oota-llvm.git
15 years agoMark the x86 fp stack registers as "reserved". This tells LiveVariables
Dan Gohman [Thu, 18 Dec 2008 01:05:09 +0000 (01:05 +0000)]
Mark the x86 fp stack registers as "reserved". This tells LiveVariables
and the RegisterScavenger not to expect traditional liveness
techniques are applicable to these registers, since we don't fully
modify the effects of push and pop after stackification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61179 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix the time regression I introduced in 464.h264ref with
Dale Johannesen [Thu, 18 Dec 2008 00:57:22 +0000 (00:57 +0000)]
Fix the time regression I introduced in 464.h264ref with
my last patch to this file.

The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop.  This was extra bad
because register pressure later forced both base and IV into
memory.  Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this.  However,
there were side effects....

It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before.  (This patch does not handle
all the cases where this can happen.)  And when inserting
new code that feeds into a PHI, it's right to put such
code at the original location rather than in the PHI's
immediate predecessor(s) when the original location is outside
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.

Everything above is exercised in
CodeGen/X86/lsr-negative-stride.ll (and ifcvt4 in ARM which is
the same IR).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61178 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoreapply this hunk from Bill's reversion in r61169, it is conservative
Chris Lattner [Thu, 18 Dec 2008 00:51:32 +0000 (00:51 +0000)]
reapply this hunk from Bill's reversion in r61169, it is conservative
and safe and orthogonal from turning off load pre.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61177 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoXFAIL on Linux.
Bill Wendling [Thu, 18 Dec 2008 00:35:21 +0000 (00:35 +0000)]
XFAIL on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61176 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agomake instnamer name unnamed blocks as well as instructions and args.
Chris Lattner [Thu, 18 Dec 2008 00:33:11 +0000 (00:33 +0000)]
make instnamer name unnamed blocks as well as instructions and args.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61175 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoDo not XFAIL.
Bill Wendling [Thu, 18 Dec 2008 00:27:15 +0000 (00:27 +0000)]
Do not XFAIL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61174 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoTemporarily revert r61027. It was causing a bootstrap failure in "release" mode
Bill Wendling [Wed, 17 Dec 2008 23:31:20 +0000 (23:31 +0000)]
Temporarily revert r61027. It was causing a bootstrap failure in "release" mode
with everyone's favorite error messages:

Comparing stages 2 and 3
warning: ./cc1-checksum.o differs
warning: ./cc1plus-checksum.o differs
Bootstrap comparison failure!
./c-decl.o differs
./cp/decl.o differs
./df-core.o differs
./gcc.o differs
./i386.o differs
./stor-layout.o differs
./tree-pretty-print.o differs
./tree.o differs
make[2]: *** [compare] Error 1
make[1]: *** [stage3-bubble] Error 2

See PR3227.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61169 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoXFAIL for now.
Devang Patel [Wed, 17 Dec 2008 22:54:54 +0000 (22:54 +0000)]
XFAIL for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61167 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoXfail these tests for now.
Devang Patel [Wed, 17 Dec 2008 22:53:09 +0000 (22:53 +0000)]
Xfail these tests for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61166 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoToday the front-ends (llvm-gcc and clang) generate multiple llvm.dbg.compile_units...
Devang Patel [Wed, 17 Dec 2008 22:39:29 +0000 (22:39 +0000)]
Today the front-ends (llvm-gcc and clang) generate multiple llvm.dbg.compile_units to identify source file for various debug entities. Each llvm.dbg.compile_unit matches one file on the disk. However, the backend only supports one DW_TAG_compile_unit per .o file. The backend selects first compile_unit from the vector to construct DW_TAG_compile_unit entry, which is not correct in all cases.

