dse, normalize_ref)
(clear_bytes_written_by, setup_live_bytes_from_ref, compute_trims)
(maybe_trim_complex_store, maybe_trim_constructor_store)
(live_bytes_read, dse_classify_store): Likewise.
* tree-ssa-sccvn.c (vn_reference_compute_hash, vn_reference_eq):
(copy_reference_ops_from_ref, ao_ref_init_from_vn_reference)
(fully_constant_vn_reference_p, valueize_refs_1): Likewise.
(vn_reference_lookup_3): Likewise.
* tree-ssa-uninit.c (warn_uninitialized_vars): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255872

poly_int: indirect_refs_may_alias_p

This patch makes indirect_refs_may_alias_p use ranges_may_overlap_p
rather than ranges_overlap_p.  Unlike the former, the latter can handle
negative offsets, so the fix for PR44852 should no longer be necessary.
It can also handle offset_int, so avoids unchecked truncations to
HOST_WIDE_INT.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssa-alias.c (indirect_ref_may_alias_decl_p)
(indirect_refs_may_alias_p): Use ranges_may_overlap_p
instead of ranges_overlap_p.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255871

poly_int: same_addr_size_stores_p

This patch makes tree-ssa-alias.c:same_addr_size_stores_p handle
poly_int sizes and offsets.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssa-alias.c (same_addr_size_stores_p): Take the offsets and
sizes as poly_int64s rather than HOST_WIDE_INTs.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255870

poly_int: fold_ctor_reference

This patch changes the offset and size arguments to
fold_ctor_reference from unsigned HOST_WIDE_INT to poly_uint64.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* gimple-fold.h (fold_ctor_reference): Take the offset and size
as poly_uint64 rather than unsigned HOST_WIDE_INT.
* gimple-fold.c (fold_ctor_reference): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255869

poly_int: DWARF locations

This patch adds support for DWARF location expressions
that involve polynomial offsets.  It adds a target hook that
says how the runtime invariants used in the offsets should be
represented in DWARF.  SVE vectors have to be a multiple of
128 bits in size, so the GCC port uses the number of 128-bit
blocks minus one as the runtime invariant.  However, in DWARF,
the vector length is exposed via a pseudo "VG" register that
holds the number of 64-bit elements in a vector.  Thus:

  indeterminate 1 == (VG / 2) - 1

The hook needs to be general enough to express this.
Note that in most cases the division and subtraction fold
away into surrounding expressions.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* target.def (dwarf_poly_indeterminate_value): New hook.
* targhooks.h (default_dwarf_poly_indeterminate_value): Declare.
* targhooks.c (default_dwarf_poly_indeterminate_value): New function.
* doc/tm.texi.in (TARGET_DWARF_POLY_INDETERMINATE_VALUE): Document.
* doc/tm.texi: Regenerate.
* dwarf2out.h (build_cfa_loc, build_cfa_aligned_loc): Take the
offset as a poly_int64.
* dwarf2out.c (new_reg_loc_descr): Move later in file.  Take the
offset as a poly_int64.
(loc_descr_plus_const, loc_list_plus_const, build_cfa_aligned_loc):
Take the offset as a poly_int64.
(build_cfa_loc): Likewise.  Use loc_descr_plus_const.
(frame_pointer_fb_offset): Change to a poly_int64.
(int_loc_descriptor): Take the offset as a poly_int64.  Use
targetm.dwarf_poly_indeterminate_value for polynomial offsets.
(based_loc_descr): Take the offset as a poly_int64.
Use strip_offset_and_add to handle (plus X (const)).
Use new_reg_loc_descr instead of an open-coded version of the
previous implementation.
(mem_loc_descriptor): Handle CONST_POLY_INT.
(compute_frame_pointer_to_fb_displacement): Take the offset as a
poly_int64.  Use strip_offset_and_add to handle (plus X (const)).

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255868

poly_int: REG_OFFSET

This patch changes the type of the reg_attrs offset field
from HOST_WIDE_INT to poly_int64 and updates uses accordingly.
This includes changing reg_attr_hasher::hash to use inchash.
(Doing this has no effect on code generation since the only
use of the hasher is to avoid creating duplicate objects.)

