Testing subversion commit access

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96948 91177308-0d34-0410-b5e6-96231b3b80d8

Fix -mtune forwarding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96944 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused lowering function LowerJumpTable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96943 91177308-0d34-0410-b5e6-96231b3b80d8

Lower BR_JT on the XCore to a jump into a series of jump instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96942 91177308-0d34-0410-b5e6-96231b3b80d8

tests: Don't make a missing llvm-gcc dir a fatal error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96938 91177308-0d34-0410-b5e6-96231b3b80d8

Switch .bc/.ll Makefile rules to use LLVM{CC,CXX} instead of LLVMG{CC,XX}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96936 91177308-0d34-0410-b5e6-96231b3b80d8

Add LLVM{CC,CXX} make variables, which specify the configured path the LLVM
capable compilers (which could be llvm-gcc or clang).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96935 91177308-0d34-0410-b5e6-96231b3b80d8

Initial configure support for using Clang as the LLVM capable compiler.

Comes in two parts:
 1. Use --with-clang=path/to/clang/compiler to select an installed clang, or
    --with-built-clang to have the makefiles use the clang which will be built
    as the LLVM capable compiler. If neither is given, --with-built-clang will
    be used if the Clang sources are checked out into the standard location
    (tools/clang).

 2. Use --with-llvmcc={llvm-gcc,clang,none} to specify which LLVM capable
    compiler to use. If not given, then llvm-gcc will be used if available,
    otherwise Clang.

Makefile support still to come.

Eric, Doug, Chris, seem reasonable?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96934 91177308-0d34-0410-b5e6-96231b3b80d8

Update mcc16 and the ancient Clang plugin for the 'cmd_line' -> 'command' change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96933 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate CFERuntimeLibDir make variable, this shouldn't be needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96932 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a thinko in the lit.cfg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96931 91177308-0d34-0410-b5e6-96231b3b80d8

Input files with empty suffixes must be passed to linker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96927 91177308-0d34-0410-b5e6-96231b3b80d8

Support -Xlinker et al.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96926 91177308-0d34-0410-b5e6-96231b3b80d8

Temporary disable response files.

They are giving us problems on Mac.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96925 91177308-0d34-0410-b5e6-96231b3b80d8

Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96924 91177308-0d34-0410-b5e6-96231b3b80d8

Trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96923 91177308-0d34-0410-b5e6-96231b3b80d8

Implement order-preserving option forwarding.

Needed to correctly handle things like 'llvmc -framework Foo foo.o -framework
Bar bar.o' - before this commit all '-framework' options would've been grouped
together in the beginning.

Due to our dependence on CommandLine this turned out to be a giant hack; we will
migrate away from CommandLine eventually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96922 91177308-0d34-0410-b5e6-96231b3b80d8

Update the test suite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96921 91177308-0d34-0410-b5e6-96231b3b80d8

Correct option forwarding: initial implementation.

Does not work, but the infrastructure changes are in place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96920 91177308-0d34-0410-b5e6-96231b3b80d8

Precompiled headers: initial support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96919 91177308-0d34-0410-b5e6-96231b3b80d8

New experimental/undocumented feature: 'works_on_empty'.

For now, just enough support to make -filelist work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96918 91177308-0d34-0410-b5e6-96231b3b80d8

Support '-install_name'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96917 91177308-0d34-0410-b5e6-96231b3b80d8

Add a way to enable '-opt=foo' forwarding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96916 91177308-0d34-0410-b5e6-96231b3b80d8

Inline and eliminate LLVMG{CC,XX}WITHPATH.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96913 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ancient Xcode project, replaced by CMake project.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96912 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead LUPGRADE make variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96911 91177308-0d34-0410-b5e6-96231b3b80d8

Kill off unused LLVMGCCLIBEXEC make variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96910 91177308-0d34-0410-b5e6-96231b3b80d8

Kill off LLVMGCCARCH and LLVMGCC_VERSION make variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96909 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate llvmgcc_version testing variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96908 91177308-0d34-0410-b5e6-96231b3b80d8

