Christoph Mallon [Wed, 8 Oct 2008 14:06:58 +0000 (14:06 +0000)]
Introduce ia32_Call, which can handle address mode.
[r22609]
Christoph Mallon [Wed, 8 Oct 2008 14:00:38 +0000 (14:00 +0000)]
No need to ignore Projs in the emitter, because there are none in the schedule.
[r22608]
Christoph Mallon [Wed, 8 Oct 2008 13:55:17 +0000 (13:55 +0000)]
Sort.
[r22606]
Michael Beck [Wed, 8 Oct 2008 11:42:56 +0000 (11:42 +0000)]
- BugFix: fix the rare case where two nodes a congruent, but have different modes
This happens for instance with AddP(x, NULL) and x.
Add a Conv than.
[r22603]
Christoph Mallon [Wed, 8 Oct 2008 10:29:13 +0000 (10:29 +0000)]
Add ia32_emitf() to concisely output ia32 instructions.
[r22602]
Christoph Mallon [Wed, 8 Oct 2008 10:12:44 +0000 (10:12 +0000)]
Make ia32_assign_exc_label() static and remove redundant test.
[r22601]
Christoph Mallon [Wed, 8 Oct 2008 09:32:42 +0000 (09:32 +0000)]
Slightly simplify jump table emitter.
[r22600]
Christoph Mallon [Wed, 8 Oct 2008 09:01:46 +0000 (09:01 +0000)]
Do not emit cld in the CopyB prologue. The ABI mandates that DF is cleared, so do not bother performing this NOP.
[r22599]
Christoph Mallon [Wed, 8 Oct 2008 08:54:21 +0000 (08:54 +0000)]
Simplify emit_CopyB_prolog() and remove an unnecessary bit filter before calling it.
[r22598]
Christoph Mallon [Wed, 8 Oct 2008 08:13:35 +0000 (08:13 +0000)]
Make it a bit more clear, what's going on in ia32_emit_x87_binop().
[r22597]
Michael Beck [Wed, 8 Oct 2008 00:23:30 +0000 (00:23 +0000)]
- fixed some comment typos
[r22596]
Michael Beck [Wed, 8 Oct 2008 00:06:47 +0000 (00:06 +0000)]
- Add support for commutative congruences: Add(a,b) and Add(b,a) are
congruent now
- add some explaining comment for the "wrong opcode after
split_by(lambda_opcode) problem
- new checker added
- improved debug output
[r22595]
Matthias Braun [Tue, 7 Oct 2008 19:38:27 +0000 (19:38 +0000)]
and again
[r22594]
Matthias Braun [Tue, 7 Oct 2008 18:51:00 +0000 (18:51 +0000)]
again
[r22593]
Matthias Braun [Tue, 7 Oct 2008 17:32:21 +0000 (17:32 +0000)]
revert change for now breaks 255.vortex and 253.perlbmk
[r22586]
Matthias Braun [Tue, 7 Oct 2008 17:30:17 +0000 (17:30 +0000)]
another test
[r22585]
Matthias Braun [Tue, 7 Oct 2008 16:56:21 +0000 (16:56 +0000)]
another fix for the delayed phi problem, mark memperms as spill+reload
[r22584]
Michael Beck [Tue, 7 Oct 2008 16:38:56 +0000 (16:38 +0000)]
- cond-eval is ok, gvn-pre has a problem with the do loop
[r22583]
Michael Beck [Tue, 7 Oct 2008 16:29:07 +0000 (16:29 +0000)]
- some strange combination of cond-eval and gvn-pre
[r22582]
Michael Beck [Tue, 7 Oct 2008 15:52:45 +0000 (15:52 +0000)]
- do not clean up keep-alives here
- ensure that really all blocks are initialized
[r22581]
Michael Beck [Tue, 7 Oct 2008 14:54:04 +0000 (14:54 +0000)]
- fixed strange code: when did it break?
[r22579]
Michael Beck [Tue, 7 Oct 2008 14:50:58 +0000 (14:50 +0000)]
- kill Bad nodes from the keep-alive list
[r22578]
Michael Beck [Tue, 7 Oct 2008 13:27:03 +0000 (13:27 +0000)]
- used new irg_walk_blkwise_dom_top_down() to calculate the value of each node
as needed in first phase of BuildSets
[r22574]
Michael Beck [Tue, 7 Oct 2008 13:23:57 +0000 (13:23 +0000)]
- we need irg_walk_blkwise_dom_top_down() instead of
irg_walk_blkwise_dom_graph()
[r22573]
Michael Beck [Tue, 7 Oct 2008 13:04:35 +0000 (13:04 +0000)]
- kill Bad nodes from the keep-alive list: these arise from the new
keep-semantics (all memory must be kept in loops)
[r22572]
Michael Beck [Tue, 7 Oct 2008 12:34:36 +0000 (12:34 +0000)]
- re-enabled global-non-null test, but place Confirms only before Cmp nodes
[r22571]
Christoph Mallon [Tue, 7 Oct 2008 09:39:20 +0000 (09:39 +0000)]
Add irn_visited_else_mark(), which combines irn_visited() and mark_irn_visited().