First step to resolve this is, record file name and directory directly in debug info for various debug entities.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61164 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoRevert r61158 for now, as it caused some test failures.
Owen Anderson [Wed, 17 Dec 2008 22:17:27 +0000 (22:17 +0000)]
Revert r61158 for now, as it caused some test failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61159 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix miscompilations caused by renumbering, and enable it as part of prealloc splitting.
Owen Anderson [Wed, 17 Dec 2008 22:06:59 +0000 (22:06 +0000)]
Fix miscompilations caused by renumbering, and enable it as part of prealloc splitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61158 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoThis adds some missing functions to the C binding:
Chris Lattner [Wed, 17 Dec 2008 21:39:50 +0000 (21:39 +0000)]
This adds some missing functions to the C binding:
- ability to insert previously created instructions using a builder
- creation of aliases
- creation of inline asm constants

Patch by Zoltan Varga!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61153 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoForgot to revert r61031 when I reverted r61019, r61030, and r61040.
Bill Wendling [Wed, 17 Dec 2008 20:59:57 +0000 (20:59 +0000)]
Forgot to revert r61031 when I reverted r61019, r61030, and r61040.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61150 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoSimplified marking code regions -- no need to use <div> to surround <pre> tags.
Misha Brukman [Wed, 17 Dec 2008 18:11:40 +0000 (18:11 +0000)]
Simplified marking code regions -- no need to use <div> to surround <pre> tags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61143 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoInstead of referring to the license file in the 2.3 release, refer to the
Misha Brukman [Wed, 17 Dec 2008 18:06:53 +0000 (18:06 +0000)]
Instead of referring to the license file in the 2.3 release, refer to the
always-current SVN version.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61142 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdded mention of the RELEASE_24 tag for the 2.4 release.
Misha Brukman [Wed, 17 Dec 2008 16:27:23 +0000 (16:27 +0000)]
Added mention of the RELEASE_24 tag for the 2.4 release.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61138 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix expansion of vsetcc to set the high bit for true instead of 1.
Mon P Wang [Wed, 17 Dec 2008 08:49:47 +0000 (08:49 +0000)]
Fix expansion of vsetcc to set the high bit for true instead of 1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61129 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoinsert some sequence points and preincrement an iterator to avoid
Chris Lattner [Wed, 17 Dec 2008 05:42:08 +0000 (05:42 +0000)]
insert some sequence points and preincrement an iterator to avoid
iterator invalidation problems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61124 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoEnhance heap sra to be substantially more aggressive w.r.t PHI
Chris Lattner [Wed, 17 Dec 2008 05:28:49 +0000 (05:28 +0000)]
Enhance heap sra to be substantially more aggressive w.r.t PHI
nodes.  This allows it to do fairly general phi insertion if a
load from a pointer global wants to be SRAd but the load is used
by (recursive) phi nodes.  This fixes a pessimization on ppc
introduced by Load PRE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61123 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoDouble the amount of memory reserved for SUnits. This is a
Dan Gohman [Wed, 17 Dec 2008 04:30:46 +0000 (04:30 +0000)]
Double the amount of memory reserved for SUnits. This is a
temporary workaround for an obscure bug. When node cloning is
used, it is possible that more SUnits will be created, and
if the SUnits std::vector has to reallocate, it will
invalidate all the graph edges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61122 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoUse getDepth() and getHeight() instead of accessing the
Dan Gohman [Wed, 17 Dec 2008 04:25:52 +0000 (04:25 +0000)]
Use getDepth() and getHeight() instead of accessing the
Depth and Height members directly, as they may not be
current.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61121 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix the comments for getDepth and getHeight.
Dan Gohman [Wed, 17 Dec 2008 04:24:23 +0000 (04:24 +0000)]
Fix the comments for getDepth and getHeight.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61120 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix for PR3225: disable a broken optimization in
Eli Friedman [Wed, 17 Dec 2008 03:35:17 +0000 (03:35 +0000)]
Fix for PR3225: disable a broken optimization in
DAGTypeLegalizer::ExpandShiftWithKnownAmountBit.

In terms of restoring the optimization, the best fix here isn't
obvious... any ideas?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61119 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoUpdate also the generated docs.
Mikhail Glushenkov [Wed, 17 Dec 2008 02:47:30 +0000 (02:47 +0000)]
Update also the generated docs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61118 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoSome enhancements for the 'case' expression.
Mikhail Glushenkov [Wed, 17 Dec 2008 02:47:01 +0000 (02:47 +0000)]
Some enhancements for the 'case' expression.