Kill off LLVMGCC_MAJVERS make variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96907 91177308-0d34-0410-b5e6-96231b3b80d8

Kill unused llvmgccmajvers testing variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96906 91177308-0d34-0410-b5e6-96231b3b80d8

add some #if 0'd out code for checking that named values in
input/output patterns have the same type.  It turns out that
this triggers all the time because we don't infer types
between these boundaries.  Until we do, don't turn this on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96905 91177308-0d34-0410-b5e6-96231b3b80d8

Reject patterns that use a name multiple times in the src or result
of a pattern and where the uses have different types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96904 91177308-0d34-0410-b5e6-96231b3b80d8

disable two patterns that are using non-sensical result pattern types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96903 91177308-0d34-0410-b5e6-96231b3b80d8

remove a confused pattern that is trying to match an address
then use it as an MMX register (!?).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96901 91177308-0d34-0410-b5e6-96231b3b80d8

reject patterns that have dead named arguments in the input pattern
this is tidier and can find bugs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96900 91177308-0d34-0410-b5e6-96231b3b80d8

remove a bunch of dead named arguments in input patterns,
though some look dubious afaict, these are all ok.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96899 91177308-0d34-0410-b5e6-96231b3b80d8

reject patterns that mention a name in the destination pattern
but not in the input.  Previously, this would trigger an abort
late in the isel logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96898 91177308-0d34-0410-b5e6-96231b3b80d8

merge some code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96896 91177308-0d34-0410-b5e6-96231b3b80d8

fix a type mismatch in this pattern, where we were using an i64 imm in a
place where an i32 imm was required, the old isel just got lucky.

This fixes CodeGen/X86/x86-64-and-mask.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96894 91177308-0d34-0410-b5e6-96231b3b80d8

more tidying up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96891 91177308-0d34-0410-b5e6-96231b3b80d8

Update memdep when load PRE inserts a new load, and add some debug output.
I don't have a small testcase for this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96890 91177308-0d34-0410-b5e6-96231b3b80d8

reduce indentation by eliminating 'else after return'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96889 91177308-0d34-0410-b5e6-96231b3b80d8

reapply my cellspu changes with a fix to not break the old isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96885 91177308-0d34-0410-b5e6-96231b3b80d8

Update LangRef to match the code; pointers default to being 64-bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96873 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 96854, 96852, and 96849, unbreaking test/CodeGen/CellSPU/i64ops.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96871 91177308-0d34-0410-b5e6-96231b3b80d8

X86InstrInfoSSE.td declares PINSRW as having type v8i16,
don't alis it in the MMX .td file with a different width,
split into two X86ISD opcodes.  This fixes an x86 testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96859 91177308-0d34-0410-b5e6-96231b3b80d8

Added versions of VCGE, VCGT, VCLE, and VCLT NEON instructions which compare to
(immediate #0) for disassembly only.

A8.6.283, A8.6.285, A8.6.287, A8.6.290

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96856 91177308-0d34-0410-b5e6-96231b3b80d8

fix hte last cellspu failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96854 91177308-0d34-0410-b5e6-96231b3b80d8

hack around more crimes in instruction selection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96852 91177308-0d34-0410-b5e6-96231b3b80d8

Convert this test to FileCheck and add a testcase for PR3574.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96851 91177308-0d34-0410-b5e6-96231b3b80d8

the cell backend is making all sorts of unsafe and incorrect assumptions
about ownership and update policies.  It isn't clear why it is doing all
this lowering at isel time instead of in legalize.  This fixes fcmp64.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96849 91177308-0d34-0410-b5e6-96231b3b80d8

really fix an off-by-one error

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96845 91177308-0d34-0410-b5e6-96231b3b80d8

fix an off-by-one error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96844 91177308-0d34-0410-b5e6-96231b3b80d8

switch the value# in OPC_CompleteMatch and OPC_EmitNode to use a
VBR encoding for the insanity being perpetrated by the spu backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96843 91177308-0d34-0410-b5e6-96231b3b80d8

Added VCEQ (immediate #0) NEON instruction for disassembly only.

