Mask directive completed with CalleeSave info
Comments for Mips directives added.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41526 91177308-0d34-0410-b5e6-96231b3b80d8

Added methods to record SPOffsets from LowerFORMAL_ARGUMENTS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41525 91177308-0d34-0410-b5e6-96231b3b80d8

Cut off crazy computation.  This helps PR1622 slightly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41522 91177308-0d34-0410-b5e6-96231b3b80d8

Use simpler test to filter loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41516 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit triples to avoid default behavior that varies by host.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41510 91177308-0d34-0410-b5e6-96231b3b80d8

Update InvokeInst to work like CallInst

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41506 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment about using libc memset/memcpy or generating inline code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41502 91177308-0d34-0410-b5e6-96231b3b80d8

Remove this test as it is too hard to fix after the
latest EH changes, and in any case it is hard to
imagine how the original bug could be reintroduced.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41497 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test: llvm-gcc-4.2 optimizes almost everything away,
resulting in zero matches, while llvm-gcc-4.1 manages to
remove one pointless selector changing the number of matches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41494 91177308-0d34-0410-b5e6-96231b3b80d8

Make DAGCombiner's global alias analysis query more precise in the case
where both pointers have non-zero offsets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41491 91177308-0d34-0410-b5e6-96231b3b80d8

Now that we don't output cleanups by default, the action
offset needs to be adjusted in this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41490 91177308-0d34-0410-b5e6-96231b3b80d8

If the source and destination pointers in an llvm.memmove are known
to not alias each other, it can be translated as an llvm.memcpy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41489 91177308-0d34-0410-b5e6-96231b3b80d8

Change comments to refer to @malloc and @free instead of %malloc and %free.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41488 91177308-0d34-0410-b5e6-96231b3b80d8

There is an impedance matching problem between LLVM and
gcc exception handling: if an exception unwinds through
an invoke, then execution must branch to the invoke's
unwind target.  We previously tried to enforce this by
appending a cleanup action to every selector, however
this does not always work correctly due to an optimization
in the C++ unwinding runtime: if only cleanups would be
run while unwinding an exception, then the program just
terminates without actually executing the cleanups, as
invoke semantics would require.  I was hoping this
wouldn't be a problem, but in fact it turns out to be the
cause of all the remaining failures in the LLVM testsuite
(these also fail with -enable-correct-eh-support, so turning
on -enable-eh didn't make things worse!).  Instead we need
to append a full-blown catch-all to the end of each
selector.  The correct way of doing this depends on the
personality function, i.e. it is language dependent, so
can only be done by gcc.  Thus this patch which generalizes
the eh.selector intrinsic so that it can handle all possible
kinds of action table entries (before it didn't accomodate
cleanups): now 0 indicates a cleanup, and filters have to be
specified using the number of type infos plus one rather than
the number of type infos.  Related gcc patches will cause
Ada to pass a cleanup (0) to force the selector to always
fire, while C++ will use a C++ catch-all (null).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41484 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit keywords and remove spurious trailing semicolons.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41482 91177308-0d34-0410-b5e6-96231b3b80d8

call libc memcpy/memset if array size is bigger then threshold.

Coping 100MB array (after a warmup) shows that glibc 2.6.1 implementation on
x86-64 (core 2) is 30% faster (from 0.270917s to 0.188079s)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41479 91177308-0d34-0410-b5e6-96231b3b80d8

Don't promote volatile loads/stores. This is needed (for example) to handle setjmp/longjmp properly.
This fixes PR1520.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41461 91177308-0d34-0410-b5e6-96231b3b80d8

Don't DSe volatile stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41456 91177308-0d34-0410-b5e6-96231b3b80d8

Revise per review comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41409 91177308-0d34-0410-b5e6-96231b3b80d8

Add APFloat interface to ConstantFPSDNode.  Change
over uses in DAGCombiner.  Fix interfaces to work
with APFloats.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41407 91177308-0d34-0410-b5e6-96231b3b80d8

Move exit condition and exit branch from exiting block into loop header and dominator info. This avoid execution of dead iteration. Loop is already filter in the beginning such that this change is safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41394 91177308-0d34-0410-b5e6-96231b3b80d8

Constant split values needs upper bound and lower bound check, just like any other split value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41389 91177308-0d34-0410-b5e6-96231b3b80d8

Allow target constants to be illegal types.  The target should
know how to handle them.  This fixes
test/CodeGen/Generic/asm-large-immediate.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41388 91177308-0d34-0410-b5e6-96231b3b80d8

