Bernd Schmidt [Sun, 23 Jan 2011 21:11:24 +0000 (21:11 +0000)]
re PR rtl-optimization/47166 (SpecCpu2000 Ammp segfaults for ARM with -O3 -mthumb)
PR rtl-optimization/47166
* reload1.c (emit_reload_insns): Disable the spill_reg_store
mechanism for PRE_MODIFY and POST_MODIFY.
(inc_for_reload): For PRE_MODIFY, return the insn that sets the
reloadreg.
Co-Authored-By: Richard Sandiford <rdsandiford@googlemail.com>
From-SVN: r169144
Ira Rosen [Sun, 23 Jan 2011 08:20:22 +0000 (08:20 +0000)]
costmodel-vect-33.c: Expect vectorization to be not profitable if...
* gcc.dg/vect/costmodel/spu/costmodel-vect-33.c: Expect
vectorization to be not profitable if peeling is used
to realign the memory access.
* gcc.dg/vect/costmodel/spu/costmodel-vect-76a.c:
Increase loop bound to avoid loop unrolling.
* gcc.dg/vect/costmodel/spu/costmodel-vect-76c.c:
Likewise.
Ulrich Weigand [Sat, 22 Jan 2011 21:24:54 +0000 (21:24 +0000)]
re PR middle-end/47401 (Support for must-not-throw regions with SJLJ exceptions broken)
PR middle-end/47401
* except.c (sjlj_assign_call_site_values): Move setting the
crtl->uses_eh_lsda flag to ...
(sjlj_mark_call_sites): ... here.
(sjlj_emit_function_enter): Support NULL dispatch label.
(sjlj_build_landing_pads): In a function with no landing pads
that still has must-not-throw regions, generate code to register
a personality function with empty LSDA.
Thomas Koenig [Sat, 22 Jan 2011 17:30:22 +0000 (17:30 +0000)]
re PR fortran/38536 (ICE with C_LOC in resolve.c due to not properly going through expr->ref)
2011-01-22 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/38536
* resolve.c (gfc_iso_c_func_interface): For C_LOC,
check for array sections followed by component references
which are illegal. Also check for coindexed arguments.
2011-01-22 Thomas Koenig <tkoenig@gcc.gnu.org>
PR fortran/38536
* gfortran.dg/c_loc_tests_16.f90: New test.
Nick Clifton [Sat, 22 Jan 2011 16:05:49 +0000 (16:05 +0000)]
rx.md (UNSPEC_CONST): New.
* config/rx/rx.md (UNSPEC_CONST): New.
(deallocate_and_return): Wrap the amount popped off the stack in
an UNSPEC_CONST in order to stop it being rejected by
-mmax-constant-size.
(pop_and_return): Add a "(return)" rtx.
(call): Drop the immediate operand.
(call_internal): Likewise.
(call_value): Likewise.
(call_value_internal): Likewise.
(sibcall_internal): Likewise.
(sibcall_value_internal): Likewise.
(sibcall): Likewise. Generate an explicit call using
sibcall_internal.
(sibcall_value): Likewise.
(mov<>): FAIL if a constant operand is not legitimate.
(addsi3_unpsec): New pattern.
* config/rx/rx.c (rx_print_operand_address): Handle UNPSEC
CONSTs.
(ok_for_max_constant): New function.
(gen_safe_add): New function.
(rx_expand_prologue): Use gen_safe_add.
(rx_expand_epilogue): Likewise.
(rx_is_legitimate_constant): Use ok_for_max_constant. Handle
UNSPEC CONSTs.
Jeff Law [Sat, 22 Jan 2011 04:09:44 +0000 (21:09 -0700)]
re PR tree-optimization/47053 (ICE: verify_flow_info failed: BB 2 can not throw but has an EH edge with -O -fnon-call-exceptions)
PR tree-optimization/47053
* tree-ssa-dse.c (need_eh_cleanup): New bitmap.
(dse_optimize_stmt): Set the appropriate bit in NEED_EH_CLEANUP when
statements are deleted.
(tree_ssa_dse): Allocate & free NEED_EH_CLEANUP. If NEED_EH_CLEANUP
is nonempty, then purge dead edges and cleanup the CFG.
PR tree-optimization/47053
* g++.dg/pr47053.C: New test.
Tobias Burnus [Fri, 21 Jan 2011 20:33:10 +0000 (21:33 +0100)]
re PR fortran/47394 (Internal compiler error when error count limit is reached)
2011-01-21 Tobias Burnus <burnus@net-b.de>
PR fortran/47394
* error.c (gfc_error_now, gfc_fatal_error, gfc_error_check):
Use defined instead of magic number exit status codes.
* scanner.c (include_line, gfc_new_file): Ditto.
Jakub Jelinek [Fri, 21 Jan 2011 19:35:42 +0000 (20:35 +0100)]
re PR middle-end/45566 (ICE: in convert_to_eh_region_ranges, at except.c:2446 with -freorder-blocks-and-partition -fnon-call-exceptions -fprofile-use)
PR middle-end/45566
* except.c (convert_to_eh_region_ranges): Emit queued no-region
notes from other section in hot/cold partitioning even if
last_action is -3. Increment call_site_base.
