Fix a bug where we were marking GEP expressions with the wrong opcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40085 91177308-0d34-0410-b5e6-96231b3b80d8

When removing instructions from the analysis, be sure to check the confirmed
flag when determining what to do with dependencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40079 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40077 91177308-0d34-0410-b5e6-96231b3b80d8

Because we promote SSE logical ops and loads to v2i64, we often end up generate
code that cross integer / floating point domains (e.g. generate pxor / pand for
logical ops on floating point value, movdqa to load / store floating point SSE
values). Given that, it's better to use movaps instead of movdqa and movups
instead of movdqu. They have the same latency but the "aps" variants are one
byte shorter.
If the domain crossing problem is a real performance issue, then we will have to
fix it with dynamic programming based isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40076 91177308-0d34-0410-b5e6-96231b3b80d8

No need for noResults anymore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40075 91177308-0d34-0410-b5e6-96231b3b80d8

Oops. These stores actually produce results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40074 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40073 91177308-0d34-0410-b5e6-96231b3b80d8

Try fixing it again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40072 91177308-0d34-0410-b5e6-96231b3b80d8

Fix custom lowering of SSE FXOR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40071 91177308-0d34-0410-b5e6-96231b3b80d8

Fix patterns so we isel the xorps, etc. for floating pt logical SSE ops. DAG combiner may fold away the (bit_convert (load)).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40070 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1553:
Change the keywords for the zext and sext parameter attributes to be
zeroext and signext so they don't conflict with the keywords for the
instructions of the same name. This gets around the ambiguity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40069 91177308-0d34-0410-b5e6-96231b3b80d8

Hush a noisy warning from GCC 4.2 about overflow during conversion by using
the type "unsigned" instead of uintptr_t for a 1-bit structure field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40066 91177308-0d34-0410-b5e6-96231b3b80d8

Make val_replace fail early, which reduces the time to optimize 403.gcc to 14.8s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40064 91177308-0d34-0410-b5e6-96231b3b80d8

Don't need the "&&" to glue lines together.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40063 91177308-0d34-0410-b5e6-96231b3b80d8

Verify loop info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40062 91177308-0d34-0410-b5e6-96231b3b80d8

Properly initialize value :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40059 91177308-0d34-0410-b5e6-96231b3b80d8

Replace mysterious code causing a g++-4.2 warning
with hopefully correct code that pleases g++-4.2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40051 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test.  We produce pretty cruddy code for this
nowadays, unfortunately it's not so easy to get good
code here while still producing correct code in exotic
situations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40050 91177308-0d34-0410-b5e6-96231b3b80d8

As pointed out by g++-4.2, the original code didn't do
what it thought it was doing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40044 91177308-0d34-0410-b5e6-96231b3b80d8

Remember to free the heap allocated array if we're not going to use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40043 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVector and DenseMap in even more places.
With this, the time to optimize 403.gcc is down to 15.1s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40042 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR1549

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40041 91177308-0d34-0410-b5e6-96231b3b80d8

Change ValueTable to use a DenseMap for mapping expressions to value numbers.
This results in a slight speedup for 403.gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40040 91177308-0d34-0410-b5e6-96231b3b80d8

Set up ground work to verify preserved analysis info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40039 91177308-0d34-0410-b5e6-96231b3b80d8

Move some sets and maps to SmallPtrSet and DenseMap respectively.  This
reduces the time to optimize 403.gcc from 17.6s to 16.4s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40036 91177308-0d34-0410-b5e6-96231b3b80d8

After a basic block is split into two parts,
second part dominates all the blocks dominated
by original basic block. And first part dominates
second part.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40035 91177308-0d34-0410-b5e6-96231b3b80d8

Now this temp. fix is not required.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40034 91177308-0d34-0410-b5e6-96231b3b80d8

Change instruction description to split OperandList into OutOperandList and
InOperandList. This gives one piece of important information: # of results
produced by an instruction.
An example of the change:
def ADD32rr  : I<0x01, MRMDestReg, (ops GR32:$dst, GR32:$src1, GR32:$src2),
                 "add{l} {$src2, $dst|$dst, $src2}",
                 [(set GR32:$dst, (add GR32:$src1, GR32:$src2))]>;
=>
def ADD32rr  : I<0x01, MRMDestReg, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
                 "add{l} {$src2, $dst|$dst, $src2}",
                 [(set GR32:$dst, (add GR32:$src1, GR32:$src2))]>;

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40033 91177308-0d34-0410-b5e6-96231b3b80d8

