2002-12-02 Geoffrey Keating <geoffk@apple.com>
* combine.c (combine_simplify_rtx): Add new canonicalizations.
* doc/md.texi (Insn Canonicalizations): Document new
canonicalizations for multiply/add combinations.
* config/rs6000/rs6000.md: Add and modify floating add/multiply
patterns to ensure they're used whenever they can be.
Index: gcc/testsuite/ChangeLog
2002-12-02 Geoffrey Keating <geoffk@apple.com>
* gcc.dg/ppc-fmadd-1.c: New file.
* gcc.dg/ppc-fmadd-2.c: New file.
* gcc.dg/ppc-fmadd-3.c: New file.
From-SVN: r59841
* combine.c (force_to_mode): Only replace with (not Y) if all bits in fuller_mask
(not just mask) are set in C.
* gcc.c-torture/execute/20021118-3.c: New test.
From-SVN: r59295
* combine.c (try_combine): When split an instruction pair, where the
first has a sign_extend src, verify that the src and dest modes match.
From-SVN: r57371
fix subreg optimization bug introduced 2002-05-25
* combine.c (simplify_set): When optimizing a subreg src with a
cc0 dest, use GET_MODE (src) for mask instead of inner_mode.
From-SVN: r57323
* combine.c (try_combine): Handle the case that undobuf.other_insn
has been turned into a return or unconditional jump, by inserting
a BARRIER if necessary.
(simplify_set): Test if a condition code setter has a constant
comparison at compile time, if so convert this insn to a no-op move
and update/simplify the condition code user (undobuf.other_insn).
From-SVN: r56955
* machmode.h (SCALAR_INT_MODE_P): New macro to test for
scaler integer mode (MODE_INT or MODE_PARTIAL_INT).
* explow.c (trunc_int_for_mode): Abort when the mode is not
a scaler integer mode.
* combine.c (expand_compound_operation): Don't expand Vector
or Complex modes into shifts.
(expand_field_assignment): Don't do bitwise arithmatic and
shifts on Vector or Complex modes.
(simplify_comparison): Don't call trunc_int_for_mode
for VOIDmode.
* recog.c (general_operand): Likewise.
(immediate_operand): Likewise.
(nonmemory_operand): Likewise.
Co-Authored-By: Steve Ellcey <sje@cup.hp.com>
From-SVN: r56443
2002-08-01 Toon Moene <toon@moene.indiv.nluug.nl>
* c-common.c (cb_register_builtins): Set
__FINITE_MATH_ONLY__ to 1 if -ffinite-math-only
is given, and to 0 otherwise.
* combine.c (simplify_if_then_else): HONOR_NANS
implies FLOAT_MODE_P.
From-SVN: r55932
2002-07-30 Toon Moene <toon@moene.indiv.nluug.nl>
* flags.h: Declare flag_finite_math_only.
Use it in definition of HONOR_NANS and
HONOR_INFINITIES.
* c-common.c (cb_register_builtins): Emit
__FINITE_MATH_ONLY__ when flag_finite_math_only
is set.
* combine.c (simplify_if_then_else): If
flag_finite_math_only is set, a == b has a
definite value.
* toplev.c: Initialize flag_finite_math_only.
(set_flags_fast_math): Set it on -ffast-math.
(flag_fast_math_set_p): Test it.
doc:
* invoke.texi: Document -ffinite-math-only.
f:
* com.c (ffe_init_options): Set
flag_finite_math_only.
* invoke.texi: Document -fno-finite-math-only.
From-SVN: r55887
* combine.c (struct_undo): Change types of recorded substitutions
to be either "int" or "rtx", instead of "unsigned int" and "rtx".
(do_SUBST_INT): Change types of the substitution from unsigned int
to int, to avoid compilation warning from SUBST_INT's only caller.
(make_extraction): Add cast to avoid compilation warning.
(force_to_mode): Remove cast to avoid compilation warning.
From-SVN: r55221
* combine.c (combine_simplify_rtx): Pass the mode of the
shift count, not the shift operation when trying to simplify
a shift on a SHIFT_COUNT_TRUNCATED target.
From-SVN: r55085
2002-06-05 David S. Miller <davem@redhat.com>
Delete SEQUENCE rtl usage outside of reorg and ssa passes.
* rtl.h (gen_sequence, emit_insns, emit_insns_before,
emit_insns_before_scope, emit_insns_after,
emit_insns_after_scope): Delete declaration.
* ada/misc.c (insert_code_for): Use emit_insn* instead of
emit_insns_foo.
* config/alpha/alpha.c (alpha_set_memflags_1): Abort on SEQUENCE.
(alpha_set_memflags): Fix comment.
(set_frame_related_p): Use get_insns instead of gen_sequence.
* config/alpha/alpha.md (setjmp receiver splitter): Avoid
emitting no insns.
* config/arm/arm.c (arm_finalize_pic): Use get_insns instead of
gen_sequence.
(arm_gen_load_multiple, arm_gen_store_multiple): Likewise.
* config/fr30/fr30.c (fr30_move_double): Likewise.