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
            Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* rtl.h (reg_attrs::offset): Change from HOST_WIDE_INT to poly_int64.
(gen_rtx_REG_offset): Take the offset as a poly_int64.
* inchash.h (inchash::hash::add_poly_hwi): New function.
* gengtype.c (main): Register poly_int64.
* emit-rtl.c (reg_attr_hasher::hash): Use inchash.  Treat the
offset as a poly_int.
(reg_attr_hasher::equal): Use must_eq to compare offsets.
(get_reg_attrs, update_reg_offset, gen_rtx_REG_offset): Take the
offset as a poly_int64.
(set_reg_attrs_from_value): Treat the offset as a poly_int64.
* print-rtl.c (print_poly_int): New function.
(rtx_writer::print_rtx_operand_code_r): Treat REG_OFFSET as
a poly_int.
* var-tracking.c (track_offset_p, get_tracked_reg_offset): New
functions.
(var_reg_set, var_reg_delete_and_set, var_reg_delete): Use them.
(same_variable_part_p, track_loc_p): Take the offset as a poly_int64.
(vt_get_decl_and_offset): Return the offset as a poly_int64.
Enforce track_offset_p for parts of a PARALLEL.
(vt_add_function_parameter): Use const_offset for the final
offset to track.  Use get_tracked_reg_offset for the parts
of a PARALLEL.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255867

poly_int: TRULY_NOOP_TRUNCATION

This patch makes TRULY_NOOP_TRUNCATION take the mode sizes as
poly_uint64s instead of unsigned ints.  The function bodies
don't need to change.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
            Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* target.def (truly_noop_truncation): Take poly_uint64s instead of
unsigned ints.  Change default to hook_bool_puint64_puint64_true.
* doc/tm.texi: Regenerate.
* hooks.h (hook_bool_uint_uint_true): Delete.
(hook_bool_puint64_puint64_true): Declare.
* hooks.c (hook_bool_uint_uint_true): Delete.
(hook_bool_puint64_puint64_true): New function.
* config/mips/mips.c (mips_truly_noop_truncation): Take poly_uint64s
instead of unsigned ints.
* config/spu/spu.c (spu_truly_noop_truncation): Likewise.
* config/tilegx/tilegx.c (tilegx_truly_noop_truncation): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255866

poly_int: create_integer_operand

This patch generalises create_integer_operand so that it accepts
poly_int64s rather than HOST_WIDE_INTs.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* optabs.h (expand_operand): Add an int_value field.
(create_expand_operand): Add an int_value parameter and use it
to initialize the new expand_operand field.
(create_integer_operand): Replace with a declaration of a function
that accepts poly_int64s.  Move the implementation to...
* optabs.c (create_integer_operand): ...here.
(maybe_legitimize_operand): For EXPAND_INTEGER, check whether
the mode preserves the value of int_value, instead of calling
const_int_operand on the rtx.  Use gen_int_mode to generate
the new rtx.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255865

poly_int: dump routines

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* dumpfile.h (dump_dec): Declare.
* dumpfile.c (dump_dec): New function.
* pretty-print.h (pp_wide_integer): Turn into a function and
declare a poly_int version.
* pretty-print.c (pp_wide_integer): New function for poly_ints.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255864

poly_int: tree constants

This patch adds a tree representation for poly_ints.  Unlike the
rtx version, the coefficients are INTEGER_CSTs rather than plain
integers, so that we can easily access them as poly_widest_ints
and poly_offset_ints.

The patch also adjusts some places that previously
relied on "constant" meaning "INTEGER_CST".  It also makes
sure that the TYPE_SIZE agrees with the TYPE_SIZE_UNIT for
vector booleans, given the existing:

/* Several boolean vector elements may fit in a single unit.  */
if (VECTOR_BOOLEAN_TYPE_P (type)
    && type->type_common.mode != BLKmode)
  TYPE_SIZE_UNIT (type)
    = size_int (GET_MODE_SIZE (type->type_common.mode));
else
  TYPE_SIZE_UNIT (type) = int_const_binop (MULT_EXPR,
   TYPE_SIZE_UNIT (innertype),
   size_int (nunits));