Add (error) and (empty).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61117 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoClarify that the scale factor from CheckForIVReuse
Dale Johannesen [Tue, 16 Dec 2008 22:16:28 +0000 (22:16 +0000)]
Clarify that the scale factor from CheckForIVReuse
can be negative.  Keep track of whether all uses of
an IV are outside the loop.  Some cosmetics; no
functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61109 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoA new dag combine; several permutations of this
Dale Johannesen [Tue, 16 Dec 2008 22:13:49 +0000 (22:13 +0000)]
A new dag combine; several permutations of this
are there under ADD, this one was missing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61107 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd code to renumber split intervals into new vregs. This is disabled for now until...
Owen Anderson [Tue, 16 Dec 2008 21:35:08 +0000 (21:35 +0000)]
Add code to renumber split intervals into new vregs.  This is disabled for now until I finish working out some iterator invalidation issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61104 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix another crash found by inspection. If we have a PHI node merging
Chris Lattner [Tue, 16 Dec 2008 21:24:51 +0000 (21:24 +0000)]
Fix another crash found by inspection.  If we have a PHI node merging
the load multiple times, make sure the check the uses of the PHI to
ensure they are transformable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61102 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agofix a crash found by inspection.
Chris Lattner [Tue, 16 Dec 2008 21:04:51 +0000 (21:04 +0000)]
fix a crash found by inspection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61101 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd a helper to remove a branch and DCE the condition, and use it
Eli Friedman [Tue, 16 Dec 2008 20:54:32 +0000 (20:54 +0000)]
Add a helper to remove a branch and DCE the condition, and use it
consistently for deleting branches.  In addition to being slightly
more readable, this makes SimplifyCFG a bit better
about cleaning up after itself when it makes conditions unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61100 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoEliminate the loop that walks the critical path. Instead, just track the
Dan Gohman [Tue, 16 Dec 2008 19:27:52 +0000 (19:27 +0000)]
Eliminate the loop that walks the critical path. Instead, just track the
position in the critical path during the main instruction walk.  This
eliminates the need for the CritialAntiDep DenseMap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61096 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoRemove empty test.
Bill Wendling [Tue, 16 Dec 2008 19:07:17 +0000 (19:07 +0000)]
Remove empty test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61095 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoTemporarily revert r61019, r61030, and r61040. These were breaking LLVM Release
Bill Wendling [Tue, 16 Dec 2008 19:06:48 +0000 (19:06 +0000)]
Temporarily revert r61019, r61030, and r61040. These were breaking LLVM Release
builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61094 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoPreserve SourceValue information when lowering produces multiple loads from
Dan Gohman [Tue, 16 Dec 2008 18:25:36 +0000 (18:25 +0000)]
Preserve SourceValue information when lowering produces multiple loads from
different offsets within the same stack slot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61093 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoWe have decided not to support inline asm where an output operand with a matching...
Evan Cheng [Tue, 16 Dec 2008 18:21:39 +0000 (18:21 +0000)]
We have decided not to support inline asm where an output operand with a matching input operand with incompatible type (i.e. either one is a floating point and the other is an integer or the sizes of the types differ). SelectionDAGBuild will catch these and exit with an error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61092 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoCMake: Added DbgInfoPrinter.cpp to lib/Analysis/CMakeFiles.txt.
Oscar Fuentes [Tue, 16 Dec 2008 12:25:04 +0000 (12:25 +0000)]
CMake: Added DbgInfoPrinter.cpp to lib/Analysis/CMakeFiles.txt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61087 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoadd testcase for -print-dbginfo
Torok Edwin [Tue, 16 Dec 2008 10:10:23 +0000 (10:10 +0000)]
add testcase for -print-dbginfo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61086 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd -print-dbginfo pass that prints LLVM IR with comments inserted to show
Torok Edwin [Tue, 16 Dec 2008 09:09:19 +0000 (09:09 +0000)]
Add -print-dbginfo pass that prints LLVM IR with comments inserted to show
which source/line a certain BB/instruction comes from, original variable names,
and original (unmangled) C++ name of functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61085 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd utility functions to search for DbgStopPointInst corresponding to an
Torok Edwin [Tue, 16 Dec 2008 09:07:36 +0000 (09:07 +0000)]
Add utility functions to search for DbgStopPointInst corresponding to an
instruction or BasicBlock, and to search for DbgDeclareInst corresponding to a
variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61084 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agouse different name for parameter to make it clear that we set DIDescriptor::GV
Torok Edwin [Tue, 16 Dec 2008 09:06:01 +0000 (09:06 +0000)]
use different name for parameter to make it clear that we set DIDescriptor::GV