While calculating upper loop bound for first loop and lower loop bound for second loop, take care of edge cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41387 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the dag scheduler to handle inline asm nodes with multi-value immediate operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41386 91177308-0d34-0410-b5e6-96231b3b80d8

The personality function might need to be declared as:

    .set  Lset0,___gxx_personality_v0-.
    .long Lset0

on some targets. Make it so!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41385 91177308-0d34-0410-b5e6-96231b3b80d8

rename isOperandValidForConstraint to LowerAsmOperandForConstraint,
changing the interface to allow for future changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41384 91177308-0d34-0410-b5e6-96231b3b80d8

Comment out declaration of operator== (undefined).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41383 91177308-0d34-0410-b5e6-96231b3b80d8

Test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41382 91177308-0d34-0410-b5e6-96231b3b80d8

Add self

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41381 91177308-0d34-0410-b5e6-96231b3b80d8

Rename llvm-gcc4 to llvm-gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41380 91177308-0d34-0410-b5e6-96231b3b80d8

Poison APFloat::operator==.  Replace existing uses with bitwiseIsEqual.
This means backing out the preceding change to Constants.cpp, alas.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41378 91177308-0d34-0410-b5e6-96231b3b80d8

Use APFloat internally for ConstantFPSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41372 91177308-0d34-0410-b5e6-96231b3b80d8

sink clone() down the class hierarchy from CmpInst into ICmpInst/FCmpInst.
This eliminates a conditional on that path, and ensures ICmpInst/FCmpInst
both have an out-of-line virtual method to home the class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41371 91177308-0d34-0410-b5e6-96231b3b80d8

Fix regression that I caused yesterday night while adding logic to select appropriate split condition branch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41365 91177308-0d34-0410-b5e6-96231b3b80d8

Disable EH generation until PPC works 100%.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41360 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41359 91177308-0d34-0410-b5e6-96231b3b80d8

It is not safe to execute split condition's true branch first all the time. If split
condition predicate is GT or GE then execute false branch first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41358 91177308-0d34-0410-b5e6-96231b3b80d8

Reject ICMP_NE as index split condition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41357 91177308-0d34-0410-b5e6-96231b3b80d8

Tightenup loop filter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41356 91177308-0d34-0410-b5e6-96231b3b80d8

Remove incomplete cost analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41354 91177308-0d34-0410-b5e6-96231b3b80d8

Revised per review feedback from previous patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41353 91177308-0d34-0410-b5e6-96231b3b80d8

silence some warnings in an optimized build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41352 91177308-0d34-0410-b5e6-96231b3b80d8

Change internal representation of ConstantFP to use APFloat.
Interface to rest of the compiler unchanged, as yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41348 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this testcase: there are two matches for
llvm.cttz.i64 because of the declaration of
the intrinsic.  Also, emit-llvm is automatic
and doesn't need to be specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41326 91177308-0d34-0410-b5e6-96231b3b80d8

add some notes on really poor codegen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41319 91177308-0d34-0410-b5e6-96231b3b80d8

new example

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41318 91177308-0d34-0410-b5e6-96231b3b80d8

Perform correct codegen for eh_dwarf_cfa intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41316 91177308-0d34-0410-b5e6-96231b3b80d8

default this to radix 10 like the toString(Un)Signed methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41311 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid hiding issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41310 91177308-0d34-0410-b5e6-96231b3b80d8

rename APInt::toString -> toStringUnsigned for symmetry with toStringSigned()

Add an APSInt::toString() method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41309 91177308-0d34-0410-b5e6-96231b3b80d8

Implement review feedback on trampoline documentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41304 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41295 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41292 91177308-0d34-0410-b5e6-96231b3b80d8

update test to check that codegen works with llvm.used in llvm.metadata section

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41289 91177308-0d34-0410-b5e6-96231b3b80d8

Cosmetic change
"True Loop" and "False Loop" naming terminology to refer two loops
after loop cloning is confusing. Instead just use A_Loop and B_Loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41287 91177308-0d34-0410-b5e6-96231b3b80d8

move this check.  ppc outputs .no_dead_strip properly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41286 91177308-0d34-0410-b5e6-96231b3b80d8

Add the PCSymbol for Darwin x86 platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41284 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor loop condition check in a separate function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41282 91177308-0d34-0410-b5e6-96231b3b80d8

Fix thinko.
Starting value of second loop's induction variable can not be lower
then starting value of original loop's induction variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41280 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for llvm.c* intrinsic failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41268 91177308-0d34-0410-b5e6-96231b3b80d8