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/generic.texi (POLY_INT_CST): Document.
* tree.def (POLY_INT_CST): New tree code.
* treestruct.def (TS_POLY_INT_CST): New tree layout.
* tree-core.h (tree_poly_int_cst): New struct.
(tree_node): Add a poly_int_cst field.
* tree.h (POLY_INT_CST_P, POLY_INT_CST_COEFF): New macros.
(wide_int_to_tree, force_fit_type): Take a poly_wide_int_ref
instead of a wide_int_ref.
(build_int_cst, build_int_cst_type): Take a poly_int64 instead
of a HOST_WIDE_INT.
(build_int_cstu, build_array_type_nelts): Take a poly_uint64
instead of an unsigned HOST_WIDE_INT.
(build_poly_int_cst, tree_fits_poly_int64_p, tree_fits_poly_uint64_p)
(ptrdiff_tree_p): Declare.
(tree_to_poly_int64, tree_to_poly_uint64): Likewise.  Provide
extern inline implementations if the target doesn't use POLY_INT_CST.
(poly_int_tree_p): New function.
(wi::unextended_tree): New class.
(wi::int_traits <unextended_tree>): New override.
(wi::extended_tree): Add a default constructor.
(wi::extended_tree::get_tree): New function.
(wi::widest_extended_tree, wi::offset_extended_tree): New typedefs.
(wi::tree_to_widest_ref, wi::tree_to_offset_ref): Use them.
(wi::tree_to_poly_widest_ref, wi::tree_to_poly_offset_ref)
(wi::tree_to_poly_wide_ref): New typedefs.
(wi::ints_for): Provide overloads for extended_tree and
unextended_tree.
(poly_int_cst_value, wi::to_poly_widest, wi::to_poly_offset)
(wi::to_wide): New functions.
(wi::fits_to_boolean_p, wi::fits_to_tree_p): Handle poly_ints.
* tree.c (poly_int_cst_hasher): New struct.
(poly_int_cst_hash_table): New variable.
(tree_node_structure_for_code, tree_code_size, simple_cst_equal)
(valid_constant_size_p, add_expr, drop_tree_overflow): Handle
POLY_INT_CST.
(initialize_tree_contains_struct): Handle TS_POLY_INT_CST.
(init_ttree): Initialize poly_int_cst_hash_table.
(build_int_cst, build_int_cst_type, build_invariant_address): Take
a poly_int64 instead of a HOST_WIDE_INT.
(build_int_cstu, build_array_type_nelts): Take a poly_uint64
instead of an unsigned HOST_WIDE_INT.
(wide_int_to_tree): Rename to...
(wide_int_to_tree_1): ...this.
(build_new_poly_int_cst, build_poly_int_cst): New functions.
(force_fit_type): Take a poly_wide_int_ref instead of a wide_int_ref.
(wide_int_to_tree): New function that takes a poly_wide_int_ref.
(ptrdiff_tree_p, tree_to_poly_int64, tree_to_poly_uint64)
(tree_fits_poly_int64_p, tree_fits_poly_uint64_p): New functions.
* lto-streamer-out.c (DFS::DFS_write_tree_body, hash_tree): Handle
TS_POLY_INT_CST.
* tree-streamer-in.c (lto_input_ts_poly_tree_pointers): Likewise.
(streamer_read_tree_body): Likewise.
* tree-streamer-out.c (write_ts_poly_tree_pointers): Likewise.
(streamer_write_tree_body): Likewise.
* tree-streamer.c (streamer_check_handled_ts_structures): Likewise.
* asan.c (asan_protect_global): Require the size to be an INTEGER_CST.
* cfgexpand.c (expand_debug_expr): Handle POLY_INT_CST.
* expr.c (expand_expr_real_1, const_vector_from_tree): Likewise.
* gimple-expr.h (is_gimple_constant): Likewise.
* gimplify.c (maybe_with_size_expr): Likewise.
* print-tree.c (print_node): Likewise.
* tree-data-ref.c (data_ref_compare_tree): Likewise.
* tree-pretty-print.c (dump_generic_node): Likewise.
* tree-ssa-address.c (addr_for_mem_ref): Likewise.
* tree-vect-data-refs.c (dr_group_sort_cmp): Likewise.
* tree-vrp.c (compare_values_warnv): Likewise.
* tree-ssa-loop-ivopts.c (determine_base_object, constant_multiple_of)
(get_loop_invariant_expr, add_candidate_1, get_computation_aff_1)
(force_expr_to_var_cost): Likewise.
* tree-ssa-loop.c (for_each_index): Likewise.
* fold-const.h (build_invariant_address, size_int_kind): Take a
poly_int64 instead of a HOST_WIDE_INT.
* fold-const.c (fold_negate_expr_1, const_binop, const_unop)
(fold_convert_const, multiple_of_p, fold_negate_const): Handle
POLY_INT_CST.
(size_binop_loc): Likewise.  Allow int_const_binop_1 to fail.
(int_const_binop_2): New function, split out from...
(int_const_binop_1): ...here.  Handle POLY_INT_CST.
(size_int_kind): Take a poly_int64 instead of a HOST_WIDE_INT.
* expmed.c (make_tree): Handle CONST_POLY_INT_P.
* gimple-ssa-strength-reduction.c (slsr_process_add)
(slsr_process_mul): Check for INTEGER_CSTs before using them
as candidates.
* stor-layout.c (bits_from_bytes): New function.
(bit_from_pos): Use it.
(layout_type): Likewise.  For vectors, multiply the TYPE_SIZE_UNIT
by BITS_PER_UNIT to get the TYPE_SIZE.
* tree-cfg.c (verify_expr, verify_types_in_gimple_reference): Allow
MEM_REF and TARGET_MEM_REF offsets to be a POLY_INT_CST.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255863

poly_int: rtx constants

This patch adds an rtl representation of poly_int values.
There were three possible ways of doing this:

(1) Add a new rtl code for the poly_ints themselves and store the
    coefficients as trailing wide_ints.  This would give constants like:

      (const_poly_int [c0 c1 ... cn])

    The runtime value would be:

      c0 + c1 * x1 + ... + cn * xn

(2) Like (1), but use rtxes for the coefficients.  This would give
    constants like:

      (const_poly_int [(const_int c0)
                       (const_int c1)
                       ...
                       (const_int cn)])

    although the coefficients could be const_wide_ints instead
    of const_ints where appropriate.