[r22563]
Christoph Mallon [Tue, 7 Oct 2008 09:32:28 +0000 (09:32 +0000)]
fild and fist can handle 16bit source/destination.
[r22562]
Christoph Mallon [Tue, 7 Oct 2008 09:31:24 +0000 (09:31 +0000)]
Use separate code to emit suffixes for integer and floating point instructions, because GAS wants different suffixes for 16bit /integer/ operands ("w" vs. "s") for them.
[r22561]
Matthias Braun [Mon, 6 Oct 2008 20:44:07 +0000 (20:44 +0000)]
fix last fix (+ optimisation: don't do anything if copy and orig are the same)
[r22552]
Michael Beck [Mon, 6 Oct 2008 20:24:08 +0000 (20:24 +0000)]
- reenabled modeb handling after condeval fix
[r22551]
Matthias Braun [Mon, 6 Oct 2008 19:02:12 +0000 (19:02 +0000)]
fix condeval ssa recondstruction
[r22550]
Andreas Zwinkau [Mon, 6 Oct 2008 17:29:52 +0000 (17:29 +0000)]
Adding a smaller variant of fehler044
[r22546]
Michael Beck [Mon, 6 Oct 2008 15:14:39 +0000 (15:14 +0000)]
- implemented irg_walk_blkwise_dom_graph()
[r22541]
Matthias Braun [Mon, 6 Oct 2008 14:43:12 +0000 (14:43 +0000)]
disable femms usage for now
[r22539]
Michael Beck [Mon, 6 Oct 2008 14:40:05 +0000 (14:40 +0000)]
- indentation fixed
[r22538]
Matthias Braun [Mon, 6 Oct 2008 13:47:35 +0000 (13:47 +0000)]
fix typo
[r22535]
Michael Beck [Mon, 6 Oct 2008 13:33:29 +0000 (13:33 +0000)]
- BugFix: fixed sc_print() for modes with bitsize non dividable by 4
[r22534]
Matthias Braun [Mon, 6 Oct 2008 13:22:08 +0000 (13:22 +0000)]
try to never put phis into pending
[r22532]
Matthias Braun [Mon, 6 Oct 2008 13:21:53 +0000 (13:21 +0000)]
more debug info
[r22531]
Christoph Mallon [Mon, 6 Oct 2008 09:36:34 +0000 (09:36 +0000)]
Typo in comment.
[r22522]
Christoph Mallon [Mon, 6 Oct 2008 09:33:18 +0000 (09:33 +0000)]
s/irn_not_visited(x)/!irn_visited(x)/.
[r22521]
Christoph Mallon [Mon, 6 Oct 2008 08:32:09 +0000 (08:32 +0000)]
Remove redundant test.
[r22518]
Michael Beck [Mon, 6 Oct 2008 00:17:27 +0000 (00:17 +0000)]
- Condeval places a Phi in a wrong block, see the AddP
generated for line 12
[r22508]
Christoph Mallon [Sun, 5 Oct 2008 17:34:39 +0000 (17:34 +0000)]
Clean up need_constraint_copy().
[r22503]
Michael Beck [Sun, 5 Oct 2008 15:07:10 +0000 (15:07 +0000)]
- add missing include
[r22499]
Christoph Mallon [Sun, 5 Oct 2008 13:32:31 +0000 (13:32 +0000)]
Good day and welcome to the FIRM XMALLOC*() macros. These macros are provided for the type safety and convenience of the IPD research facility personnel. The time is 15:32. Check xmalloc.h for details.
[r22494]
Christoph Mallon [Sun, 5 Oct 2008 12:19:18 +0000 (12:19 +0000)]
Plug memory leak.
[r22493]
Christoph Mallon [Sun, 5 Oct 2008 11:41:37 +0000 (11:41 +0000)]
Do not bother zeroing an object just before freeing it.
[r22492]
Christoph Mallon [Sun, 5 Oct 2008 06:56:52 +0000 (06:56 +0000)]
Panic, when encountering a to be lowered ASM node. This at least prevents the endless loop in fehler132.
[r22490]
Michael Beck [Sun, 5 Oct 2008 00:45:03 +0000 (00:45 +0000)]
- improved is_non_null_Confirm()
- disabled buggy parts for now
[r22487]
Moritz Kroll [Sun, 5 Oct 2008 00:44:22 +0000 (00:44 +0000)]
Let some optimizations return non-zero, if they changed something (for fixpoint iteration). Updated Visual Studio project.