Test handling of complex constants as lvalues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41266 91177308-0d34-0410-b5e6-96231b3b80d8

Dror GCCLibraries: they are obsolete since 2.0 release. This also reduces number of licenses used in LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41257 91177308-0d34-0410-b5e6-96231b3b80d8

Rename bunch of variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41250 91177308-0d34-0410-b5e6-96231b3b80d8

Preserve LCSSA.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41246 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41230 91177308-0d34-0410-b5e6-96231b3b80d8

No need to hardcode SmallVector size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41228 91177308-0d34-0410-b5e6-96231b3b80d8

InlineAsm asm support for integer registers added

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41225 91177308-0d34-0410-b5e6-96231b3b80d8

Instruction Itinerary attribution fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41224 91177308-0d34-0410-b5e6-96231b3b80d8

Doxygenize comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41222 91177308-0d34-0410-b5e6-96231b3b80d8

Fix potentially N^2 behavior handling arrays with many of the
same value which get RAUW'd.  This speeds up reading the .bc
file in PR1616 from 852s to 0.19s on my G5 with a debug build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41209 91177308-0d34-0410-b5e6-96231b3b80d8

Use only 1 knob to enable exceptions on Darwin :).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41208 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVector instead of std::vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41207 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code, improve a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41205 91177308-0d34-0410-b5e6-96231b3b80d8

s/ExitBlock/ExitingBlock/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41204 91177308-0d34-0410-b5e6-96231b3b80d8

initial checkin of Neil's APFloat work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41203 91177308-0d34-0410-b5e6-96231b3b80d8

clarify precedence, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41199 91177308-0d34-0410-b5e6-96231b3b80d8

add reverse iterators to smallvector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41198 91177308-0d34-0410-b5e6-96231b3b80d8

- Use correct header for SCEV inside LoopPass.cpp
- Move SCEVExpander::expand() out-of-line workarounding possible toolchain bug

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41197 91177308-0d34-0410-b5e6-96231b3b80d8

Replace indunction variable with split value in loop body.
This fixes art miscompile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41195 91177308-0d34-0410-b5e6-96231b3b80d8

Do not split loops rejected by processOneIterationLoop().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41194 91177308-0d34-0410-b5e6-96231b3b80d8

Add FilteredPassNameParser along with PassArgFilter to filter passes
based on their Arg members.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41192 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanups to reduce some spurious differences between different
scheduler implementations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41191 91177308-0d34-0410-b5e6-96231b3b80d8

Add Type::isIntOrIntVector, like Type::isFPOrFPVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41190 91177308-0d34-0410-b5e6-96231b3b80d8

When Intrinsic::getName is constructing names for overloaded intrinsics,
use the ValueType name instead of the llvm type name, to match what the
verifier expects. For integers these are the same, but for floating-point
values the intrinsics use f32/f64 instead of float/double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41189 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test for gcc-4.2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41187 91177308-0d34-0410-b5e6-96231b3b80d8

Partial implementation of calling functions with byval arguments:
 *) The needed information is propagated to the DAG
 *) The X86-64 backend detects it and aborts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41179 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41178 91177308-0d34-0410-b5e6-96231b3b80d8

update the xcode project file, patch by Gordon Henriksen

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41175 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1611 - Visibility should be ignored for a declaration
when a definition's visibility is different.  Likewise, the
visibility of two declarations mismatching is not an error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41174 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code and print visibility for declarations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41173 91177308-0d34-0410-b5e6-96231b3b80d8

add mips target to builder.  I'd appreciate it if someone with
the right version of autoconf could regenerate the configure script.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41172 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, remove assert that wasn't meant to be committed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41170 91177308-0d34-0410-b5e6-96231b3b80d8

Never insert duplicate edges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41169 91177308-0d34-0410-b5e6-96231b3b80d8

Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41168 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41167 91177308-0d34-0410-b5e6-96231b3b80d8

Compute the argument list as lazily as possible.  This ensures that clients
that don't use it don't have to pay the memory cost for the arguments.  This
allows us to avoid creating Argument nodes for many prototypes and for clients
who lazily deserialize code from a bytecode file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41166 91177308-0d34-0410-b5e6-96231b3b80d8

use a nicer accessor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41165 91177308-0d34-0410-b5e6-96231b3b80d8

Test dag xform: Fold C ? 0 : 1 to ~C or zext(~C) or trunc(~C)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41164 91177308-0d34-0410-b5e6-96231b3b80d8

Fold C ? 0 : 1 to ~C or zext(~C) or trunc(~C) depending the types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41163 91177308-0d34-0410-b5e6-96231b3b80d8