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61083 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoGeneralize support for analyzing loops to include SLE/SGE loop exit conditions
Nick Lewycky [Tue, 16 Dec 2008 08:30:01 +0000 (08:30 +0000)]
Generalize support for analyzing loops to include SLE/SGE loop exit conditions
and support for non-unit strides with signed exit conditions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61082 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoswitch some std::set/std::map to SmallPtrSet/DenseMap.
Chris Lattner [Tue, 16 Dec 2008 07:34:30 +0000 (07:34 +0000)]
switch some std::set/std::map to SmallPtrSet/DenseMap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61081 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agofix PR3217: fully cached queries need to be verified against the
Chris Lattner [Tue, 16 Dec 2008 07:10:09 +0000 (07:10 +0000)]
fix PR3217: fully cached queries need to be verified against the
visited set before they are used.  If used, their blocks need to be
added to the visited set so that subsequent queries don't use conflicting
pointer values in the cache result blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61080 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoEnable anti-dependence breaking by default when post-RA scheduling is enabled.
Dan Gohman [Tue, 16 Dec 2008 06:21:45 +0000 (06:21 +0000)]
Enable anti-dependence breaking by default when post-RA scheduling is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61078 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoWhen breaking an anti-dependency, don't use a register which has seen
Dan Gohman [Tue, 16 Dec 2008 06:20:58 +0000 (06:20 +0000)]
When breaking an anti-dependency, don't use a register which has seen
one of its aliases defined. This is conservative, but tricky subreg
corner cases are outside the primary aim of this pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61077 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd initial support for back-scheduling address computations,
Dan Gohman [Tue, 16 Dec 2008 03:35:01 +0000 (03:35 +0000)]
Add initial support for back-scheduling address computations,
especially in the case of addresses computed from loop induction
variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61075 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoRemove some special-case logic in ScheduleDAGSDNodes's
Dan Gohman [Tue, 16 Dec 2008 03:31:11 +0000 (03:31 +0000)]
Remove some special-case logic in ScheduleDAGSDNodes's
latency computation code that is no longer needed with the
new method for handling latencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61074 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix some register-alias-related bugs in the post-RA scheduler liveness
Dan Gohman [Tue, 16 Dec 2008 03:25:46 +0000 (03:25 +0000)]
Fix some register-alias-related bugs in the post-RA scheduler liveness
computation code. Also, avoid adding output-depenency edges when both
defs are dead, which frequently happens with EFLAGS defs.

Compute Depth and Height lazily, and always in terms of edge latency
values. For the schedulers that don't care about latency, edge latencies
are set to 1.

Eliminate Cycle and CycleBound, and LatencyPriorityQueue's Latencies array.
These are all subsumed by the Depth and Height fields.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61073 91177308-0d34-0410-b5e6-96231b3b80d8

15 years ago* Comply with HTML 4.01 Strict
Misha Brukman [Tue, 16 Dec 2008 03:07:49 +0000 (03:07 +0000)]
* Comply with HTML 4.01 Strict
* Added code-region markings to code sections to make them stand out
* Added pre.doc_code class to llvm.css to simplify marking code regions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61072 91177308-0d34-0410-b5e6-96231b3b80d8