* config/i386/i386.c (ix86_expand_int_movcc, ix86_expand_movstr):
Likewise.
* config/ia64/ia64.c (spill_restore_mem): Likewise.
* config/ia64/ia64.md (conditional move spliiter): Avoid emitting
no insns.
* config/m32r/m32r.c (gen_split_move_double): Use get_insns
instead of gen_sequence.
* config/mips/mips.c (embedded_pic_fnaddr_reg): Likewise.
(mips_expand_prologue, mips16_gp_pseudo_reg): Likewise.
* config/sh/sh.c (sh_need_epilogue): Likewise.
* config/sparc/sparc.md (current_function_calls_alloca, flat): New
attributes.
(setjmp pattern and split): Use them to avoid splitter which emits
no RTL.
* genattrtab.c (main): Emit include of function.h
* config/stormy16/stormy16.c (xstormy16_split_cbranch): Use
get_insns instead of gen_sequence.
* config/cris/cris.c (cris_split_movdx): Likewise.
* emit-rtl.c (emit_insns*): Kill.
(try_split): Expect insn list instead of SEQUENCE.
(make_jump_insn_raw, make_call_insn_raw): Fix comments.
(emit_*insn*): Reimplement to work with INSN lists and PATTERNs.
Make them abort if a SEQUENCE is given and RTL checking is
enabled.
(emit_*_scope): Don't forget to set scope on final insn.
(gen_sequence): Move from here...
* ssa.c (gen_sequence): To here as private function.
* builtins.c (expand_builtin_apply_args): Use emit_insn_foo, fix
comments.
(expand_builtin_return, expand_builtin_mathfn): Likewise.
(expand_builtin_strlen): Use get_insns instead of gen_sequence.
(expand_builtin_saveregs): Use emit_insn_foo, fix comments.
(expand_builtin_expect_jump): Use get_insns and fix comments.
* calls.c (try_to_integrate): Use emit_insn_foo.
(expand_call, emit_library_call_value_1): Likewise.
* expr.c (emit_queue): Handle insn lists instead of SEQUENCE.
(emit_move_insn_1): Use get_insns instead of gen_sequence.
(expand_expr): Use emit_insn_foo.
* cfgrtl.c (commit_one_edge_insertion): Use emit_insn_foo.
* except.c (build_post_landing_pads): Likewise.
* flow.c (attempt_auto_inc): Likewise.
* stmt.c (expand_fixup, fixup_gotos, expand_nl_handler_label,
expand_nl_goto_receivers, expand_decl_cleanup): Likewise.
* function.c (fixup_var_refs_insn): Use get_insns instead of
gen_sequence.
(fixup_var_refs_1): Likewise and expect insn list from gen_foo.
(fixup_memory_subreg): Use get_insns instead of gen_sequence.
(fixup_stack_1, purge_addressof_1, expand_main_function,
get_arg_pointer_save_area): Likewise.
(optimize_bit_field, instantiate_virtual_regs_1, assign_parms,
expand_function_end): Use emit_insn_foo.
(record_insns, keep_stack_depressed): Work with insn list instead
of SEQUENCE, fix comments.
* ifcvt.c (noce_emit_store_flag, noce_try_store_flag,
noce_try_store_flag_constants, noce_try_store_flag_inc,
noce_try_store_flag_mask, noce_emit_cmove, noce_try_cmove_arith,
noce_try_minmax, noce_try_abs): Use emit_insn_foo.
(noce_process_if_block): Use get_insns instead of gen_sequence.
* optabs.c (add_equal_note): Work with insn list, fix comments.
(expand_binop): Expect insn list from GEN_FCN(), use emit_insn_foo.
(expand_unop, expand_complex_abs, expand_unop_insn,
expand_no_conflict_block): Likewise.
(gen_move_insn): Use get_insns instead of gen_sequence.
(gen_cond_trap): Likewise.
* integrate.c (copy_rtx_and_substitute): Likewise.
(emit_initial_value_sets): Use emit_insn_foo.
* reload1.c (emit_output_reload_insns, emit_reload_insns): Likewise.
(fixup_abnormal_edges): Avoid losing REG_NOTES more intelligently
now that RTL generators give insn lists.
* sibcall.c (replace_call_placeholder): Use emit_insn_foo.
* doloop.c (doloop_modify, doloop_modify_runtime): Use get_insns
instead of gen_sequence.
(doloop_optimize): Work with insn lists instead of SEQUENCE rtl.
* explow.c (emit_stack_save, emit_stack_restore): Use get_insns
instead of gen_sequence.
* loop.c (move_movables, emit_prefetch_instructions,
gen_add_mult, check_dbra_loop, gen_load_of_final_value):
Likewise.
(loop_regs_update): Work with insn list instead of SEQUENCE rtl.
(product_cheap_p): Likewise, and add commentary about RTL wastage
here.
* lcm.c (optimize_mode_switching): Use get_insns instead of
gen_sequence.
* profile.c (gen_edge_profiler): Likewise.
* regmove.c (copy_src_to_dest): Likewise.
* reg-stack.c (compensate_edge): Likewise and fix comment.