Only adjust esp around calls in presence of alloca.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40030 91177308-0d34-0410-b5e6-96231b3b80d8

Only adjust esp around calls in presence of alloca.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40028 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40025 91177308-0d34-0410-b5e6-96231b3b80d8

Fix dominator info update to accommodate CFG changes.
This fixes PR1559.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40024 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40023 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40022 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40020 91177308-0d34-0410-b5e6-96231b3b80d8

Use MOV instead of LEA to restore ESP if callee-saved frame size is 0; if previous instruction updates esp, fold it in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40018 91177308-0d34-0410-b5e6-96231b3b80d8

Add constructor overloads for LoadInst and StoreInst that insert at the
end of a BasicBlock and have an alignment parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40016 91177308-0d34-0410-b5e6-96231b3b80d8

Implement initial memory alignment awareness for SSE instructions. Vector loads
and stores that have a specified alignment of less than 16 bytes now use
instructions that support misaligned memory references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40015 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an issue where assignments that caused a SmallPtrSet to become non-small
would result in calling realloc() on a null pointer.  Instead, if we encounter
this situation, make a normal call to malloc().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40014 91177308-0d34-0410-b5e6-96231b3b80d8

It's not necessary to do rounding for alloca operations when the requested
alignment is equal to the stack alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40004 91177308-0d34-0410-b5e6-96231b3b80d8

New entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39998 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a sys::Path API error caught by a gcc 4.2 warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39995 91177308-0d34-0410-b5e6-96231b3b80d8

Work around a bogus gcc 4.2 warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39993 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test to not rely on the path but to use the
configured llvm-gcc instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39992 91177308-0d34-0410-b5e6-96231b3b80d8

The property is svn:executable not svn:execute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39991 91177308-0d34-0410-b5e6-96231b3b80d8

Alphabetize. Document -mlimit parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39990 91177308-0d34-0410-b5e6-96231b3b80d8

Turn on FastDSE by default.

Note: FastDSE now equals or exceeds the results of old DSE on all of SPEC2000 and SPEC2006. Unless major problems
show up in the testers, it will likely completely replace old DSE in the near future.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39986 91177308-0d34-0410-b5e6-96231b3b80d8

Fold prologue esp update when possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39984 91177308-0d34-0410-b5e6-96231b3b80d8

support poolalloc as checked out from svn

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39983 91177308-0d34-0410-b5e6-96231b3b80d8

Change sroa threshold back.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39980 91177308-0d34-0410-b5e6-96231b3b80d8

Dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39979 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure not to break eh_return.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39978 91177308-0d34-0410-b5e6-96231b3b80d8

Update.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39977 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39976 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use * in XFAIL line unless you want it to XFAIL on everything. I don't believe you can use true regular expressions here. It will just look for those keywords. Please test this on sparc and see if they still are xfailed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39975 91177308-0d34-0410-b5e6-96231b3b80d8

Missed the case where alloca is used but the stack size (not including callee-saved portion) is zero. Thanks Dan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39974 91177308-0d34-0410-b5e6-96231b3b80d8

detect invalid combination of sret and byval

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39971 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL these on sparc-solaris

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39969 91177308-0d34-0410-b5e6-96231b3b80d8

Use push / pop for prologues and epilogues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39967 91177308-0d34-0410-b5e6-96231b3b80d8

Make use of the process_builder_args library function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39965 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1558:
Move tests that have C/C++ sources into the appropriate directory. This
allows them to be selected for testing based on whether llvm-gcc is
present or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39963 91177308-0d34-0410-b5e6-96231b3b80d8

no email addrs in file headers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39962 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the build by putting calls to free into the implementation file and
having that implementation file #include <cstdlib>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39952 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily set SROA threshold to 512.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39950 91177308-0d34-0410-b5e6-96231b3b80d8

Take advantage of undefined behavior if the source program tries to GEP
beyond the end of an alloca to make FastDSE faster and more aggressive.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39945 91177308-0d34-0410-b5e6-96231b3b80d8

Normally, we don't want to install and we only need to build tools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39944 91177308-0d34-0410-b5e6-96231b3b80d8

No, we're not looking for a directory named /config.status.
Sigh .. fix copy & paste error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39939 91177308-0d34-0410-b5e6-96231b3b80d8

In ModuleInfo.txt, make sure that the script being executed can be
found by prepending a .
In build-for-llvm-top.sh, dereference variables for exansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39936 91177308-0d34-0410-b5e6-96231b3b80d8