15 years ago* Comply with HTML 4.01 Strict standard
Misha Brukman [Tue, 16 Dec 2008 02:54:22 +0000 (02:54 +0000)]
* Comply with HTML 4.01 Strict standard
* Converted absolute links to llvm.org/docs to relative links
* Fixed spelling and s/;/:/, as needed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61071 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd a simple target-independent heuristic to allow targets with no
Dan Gohman [Tue, 16 Dec 2008 02:38:22 +0000 (02:38 +0000)]
Add a simple target-independent heuristic to allow targets with no
instruction itinerary data to back-schedule loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61070 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoMove addPred and removePred out-of-line.
Dan Gohman [Tue, 16 Dec 2008 01:05:52 +0000 (01:05 +0000)]
Move addPred and removePred out-of-line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61067 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoMake addPred and removePred return void, since the return value is not
Dan Gohman [Tue, 16 Dec 2008 01:00:55 +0000 (01:00 +0000)]
Make addPred and removePred return void, since the return value is not
currently used by anything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61066 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoThis getEdgeAttributes doesn't need a template argument.
Dan Gohman [Tue, 16 Dec 2008 00:55:00 +0000 (00:55 +0000)]
This getEdgeAttributes doesn't need a template argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61065 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd a getSDep() access function to SUnitIterator to return the
Dan Gohman [Tue, 16 Dec 2008 00:51:33 +0000 (00:51 +0000)]
Add a getSDep() access function to SUnitIterator to return the
current SDep.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61064 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoReorder some SDep methods. No functionality change.
Dan Gohman [Tue, 16 Dec 2008 00:48:53 +0000 (00:48 +0000)]
Reorder some SDep methods. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61063 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoSDep's operator== should compare the Latency field too.
Dan Gohman [Tue, 16 Dec 2008 00:47:19 +0000 (00:47 +0000)]
SDep's operator== should compare the Latency field too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61062 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoChange so that buildit won't need a parent directory with only the build scripts...
Bill Wendling [Tue, 16 Dec 2008 00:42:25 +0000 (00:42 +0000)]
Change so that buildit won't need a parent directory with only the build scripts in them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61061 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoadd testcase for r61051
Chris Lattner [Mon, 15 Dec 2008 21:46:23 +0000 (21:46 +0000)]
add  testcase for r61051

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61052 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoenhance heap-sra to apply to fixed sized array allocations, not just
Chris Lattner [Mon, 15 Dec 2008 21:44:34 +0000 (21:44 +0000)]
enhance heap-sra to apply to fixed sized array allocations, not just
variable sized array allocations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61051 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdded support for splitting and scalarizing vector shifts.
Mon P Wang [Mon, 15 Dec 2008 21:44:00 +0000 (21:44 +0000)]
Added support for splitting and scalarizing vector shifts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61050 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoUse stripPointerCasts.
Chris Lattner [Mon, 15 Dec 2008 21:20:32 +0000 (21:20 +0000)]
Use stripPointerCasts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61047 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agominor tweaks for formatting, allow bitcast in ValueIsOnlyUsedLocallyOrStoredToOneGlobal.
Chris Lattner [Mon, 15 Dec 2008 21:08:54 +0000 (21:08 +0000)]
minor tweaks for formatting, allow bitcast in ValueIsOnlyUsedLocallyOrStoredToOneGlobal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61046 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agorefactor some code into a new TryToOptimizeStoreOfMallocToGlobal function.
Chris Lattner [Mon, 15 Dec 2008 21:02:25 +0000 (21:02 +0000)]
refactor some code into a new TryToOptimizeStoreOfMallocToGlobal function.
Use GetElementPtrInst::hasAllZeroIndices where possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61045 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoadd a basic test for heap-sra
Chris Lattner [Mon, 15 Dec 2008 19:42:05 +0000 (19:42 +0000)]
add a basic test for heap-sra