* gcse.c (process_insert_insn): Likewise.
(insert_insn_end_bb): Work with insn list instead of SEQUENCE rtl.
* jump.c (delete_prior_computation): Update comment.
* genemit.c (gen_expand, gen_split, main): Use get_insns instead
of gen_sequence, update comments to match.
* recog.c (peephole2_optimize): Work with insn lists instead of
SEQUENCE rtl.
* sched-vis.c (print_pattern): Abort on SEQUENCE.
* unroll.c (unroll_loop, find_splittable_givs, final_giv_value):
Use get_insns instead of gen_sequence.
(copy_loop_body): Likewise and don't emit dummy NOTE.
* genrecog.c: Don't mention SEQUENCE rtl in comments.
* combine.c (try_combine): Expect insn lists from split generator.
* reorg.c (relax_delay_slots): Emit SEQUENCE into insn list by
hand.
From-SVN: r54497
* emit-rtl.c (global_rtl): Update comment.
(const_double_htab, const_double_htab_hash,
const_double_htab_hash, lookup_const_double): New.
(const_int_htab_hash, const_int_htab_eq): Remove const
qualifiers, which cause tons of warnings with RTL checking on.
(gen_rtx_CONST_DOUBLE): Deleted.
(const_double_from_real_value): New function - bears some
resemblance to the former immed_real_const_1.
(immed_double_const): Moved here from varasm.c and
simplified.
(gen_rtx_REG): Make REGNO unsigned to squelch warnings.
(gen_rtx_SUBREG): Use gen_rtx_raw_SUBREG.
(gen_rtx): Use immed_double_const.
(init_emit_once): Initialize the const_double_htab. Use
REAL_VALUE_FROM_INT where possible. Can now use
CONST_DOUBLE_FROM_REAL_VALUE when setting up const_tiny_rtx.
* varasm.c (struct varasm_status): Remove x_const_double_chain.
(const_double_chain, immed_real_const, clear_const_double_mem): Delete.
(immed_double_const, immed_real_const_1): Moved to emit-rtl.c.
(init_varasm_status, mark_varasm_status): Don't touch
x_const_double_chain.
* output.h: Delete prototype for clear_const_double_mem.
* real.h: Make REAL_VALUE_TYPE a macro again. Remove leading
'0' slot from all CONST_DOUBLE_FORMAT definitions. Prototype
const_double_from_real_value, not immed_real_const_1, and use
it to define CONST_DOUBLE_FROM_REAL_VALUE. Define new macro
CONST_DOUBLE_ATOF.
* rtl.h (CONST_DOUBLE_CHAIN): Kill.
(CONST_DOUBLE_LOW, CONST_DOUBLE_HIGH): Adjust.
(gen_rtx_CONST_DOUBLE, immed_real_const): Delete prototypes.
(gen_rtx_REG): Second arg is unsigned.
* gengenrtl.c (special_rtx): Take out CONST_DOUBLE.
(excluded_rtx): New, return true for CONST_DOUBLE.
(genmacro): Write nothing for excluded codes.
* combine.c (combine_simplify_rtx): Use CONST_DOUBLE_FROM_REAL_VALUE.
* expr.c (expand_expr): Likewise.
* ggc-common.c (ggc_mark_rtx_children_1): Don't mark the
CONST_DOUBLE_CHAIN.
* toplev.c (rest_of_compilation): Don't call
clear_const_double_mem.
* config/rs6000/rs6000.c (rs6000_float_const): Delete.
(rs6000_hash_constant): Remove CONST_DOUBLE special case.
(toc_hash_eq): Remove CONST_DOUBLE and LABEL_REF special cases.
* config/rs6000/rs6000-protos.h: Don't prototype rs6000_float_const.
* config/c4x/c4x.md, config/rs6000/rs6000.md: Use CONST_DOUBLE_ATOF.
* config/dsp16xx/dsp16xx.md, config/mips/mips.md,
config/pa/pa.md: Use CONST_DOUBLE_FROM_REAL_VALUE.
* config/sparc/sparc.md, config/sparc/sparc.c: Use immed_double_const.
From-SVN: r53409
* combine.c (simplify_comparison): Avoid narrowing a comparison
with a paradoxical subreg when doing so would drop signficant bits.
Co-Authored-By: Hans-Peter Nilsson <hp@bitrange.com>
From-SVN: r51785
PR optimization/6086
* combine.c (combine_simplify_rtx): If simplify_rtx failed because
of SUBREG of volatile MEM or because the MEM was mode dependent,
return CLOBBER instead of unmodified SUBREG.
From-SVN: r51606
* combine.c (set_nonzero_bits_and_sign_copies): Don't call
nonzero_bits if not needed.
(nonzero_bits) [XOR]: Likewise.
(nonzero_bits) [REG]: Use reg_last_set_nonzero_bits even if
reg_last_set_mode and mode are both MODE_INT, but not equal.
(record_value_for_reg): Compute reg_last_set_nonzero_bits
in nonzero_bits_mode for MODE_INT modes.
From-SVN: r51551