(3) Add a new rtl code for the polynomial indeterminates,
    then use them in const wrappers.  A constant like c0 + c1 * x1
    would then look like:

      (const:M (plus:M (mult:M (const_param:M x1)
                               (const_int c1))
                       (const_int c0)))

There didn't seem to be that much to choose between them.  The main
advantage of (1) is that it's a more efficient representation and
that we can refer to the cofficients directly as wide_int_storage.

2017-12-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* doc/rtl.texi (const_poly_int): Document.  Also document the
rtl sharing behavior.
* gengenrtl.c (excluded_rtx): Return true for CONST_POLY_INT.
* rtl.h (const_poly_int_def): New struct.
(rtx_def::u): Add a cpi field.
(CASE_CONST_UNIQUE, CASE_CONST_ANY): Add CONST_POLY_INT.
(CONST_POLY_INT_P, CONST_POLY_INT_COEFFS): New macros.
(wi::rtx_to_poly_wide_ref): New typedef
(const_poly_int_value, wi::to_poly_wide, rtx_to_poly_int64)
(poly_int_rtx_p): New functions.
(trunc_int_for_mode): Declare a poly_int64 version.
(plus_constant): Take a poly_int64 instead of a HOST_WIDE_INT.
(immed_wide_int_const): Take a poly_wide_int_ref rather than
a wide_int_ref.
(strip_offset): Declare.
(strip_offset_and_add): New function.
* rtl.def (CONST_POLY_INT): New rtx code.
* rtl.c (rtx_size): Handle CONST_POLY_INT.
(shared_const_p): Use poly_int_rtx_p.
* emit-rtl.h (gen_int_mode): Take a poly_int64 instead of a
HOST_WIDE_INT.
(gen_int_shift_amount): Likewise.
* emit-rtl.c (const_poly_int_hasher): New class.
(const_poly_int_htab): New variable.
(init_emit_once): Initialize it when NUM_POLY_INT_COEFFS > 1.
(const_poly_int_hasher::hash): New function.
(const_poly_int_hasher::equal): Likewise.
(gen_int_mode): Take a poly_int64 instead of a HOST_WIDE_INT.
(immed_wide_int_const): Rename to...
(immed_wide_int_const_1): ...this and make static.
(immed_wide_int_const): New function, taking a poly_wide_int_ref
instead of a wide_int_ref.
(gen_int_shift_amount): Take a poly_int64 instead of a HOST_WIDE_INT.
(gen_lowpart_common): Handle CONST_POLY_INT.
* cse.c (hash_rtx_cb, equiv_constant): Likewise.
* cselib.c (cselib_hash_rtx): Likewise.
* dwarf2out.c (const_ok_for_output_1): Likewise.
* expr.c (convert_modes): Likewise.
* print-rtl.c (rtx_writer::print_rtx, print_value): Likewise.
* rtlhash.c (add_rtx): Likewise.
* explow.c (trunc_int_for_mode): Add a poly_int64 version.
(plus_constant): Take a poly_int64 instead of a HOST_WIDE_INT.
Handle existing CONST_POLY_INT rtxes.
* expmed.h (expand_shift): Take a poly_int64 instead of a
HOST_WIDE_INT.
* expmed.c (expand_shift): Likewise.
* rtlanal.c (strip_offset): New function.
(commutative_operand_precedence): Give CONST_POLY_INT the same
precedence as CONST_DOUBLE and put CONST_WIDE_INT between that
and CONST_INT.
* rtl-tests.c (const_poly_int_tests): New struct.
(rtl_tests_c_tests): Use it.
* simplify-rtx.c (simplify_const_unary_operation): Handle
CONST_POLY_INT.
(simplify_const_binary_operation): Likewise.
(simplify_binary_operation_1): Fold additions of symbolic constants
and CONST_POLY_INTs.
(simplify_subreg): Handle extensions and truncations of
CONST_POLY_INTs.
(simplify_const_poly_int_tests): New struct.
(simplify_rtx_c_tests): Use it.
* wide-int.h (storage_ref): Add default constructor.
(wide_int_ref_storage): Likewise.
(trailing_wide_ints): Use GTY((user)).
(trailing_wide_ints::operator[]): Add a const version.
(trailing_wide_ints::get_precision): New function.
(trailing_wide_ints::extra_size): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r255862