[r22486]
Michael Beck [Sat, 4 Oct 2008 23:06:40 +0000 (23:06 +0000)]
- BugFix: fixed effect of r22470: Immediate is always the RIGHT operand
(where this is specified?), so do NOT overwrite it (analog to
turn_back_am())
[r22485]
Michael Beck [Sat, 4 Oct 2008 21:38:35 +0000 (21:38 +0000)]
- fixed r22480: Proj numbers are now different
[r22483]
Christoph Mallon [Sat, 4 Oct 2008 18:47:20 +0000 (18:47 +0000)]
Micro optimisation of the day: Remove ia32_Test, which tests the high result of ia32_Mul.
[r22480]
Christoph Mallon [Sat, 4 Oct 2008 17:38:35 +0000 (17:38 +0000)]
peephole_ia32_Test() would due to a bug only remove an ia32_Test if its direct successor was the desired node. If the desired node was an indirect successor, it would incorrectly identify it as flags destroyer.
[r22479]
Christoph Mallon [Sat, 4 Oct 2008 16:57:08 +0000 (16:57 +0000)]
Treat conversions between the same size as downconv to get rid of unnecessary conversions.
[r22477]
Christoph Mallon [Sat, 4 Oct 2008 15:35:14 +0000 (15:35 +0000)]
Use turn_back_am() to implement peephole_ia32_Imul_split().
[r22471]
Christoph Mallon [Sat, 4 Oct 2008 15:15:17 +0000 (15:15 +0000)]
Only deactivate AM support for a node with an immediate, if the match flags do not include match_am_and_immediates. This allows folding reloads into IMul with immediate.
[r22470]
Christoph Mallon [Sat, 4 Oct 2008 13:15:35 +0000 (13:15 +0000)]
Fix fehler157: Lowering Shrs was missing a Conv.
[r22469]
Christoph Mallon [Sat, 4 Oct 2008 12:59:55 +0000 (12:59 +0000)]
fehler157: Wrong mode for lower half of lowered 64bit Mul.
[r22468]
Christoph Mallon [Sat, 4 Oct 2008 12:52:00 +0000 (12:52 +0000)]
Use gen_binop() in gen_Mulh().
[r22467]
Christoph Mallon [Sat, 4 Oct 2008 12:26:31 +0000 (12:26 +0000)]
Ensure that empty functions get their ret $0 by skipping iro_Start and beo_RegParams.
[r22465]
Christoph Mallon [Sat, 4 Oct 2008 11:50:45 +0000 (11:50 +0000)]
Simplify node construction in gen_Mulh().
[r22464]
Christoph Mallon [Sat, 4 Oct 2008 11:43:37 +0000 (11:43 +0000)]
Merge ia32_am_type_t and ia32_am_arity_t, because (type == ia32_am_None) == (arity == ia32_am_arity_none).
[r22463]
Christoph Mallon [Sat, 4 Oct 2008 11:06:30 +0000 (11:06 +0000)]
set_ia32_op_type() want ia32_AddrModeS, not ia32_am_Source (though they happen to be the same value).
[r22461]
Christoph Mallon [Sat, 4 Oct 2008 10:34:00 +0000 (10:34 +0000)]
Make ia32_get_admissible_noreg() static.
[r22459]
Christoph Mallon [Sat, 4 Oct 2008 10:17:17 +0000 (10:17 +0000)]
fehler156: aligning the stack does not work.
[r22458]
Christoph Mallon [Sat, 4 Oct 2008 09:51:49 +0000 (09:51 +0000)]
Enable folding reloads into unary ia32 instructions (only binary was possible).
[r22457]
Christoph Mallon [Sat, 4 Oct 2008 07:59:00 +0000 (07:59 +0000)]
Remove ia32_am_ternary. The only users were Div an IDiv, which are perfectly fine with being declared unary. Further the only place, where the value was tested (in turn_back_am()) set the wrong operand then. Though, effectively this was unreachable code, because Div/IDiv generate no sensible flags.
[r22456]
Christoph Mallon [Sat, 4 Oct 2008 07:43:39 +0000 (07:43 +0000)]
Use ia32_copy_am_attrs() for splitting IMul reg, mem, imm and also reset base and index.
[r22455]
Christoph Mallon [Sat, 4 Oct 2008 07:42:14 +0000 (07:42 +0000)]
Clean up turn_back_am(), panic on unknown arity.
[r22454]
Christoph Mallon [Sat, 4 Oct 2008 07:39:36 +0000 (07:39 +0000)]
Small simplification.
[r22453]
Christoph Mallon [Sat, 4 Oct 2008 07:11:14 +0000 (07:11 +0000)]
Put the flags, which are modified by Inc/Dec, into a variable.