Jakub Jelinek [Fri, 21 Jan 2011 19:34:03 +0000 (20:34 +0100)]
re PR rtl-optimization/47366 (ICE: verify_flow_info failed: missing REG_EH_REGION note in the end of bb 3 with -fnon-call-exceptions -fno-tree-ccp -fno-tree-forwprop)
PR rtl-optimization/47366
* fwprop.c (forward_propagate_into): Return bool. If
any changes are made, -fnon-call-exceptions is used and
REG_EH_REGION note is present, call purge_dead_edges
and return true if it purged anything.
(fwprop_addr): Adjust callers, call cleanup_cfg (0) if
any EH edges were purged.
Jeff Law [Fri, 21 Jan 2011 16:49:31 +0000 (09:49 -0700)]
re PR rtl-optimization/41619 (ICE in insert_save (caller-save.c) for SPEC CPU2000's 252.eon)
PR rtl-optimization/41619
* caller-save.c (setup_save_areas): Break out code to determine
which hard regs are live across calls by examining the reload chains
so that it is always used.
Eliminate code which checked REG_N_CALLS_CROSSED.
Alexandre Oliva [Fri, 21 Jan 2011 14:57:33 +0000 (14:57 +0000)]
re PR debug/47106 (-fcompare-debug failure (length) with -fpartial-inlining -flto -fconserve-stack)
PR debug/47106
* cfgexpand.c (account_used_vars_for_block): Only account vars
that are annotated as used.
(estimated_stack_frame_size): Don't set TREE_USED.
* tree-dfa.c (create_var_ann): Mark variable as used.
Anatoly Sokolov [Thu, 20 Jan 2011 17:37:13 +0000 (20:37 +0300)]
xtensa.h (GO_IF_MODE_DEPENDENT_ADDRESS): Remove.
* config/xtensa/xtensa.h (GO_IF_MODE_DEPENDENT_ADDRESS): Remove.
* config/xtensa/xtensa-protos.h (constantpool_address_p): Remove.
* config/xtensa/xtensa.c (TARGET_MODE_DEPENDENT_ADDRESS_P): Define.
(xtensa_mode_dependent_address_p): New function.
(constantpool_address_p): Make static. Change return type to bool.
Change argument type to const_rtx. Use CONST_INT_P predicate.
Jakub Jelinek [Thu, 20 Jan 2011 12:02:33 +0000 (13:02 +0100)]
re PR tree-optimization/46130 (ICE: SIGSEGV in walk_stmt_load_store_addr_ops (gimple.c:4894) with -O2 -fno-tree-dce)
PR tree-optimization/46130
* ipa-split.c (consider_split): If return_bb contains non-virtual
PHIs other than for retval or if split_function would not adjust it,
refuse to split.
* gcc.dg/pr46130-1.c: New test.
* gcc.dg/pr46130-2.c: New test.
Dodji Seketeli [Wed, 19 Jan 2011 22:57:47 +0000 (22:57 +0000)]
re PR c++/47291 (ICE: SIGSEGV in size_binop_loc (fold-const.c:1427) with -g)
Fix PR c++/47291
gcc/
PR c++/47291
* dwarf2out.c (generic_type_p, schedule_generic_params_dies_gen)
(gen_scheduled_generic_parms_dies): New functions.
(gen_struct_or_union_type_die): Schedule template parameters DIEs
generation for the end of CU compilation.
(dwarf2out_finish): Generate template parameters DIEs here.
gcc/testsuite/
PR c++/47291
* g++.dg/debug/dwarf2/template-params-10.C: New test.
Now that we properly track the life of MDR, we can emit
the RETF instruction if MDR has not been modified. This
insn is 3-4 cycles faster since the return address is
already loaded.
There's little reason to greatly complicate things by splitting
the pic_load patterns and using complex rtl to make it work out.
Instead, use the %= marker to generate unique numbers and emit
the entire load_pic sequence at once.
At the same time, collect all references to outgoing_args_size
into mn10300_frame_size, and all computations of register save
area size into mn10300_initial_offset.
For addition and logicals, define an operation-plus-flags update pattern
in preparation for compare elimination. In addition, clean up the way
we compare and validate CC_MODEs. Define NEG in terms of NOT; this is
smaller and allows a non-clobbering destination alternative.
Allow REG+REG and POST_MODIFY addressing for AM33. Fix AM33 base and
index register classes. Remove a bunch of register class combinations
that aren't really useful after this cleanup.
Use the "D" and "A" constraints, and the enabled attribute to
unify all ofthe integer move patterns. Delete the fake double
word move patterns; let the middle-end generate subregs as required.
Unfortunately, this somehow exposes a register pressure problem
with the udivmod pattern. This is properly fixed with subsequent
patches that expose the MDR register.
In the meantime it is highly desirable to to preserve bisect-ability
of the patch series, so disable this pattern for AM30.