A8.6.281

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96838 91177308-0d34-0410-b5e6-96231b3b80d8

Dead code elimination

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96837 91177308-0d34-0410-b5e6-96231b3b80d8

Fix viewCFG on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96834 91177308-0d34-0410-b5e6-96231b3b80d8

add a new Push2 opcode for targets (like cellspu) which have
ridiculously ginormous patterns and need more than one byte
of displacement for encodings.  This fixes CellSPU/fdiv.ll.
SPU is still doing something else ridiculous though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96833 91177308-0d34-0410-b5e6-96231b3b80d8

These should not have been committed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96827 91177308-0d34-0410-b5e6-96231b3b80d8

no need to run llvm-as here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96826 91177308-0d34-0410-b5e6-96231b3b80d8

Instcombine constant folding can normalize gep with negative index to index with large offset. When instcombine objsize checking transformation sees these geps where the offset seemingly point out of bound, it should just return "i don't know" rather than asserting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96825 91177308-0d34-0410-b5e6-96231b3b80d8

Change ComplexPattern handling to push the node being matched as
well as the operands produced when the pattern is matched.  This
allows CheckSame to work correctly when matching replicated
names involving ComplexPatterns.  This fixes a bunch of MSP430
failures, we're down to 13 failures, two of which are
due to a sched bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96824 91177308-0d34-0410-b5e6-96231b3b80d8

Updated version of r96634 (which was reverted due to failing 176.gcc and
126.gcc nightly tests. These failures uncovered latent bugs that machine DCE
could remove one half of a stack adjust down/up pair, causing PEI to assert.
This update fixes that, and the tests now pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96822 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test for canonicalizing ConstantExpr operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96820 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up a bit and fix for when SPAdj != 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96818 91177308-0d34-0410-b5e6-96231b3b80d8

The predicate index isn't fixed, so scan for it to make sure we get the proper
value.

Thumb2 uses the tADJCALLSTACK* instructions, and doesn't need t2 versions, so
remove the FIXME entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96817 91177308-0d34-0410-b5e6-96231b3b80d8

Canonicalize ConstantInts to the right operand of commutative
operators.

The test difference is just due to the multiplication operands
being commuted (and thus requiring a more elaborate match). In
optimized code, that expression would be folded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96816 91177308-0d34-0410-b5e6-96231b3b80d8

expand my hack to work with nodes that have flags but no chains and the
isel doesn't know the correct # results.  This fixes 8 codegen tests,
down to 22 failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96815 91177308-0d34-0410-b5e6-96231b3b80d8

add a new CheckMultiOpcode opcode for checking that a node
has one of the list of acceptable opcodes for a complex
pattern.  This fixes 4 regtest failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96814 91177308-0d34-0410-b5e6-96231b3b80d8

When matching patterns that have a complex pattern as their root, make
sure to only run the complex pattern on nodes where the target opts in.
This patch only handles targets with one opcode specified so far, but
fixes 16 failures, only 34 left.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96813 91177308-0d34-0410-b5e6-96231b3b80d8

add some debug hooks for tracking the behavior of the isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96812 91177308-0d34-0410-b5e6-96231b3b80d8

remove dupes now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96811 91177308-0d34-0410-b5e6-96231b3b80d8

move #includes earlier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96810 91177308-0d34-0410-b5e6-96231b3b80d8

MC/Mach-O: Remove non-sensical comment, and add a missing AddValueSymbols call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96809 91177308-0d34-0410-b5e6-96231b3b80d8

Minor formatting cleanup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96808 91177308-0d34-0410-b5e6-96231b3b80d8

Use Instruction::isCommutative instead of duplicating it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96807 91177308-0d34-0410-b5e6-96231b3b80d8

Added SEL, SXTB16, SXTAB16, UXTAB16, SMMULR, SMMLAR, SMMLSR, SMUAD, and SMUSD,
for disassembly only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96806 91177308-0d34-0410-b5e6-96231b3b80d8