Prepare for the future when llvm doesn't depend on llvm-gcc-4-0.
This also avoids a fork bomb from the llvm-top/build script which
doesn't know how to handle cycles in the dependence graph.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39933 91177308-0d34-0410-b5e6-96231b3b80d8

Remove insignificant test no longer needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39931 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for walking up memory def chains, which enables finding many more
dead stores on 400.perlbench.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39929 91177308-0d34-0410-b5e6-96231b3b80d8

Use realloc() to (potentially) resize the contents of SmallPtrSet in place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39926 91177308-0d34-0410-b5e6-96231b3b80d8

Return Undef if the block has no dominator. This was required to allow
llvm-gcc build to succeed. Without this change it fails in libstdc++
compilation. This causes no regressions in dejagnu tests. However,
someone who knows this code better might want to review it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39924 91177308-0d34-0410-b5e6-96231b3b80d8

Use ConstantFoldFP for folding all unary floating-point operations which may
have an error, and refector out the code for binary operators into
ConstantFoldBinaryFP and use it for all binary floating-point operations
which may have an error. These functions still rely exclusively on errno
to detect errors though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39923 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the documentation for llvm.sqrt so that sqrt(0) has defined behavior,
which appears to be the intent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39922 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comments about vectors to use the current wording.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39921 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the trailing semicolon from function declarations in LLVM,
where it's interpreted as a comment, not part of the syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39920 91177308-0d34-0410-b5e6-96231b3b80d8

Fix validation errors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39917 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up some formatting

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39916 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up some formatting. Add some doc_code div tags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39915 91177308-0d34-0410-b5e6-96231b3b80d8

Prepare for building from llvm-top. Unfortunately, this doesn't work because
of the cyclic depndency between llvm and llvm-gcc-4-0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39910 91177308-0d34-0410-b5e6-96231b3b80d8

Repair a regression in Transforms/InstCombine/mul.ll that Reid noticed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39896 91177308-0d34-0410-b5e6-96231b3b80d8

rename function to avoid llvm-upgrade warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39895 91177308-0d34-0410-b5e6-96231b3b80d8

Start adding and cleaning up comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39894 91177308-0d34-0410-b5e6-96231b3b80d8

Handle decrementing loops properly. Fixes PR1533.

Always pass the constant as the second parameter to HowManyLessThans.

Remove obsolete "isSigned" parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39893 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39885 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39882 91177308-0d34-0410-b5e6-96231b3b80d8

Implement shift-simplify.ll:test[45].

First teach instcombine that sign bit checks only demand the
sign bit, this allows simplify demanded bits to hack on
expressions better.

Second, teach instcombine that ashr is useless if only the
sign bit is demanded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39880 91177308-0d34-0410-b5e6-96231b3b80d8

Implement shift-simplify.ll:test3, turning:
(X << 31) <s 0  --> (X&1) != 0

This happens dozens of times in the CFE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39879 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify the language. Pointed out by Duncan Sands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39857 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the build. Patch from Holger Schurig.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39856 91177308-0d34-0410-b5e6-96231b3b80d8

Long live the exception handling!

This patch fills the last necessary bits to enable exceptions
handling in LLVM. Currently only on x86-32/linux.

In fact, this patch adds necessary intrinsics (and their lowering) which
represent really weird target-specific gcc builtins used inside unwinder.

After corresponding llvm-gcc patch will land (easy) exceptions should be
more or less workable. However, exceptions handling support should not be
thought as 'finished': I expect many small and not so small glitches
everywhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39855 91177308-0d34-0410-b5e6-96231b3b80d8

Use maximal intersection algorithm exclusively. Fixes miscompile bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39852 91177308-0d34-0410-b5e6-96231b3b80d8

Add alternate ConstantRange intersection algorithm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39851 91177308-0d34-0410-b5e6-96231b3b80d8

Added test case from PR1540.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39845 91177308-0d34-0410-b5e6-96231b3b80d8

Make LCSSA a loop pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39844 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR1540: Specify F0, F1 are sub-registers of D0, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39843 91177308-0d34-0410-b5e6-96231b3b80d8

Handle GEPs with all-zero indices in the same way we handle pointer-pointer bitcasts. Also, fix a potentia infinite loop.

This brings FastDSE to parity with old DSE on 175.vpr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39839 91177308-0d34-0410-b5e6-96231b3b80d8

Disable claims to preserve analysis until open issues are resolved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39834 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate an unused parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@39828 91177308-0d34-0410-b5e6-96231b3b80d8