[r22452]
Christoph Mallon [Fri, 3 Oct 2008 21:47:33 +0000 (21:47 +0000)]
Improve upper_bits_clean(): Recurse on Projs and failed Shr.
[r22450]
Matthias Braun [Fri, 3 Oct 2008 18:09:25 +0000 (18:09 +0000)]
convert bitfield initializer tarvals before using them
[r22441]
Matthias Braun [Fri, 3 Oct 2008 17:30:39 +0000 (17:30 +0000)]
fix firm outputting additional .skips after .string
[r22438]
Matthias Braun [Fri, 3 Oct 2008 17:21:17 +0000 (17:21 +0000)]
mac assembler doesn't know .skip, use .space
[r22437]
Christoph Mallon [Thu, 2 Oct 2008 17:32:13 +0000 (17:32 +0000)]
Let is_non_null_Confirm() work as advertised for SymConsts.
[r22417]
Michael Beck [Thu, 2 Oct 2008 14:02:28 +0000 (14:02 +0000)]
- reduce the number of created confirms further
[r22415]
Michael Beck [Thu, 2 Oct 2008 13:55:27 +0000 (13:55 +0000)]
reduce the number of created Confirms
[r22414]
Michael Beck [Thu, 2 Oct 2008 13:50:34 +0000 (13:50 +0000)]
- BugFix: ensure, that no self-loops are constructed (this work-aound might be too restrictive)
- typos fixed
[r22413]
Michael Beck [Thu, 2 Oct 2008 13:34:00 +0000 (13:34 +0000)]
- BugFix: remove_confirms() now does NOT deactivate confirms for all graphs except the first one
- added debug output
- implemnted global NULL pointer test elimination using Confirm magic
[r22412]
Michael Beck [Thu, 2 Oct 2008 13:32:04 +0000 (13:32 +0000)]
add global_null_ptr_elimination flag
[r22411]
Andreas Zwinkau [Thu, 2 Oct 2008 13:09:13 +0000 (13:09 +0000)]
Test case for historical reasons
[r22410]
Michael Beck [Thu, 2 Oct 2008 11:55:23 +0000 (11:55 +0000)]
test for a rather simply gcc optimization
[r22409]
Michael Beck [Thu, 2 Oct 2008 11:18:31 +0000 (11:18 +0000)]
BugFix of r22380 (fixes 176.gcc):
- smaller_mode() had the meaning "can be represented", restored
- renamed new implemnetation into values_in_mode() with the meaning
"can be converted into in back" which is different from smaller_mode()
[r22408]
Sebastian Buchwald [Thu, 2 Oct 2008 11:12:29 +0000 (11:12 +0000)]
Reuse prevents_AM check in PBQP transformer.
[r22407]
Andreas Zwinkau [Wed, 1 Oct 2008 14:25:05 +0000 (14:25 +0000)]
Panic if proj num is unexpected
[r22394]
Matthias Braun [Wed, 1 Oct 2008 14:05:41 +0000 (14:05 +0000)]
fix endless loops
[r22393]
Matthias Braun [Wed, 1 Oct 2008 14:00:49 +0000 (14:00 +0000)]
simplify and speedup belady/loopana a bit
[r22392]
Matthias Braun [Wed, 1 Oct 2008 11:38:47 +0000 (11:38 +0000)]
create an index, as this heavily speeds up some queries
[r22391]
Sebastian Buchwald [Wed, 1 Oct 2008 11:23:05 +0000 (11:23 +0000)]
PBQP now doesn't find a solution, if we modified the costs a little bit.
- removed unused global variables
- added documentation :-)
now the PBQP solver need some new features...
[r22390]
Michael Beck [Tue, 30 Sep 2008 23:43:40 +0000 (23:43 +0000)]
- BugFix: cleanup cannot work with a precalculated step, use antic_in as initial clean nodes
- use avail_out as the closure for phi_translate()
- add missing lookup() in insert_nodes()
[r22387]
Sebastian Buchwald [Tue, 30 Sep 2008 23:04:21 +0000 (23:04 +0000)]
Added users for symconst.
[r22386]
Sebastian Buchwald [Tue, 30 Sep 2008 21:22:55 +0000 (21:22 +0000)]
Don't panic, the train is waiting for you...
[r22385]
Sebastian Buchwald [Tue, 30 Sep 2008 19:31:40 +0000 (19:31 +0000)]
Add users for symconst.
[r22384]
Moritz Kroll [Tue, 30 Sep 2008 19:01:17 +0000 (19:01 +0000)]
Optimize store after stores with partial overwrites with same values
[r22382]
Moritz Kroll [Tue, 30 Sep 2008 18:55:17 +0000 (18:55 +0000)]
Push ConvP through AddI
[r22381]
Moritz Kroll [Tue, 30 Sep 2008 18:54:08 +0000 (18:54 +0000)]
Mallon: Fixed smaller_mode()
[r22380]