Erase deleted instructions from GVN's ValueTable.  This fixes assertion
failures from ValueTable::verifyRemoved() when using -debug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96805 91177308-0d34-0410-b5e6-96231b3b80d8

Actually enable the -enable-unsafe-fp-math tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96796 91177308-0d34-0410-b5e6-96231b3b80d8

Added a bunch of instructions for disassembly only:

o signed/unsigned add/subtract
o signed/unsigned halving add/subtract
o unsigned sum of absolute difference [and accumulate]
o signed/unsigned saturate
o signed multiply accumulate/subtract [long] dual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96795 91177308-0d34-0410-b5e6-96231b3b80d8

Mark the return address stack slot as mutable when moving the return address
during a tail call. A parameter might overwrite this stack slot during the tail
call.

The sequence during a tail call is:
1.) load return address to temp reg
2.) move parameters (might involve storing to return address stack slot)
3.) store return address to new location from temp reg

If the stack location is marked immutable CodeGen can colocate load (1) with the
store (3).

This fixes bug 6225.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96783 91177308-0d34-0410-b5e6-96231b3b80d8

LLVMC/MultiplePluginPriorities.td: Generally XFAIL this test for now, it is
still failing during (one) llvm-gcc powerpc build, and is also failing on my
x86_64-apple-darwin10.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96781 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused variables and parameters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96780 91177308-0d34-0410-b5e6-96231b3b80d8

Fix various doxygen warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96779 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96778 91177308-0d34-0410-b5e6-96231b3b80d8

Constant-fold certain comparisons with infinity and negative infinity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96777 91177308-0d34-0410-b5e6-96231b3b80d8

Rename a variable to avoid a -Wshadow warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96776 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the logic for reasoning about NaNs from the code that forms
SSE min and max instructions. The real thing this code needs to be
concerned about is negative zero.

Update the sse-minmax.ll test accordingly, and add tests for
-enable-unsafe-fp-math mode as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96775 91177308-0d34-0410-b5e6-96231b3b80d8

When emitting an instruction which depends on both a post-incremented
induction variable value and a loop-variant value, don't force the
insert position to be at the post-increment position, because it may
not be dominated by the loop-variant value. This fixes a
use-before-def problem noticed on PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96774 91177308-0d34-0410-b5e6-96231b3b80d8

This cast<Instruction> is unnecessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96771 91177308-0d34-0410-b5e6-96231b3b80d8

enhance my hack for flags handling, this allows us to pass
CodeGen/X86/pr2182.ll.  Down to 13 x86 failures out of ~1100

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96770 91177308-0d34-0410-b5e6-96231b3b80d8

fix an incorrect VT: eflags is always i32.  The bug was causing us to
create an X86ISD::Cmp node with result type i64 on the
CodeGen/X86/shift-i256.ll testcase and the new isel was assert on it
downstream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96768 91177308-0d34-0410-b5e6-96231b3b80d8

fix most of the failures in the x86 suite by handling multiple
result nodes correctly.  Note that this includes a horrible hack
in DAGISelHeader which cannot be fixed reasonably without
eliminating (parallel) from input patterns.  That, in turn,
can't be done until we support writing multiple result patterns
for the X86and_flag and related multiple-result nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96767 91177308-0d34-0410-b5e6-96231b3b80d8

MC/X86: Add stub AsmBackend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96763 91177308-0d34-0410-b5e6-96231b3b80d8

MC: Sketch registry support for target specific assembler backends.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96762 91177308-0d34-0410-b5e6-96231b3b80d8

Formatting tweaks (trailing whitespace, ordering, comments).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96761 91177308-0d34-0410-b5e6-96231b3b80d8

fix an odd thinko in an assertion, all arm tests pass now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96758 91177308-0d34-0410-b5e6-96231b3b80d8

Always emit register class id references as i32 like
DAGISelEmitter does.  This fixes 11 arm failures (8
left).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96757 91177308-0d34-0410-b5e6-96231b3b80d8

add some no-unwinds, other minor cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96756 91177308-0d34-0410-b5e6-96231b3b80d8