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61041 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoTeach basicaa to use the nocapture attribute when possible. When the
Chris Lattner [Mon, 15 Dec 2008 18:59:22 +0000 (18:59 +0000)]
Teach basicaa to use the nocapture attribute when possible.  When the
intrinsics are properly marked nocapture, the fixme should be addressed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61040 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix printing of PseudoSourceValues in SDNode graphs.
Dan Gohman [Mon, 15 Dec 2008 17:28:10 +0000 (17:28 +0000)]
Fix printing of PseudoSourceValues in SDNode graphs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61036 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix a typo in a comment.
Dan Gohman [Mon, 15 Dec 2008 17:26:50 +0000 (17:26 +0000)]
Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61035 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoadd some more notes.
Chris Lattner [Mon, 15 Dec 2008 08:32:28 +0000 (08:32 +0000)]
add some more notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61033 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd a testcase for GCC PR 23455, which lpre handles now. Add some
Chris Lattner [Mon, 15 Dec 2008 07:49:24 +0000 (07:49 +0000)]
Add a testcase for GCC PR 23455, which lpre handles now.  Add some
comments about why we're not getting other cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61032 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoUpdate generated files after nocapture syntax change.
Nick Lewycky [Mon, 15 Dec 2008 07:31:07 +0000 (07:31 +0000)]
Update generated files after nocapture syntax change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61031 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoIt turns out that "align 1" and unaligned are different. Add a bias to the
Nick Lewycky [Mon, 15 Dec 2008 07:29:55 +0000 (07:29 +0000)]
It turns out that "align 1" and unaligned are different. Add a bias to the
alignment attribute such that 0 means unaligned.

This will probably require a rebuild of llvm-gcc because of the change to
Attributes.h. If you see many test failures on "make check", please rebuild
your llvm-gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61030 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdded support to LegalizeType for expanding the operands of scalar to vector
Mon P Wang [Mon, 15 Dec 2008 06:57:02 +0000 (06:57 +0000)]
Added support to LegalizeType for expanding the operands of scalar to vector
and insert vector element.  Modified extract vector element to extend the
result to match the expected promoted type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61029 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agogvn now hoists this load out of the hot non-call path.
Chris Lattner [Mon, 15 Dec 2008 06:34:48 +0000 (06:34 +0000)]
gvn now hoists this load out of the hot non-call path.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61028 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoEnable Load PRE. This teaches GVN to push partially redundant loads up the
Chris Lattner [Mon, 15 Dec 2008 05:28:29 +0000 (05:28 +0000)]
Enable Load PRE.  This teaches GVN to push partially redundant loads up the
CFG when there is exactly one predecessor where the load is not available.
This is designed to not increase code size but still eliminate partially
redundant loads.  This fires 1765 times on 403.gcc even though it doesn't
do critical edge splitting yet (the most common reason for it to fail).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61027 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoif we have a phi translation failure of the start block,
Chris Lattner [Mon, 15 Dec 2008 04:58:29 +0000 (04:58 +0000)]
if we have a phi translation failure of the start block,
return *just* a clobber of the start block, not other
random stuff as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61026 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdjust testcase to make it more stable across visitation order changes,
Chris Lattner [Mon, 15 Dec 2008 04:42:00 +0000 (04:42 +0000)]
Adjust testcase to make it more stable across visitation order changes,
unbreaking it after r61024.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61025 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoIfdef out some code that I didn't mean to enable by default yet.
Owen Anderson [Mon, 15 Dec 2008 03:52:17 +0000 (03:52 +0000)]
Ifdef out some code that I didn't mean to enable by default yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61024 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agomake GVN try to rename inputs to the resultant replaced values, which
Chris Lattner [Mon, 15 Dec 2008 03:46:38 +0000 (03:46 +0000)]
make GVN try to rename inputs to the resultant replaced values, which
cleans up the generated code a bit.  This should have the added benefit of
not randomly renaming functions/globals like my previous patch did. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61023 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoImplement initial support for PHI translation in memdep. This means that
Chris Lattner [Mon, 15 Dec 2008 03:35:32 +0000 (03:35 +0000)]
Implement initial support for PHI translation in memdep.  This means that
memdep keeps track of how PHIs affect the pointer in dep queries, which
allows it to eliminate the load in cases like rle-phi-translate.ll, which
basically end up being:

BB1:
   X = load P
   br BB3
BB2:
   Y = load Q
   br BB3
BB3:
   R = phi [P] [Q]
   load R

turning "load R" into a phi of X/Y.  In addition to additional exposed
opportunities, this makes memdep safe in many cases that it wasn't before
(which is required for load PRE) and also makes it substantially more
efficient.  For example, consider:

bb1:  // has many predecessors.
   P = some_operator()
   load P

In this example, previously memdep would scan all the predecessors of BB1
to see if they had something that would mustalias P.  In some cases (e.g.
test/Transforms/GVN/rle-must-alias.ll) it would actually find them and end
up eliminating something.  In many other cases though, it would scan and not
find anything useful.  MemDep now stops at a block if the pointer is defined
in that block and cannot be phi translated to predecessors.  This causes it
to miss the (rare) cases like rle-must-alias.ll, but makes it faster by not
scanning tons of stuff that is unlikely to be useful.  For example, this
speeds up GVN as a whole from 3.928s to 2.448s (60%)!.  IMO, scalar GVN
should be enhanced to simplify the rle-must-alias pointer base anyway, which
would allow the loads to be eliminated.

In the future, this should be enhanced to phi translate through geps and
bitcasts as well (as indicated by FIXMEs) making memdep even more powerful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61022 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoAdd support for slow-path GVN with full phi construction for scalars. This is disabl...
Owen Anderson [Mon, 15 Dec 2008 02:03:00 +0000 (02:03 +0000)]
Add support for slow-path GVN with full phi construction for scalars.  This is disabled for now, as it actually pessimizes code in the abscence
of phi translation for load elimination.  This slow down GVN a bit, by about 2% on 403.gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61021 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFix whitespace in comment.
Nick Lewycky [Mon, 15 Dec 2008 01:35:36 +0000 (01:35 +0000)]
Fix whitespace in comment.
Remove TODO; icmp isn't a binary operator, so this function will never deal
with them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61020 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoIntroducing nocapture, a parameter attribute for pointers to indicate that the
Nick Lewycky [Mon, 15 Dec 2008 01:34:58 +0000 (01:34 +0000)]
Introducing nocapture, a parameter attribute for pointers to indicate that the
callee will not introduce any new aliases of that pointer.

The attributes had all bits allocated already, so I decided to collapse
alignment. Alignment was previously stored as a 16-bit integer from bits 16 to
32 of the attribute, but it was required to be a power of 2. Now it's stored in
log2 encoded form in five bits from 16 to 21. That gives us 11 more bits of
space.

You may have already noticed that you only need four bits to encode a 16-bit
power of two, so why five bits? Because the AsmParser accepted 32-bit
alignments, even though we couldn't store them (they were silently discarded).
Now we can store them in memory, but not in the bitcode.

The bitcode format was already storing these as 64-bit VBR integers. So, the
bitcode format stays the same, keeping the alignment values stored as 16 bit
raw values. There's some hideous code in the reader and writer that deals with
this, waiting to be ripped out the moment we run out of bits again and have to
replace the parameter attributes table encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61019 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agosilence warning when asserts disabled.
Chris Lattner [Sun, 14 Dec 2008 21:38:24 +0000 (21:38 +0000)]
silence warning when asserts disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61014 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agosilence warning when asserts disabled.
Chris Lattner [Sun, 14 Dec 2008 21:37:33 +0000 (21:37 +0000)]
silence warning when asserts disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61013 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoeliminate warning when asserts disabled.
Chris Lattner [Sun, 14 Dec 2008 21:36:23 +0000 (21:36 +0000)]
eliminate warning when asserts disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61012 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoanother random testcase that shouldn't crash gvn and is
Chris Lattner [Sun, 14 Dec 2008 21:20:46 +0000 (21:20 +0000)]
another random testcase that shouldn't crash gvn and is
good for coverage with future changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61011 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoClarify the meaning of the NoAlias response. The plan is to refer to this from
Nick Lewycky [Sun, 14 Dec 2008 21:08:48 +0000 (21:08 +0000)]
Clarify the meaning of the NoAlias response. The plan is to refer to this from
a future version of LangRef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61010 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoGeneralize GVN's phi construciton routine to work for things other than loads.
Owen Anderson [Sun, 14 Dec 2008 19:10:35 +0000 (19:10 +0000)]
Generalize GVN's phi construciton routine to work for things other than loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61009 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoReapply r60997, this time without forgetting that
Duncan Sands [Sun, 14 Dec 2008 09:43:15 +0000 (09:43 +0000)]
Reapply r60997, this time without forgetting that
target constants are allowed to have an illegal
type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61006 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoFixed HTML validator warning.
Misha Brukman [Sun, 14 Dec 2008 08:01:51 +0000 (08:01 +0000)]
Fixed HTML validator warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61005 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoCSS cleanup:
Misha Brukman [Sun, 14 Dec 2008 07:20:36 +0000 (07:20 +0000)]
CSS cleanup:
* Lowercased all HTML element names
* Standardized spacing around { and }
* removed class "doc_table_nw": grep finds no uses

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61004 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoTemporarily revert r60997. It was causing this failure:
Bill Wendling [Sat, 13 Dec 2008 23:53:00 +0000 (23:53 +0000)]
Temporarily revert r60997. It was causing this failure:

Running /Users/void/llvm/llvm.src/test/CodeGen/Generic/dg.exp ...
FAIL: /Users/void/llvm/llvm.src/test/CodeGen/Generic/asm-large-immediate.ll
Failed with exit(1) at line 1
while running:  llvm-as < /Users/void/llvm/llvm.src/test/CodeGen/Generic/asm-large-immediate.ll |  llc | /usr/bin/grep 68719476738
Assertion failed: ((TypesNeedLegalizing || getTypeAction(VT) == Legal) && "Illegal type introduced after type legalization?"), function HandleOp, file /Users/void/llvm/llvm.src/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp, line 493.
0   llc               0x0085392e char const* std::find<char const*, char>(char const*, char const*, char const&) + 98
1   llc               0x00853e63 llvm::sys::PrintStackTraceOnErrorSignal() + 593
2   libSystem.B.dylib 0x96cac09b _sigtramp + 43
3   libSystem.B.dylib 0xffffffff _sigtramp + 1765097359
4   libSystem.B.dylib 0x96d24ec2 raise + 26
5   libSystem.B.dylib 0x96d3447f abort + 73
6   libSystem.B.dylib 0x96d26063 __assert_rtn + 101
7   llc               0x004f9018 llvm::cast_retty<llvm::SubprogramDesc, llvm::DebugInfoDesc*>::ret_type llvm::cast<llvm::Sub
...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61001 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoLegalizeDAG is not supposed to introduce illegal
Duncan Sands [Sat, 13 Dec 2008 22:33:38 +0000 (22:33 +0000)]
LegalizeDAG is not supposed to introduce illegal
types into the DAG if they were not already there.
Check this with an assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60997 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoRLE isn't smart enough to eliminate this safely yet.
Chris Lattner [Sat, 13 Dec 2008 21:04:20 +0000 (21:04 +0000)]
RLE isn't smart enough to eliminate this safely yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60994 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agorename some tests to be more uniform in naming convention.
Chris Lattner [Sat, 13 Dec 2008 18:47:40 +0000 (18:47 +0000)]
rename some tests to be more uniform in naming convention.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60988 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agogvn should never crash on this.
Chris Lattner [Sat, 13 Dec 2008 18:39:44 +0000 (18:39 +0000)]
gvn should never crash on this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60987 91177308-0d34-0410-b5e6-96231b3b80d8

15 years agoThese messages should always be emitted when NDEBUG is unset, not when
Chris Lattner [Sat, 13 Dec 2008 18:37:58 +0000 (18:37 +0000)]
These messages should always be emitted when NDEBUG is unset, not when
NDEBUG is unset and -debug is passed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60986 91177308-0d34-0410-b5e6-96231b3b80d8