A few small clean-ups, and the addition of an LCSSA statistic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28512 91177308-0d34-0410-b5e6-96231b3b80d8

Make CALL node consistent with RET node. Signness of value has type MVT::i32
instead of MVT::i1. Either is fine except MVT::i32 is probably a legal type
for most (if not all) platforms while MVT::i1 is not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28511 91177308-0d34-0410-b5e6-96231b3b80d8

Change RET node to include signness information of the return values. i.e.
RET chain, value1, sign1, value2, sign2, ...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28510 91177308-0d34-0410-b5e6-96231b3b80d8

Change RET node to include signness information of the return values. e.g.
RET chain, value1, sign1, value2, sign2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28509 91177308-0d34-0410-b5e6-96231b3b80d8

Trivial testcase that LCSSA can already handle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28508 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a copy-and-paste-o that would break some compilers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28507 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up and refactor LCSSA a bunch.  It should also run faster now, though
there's still a lot of work to be done on it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28506 91177308-0d34-0410-b5e6-96231b3b80d8

Vector argument must be passed in memory location aligned on 16-byte boundary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28505 91177308-0d34-0410-b5e6-96231b3b80d8

Mac OS X ABI document lied. The first four XMM registers are used to pass
vector arguments, not three.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28504 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/InstCombine/store.ll:test2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28503 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for trivial DSE

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28502 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in the bc reader/writer: we were not correctly encoding varargs
nonccc calls (we were dropping the CC and tail flag).  This broke several
FORTRAN programs.

Testcase here: Regression/Assembler/2006-05-26-VarargsCallEncode.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28501 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, check that the bc file correctly encodes varargs nonccc calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28500 91177308-0d34-0410-b5e6-96231b3b80d8

Minor update to make the code more clear

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28499 91177308-0d34-0410-b5e6-96231b3b80d8

Update more comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28498 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28497 91177308-0d34-0410-b5e6-96231b3b80d8

No need to handle illegal types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28496 91177308-0d34-0410-b5e6-96231b3b80d8

Skeletal LCSSA pass.  This is currently non-functional.  Expect functionality
and documentation updates soo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28495 91177308-0d34-0410-b5e6-96231b3b80d8

On ARM, alignment is in bits
Add lr as a hard coded operand of bx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28494 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a couple of bogus casts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28493 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a bogus cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28492 91177308-0d34-0410-b5e6-96231b3b80d8

Minor bug caught by Ashwin Chandra

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28491 91177308-0d34-0410-b5e6-96231b3b80d8

Transform things like (splat(splat)) -> splat

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28490 91177308-0d34-0410-b5e6-96231b3b80d8

Introduce a helper function that simplifies interpretation of shuffle masks.
No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28489 91177308-0d34-0410-b5e6-96231b3b80d8

Consistency

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28488 91177308-0d34-0410-b5e6-96231b3b80d8

Turn (cast (shuffle (cast)) -> shuffle (cast) if it reduces the # casts in
the program.  This exposes more opportunities for the instcombiner, and implements
vec_shuffle.ll:test6

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28487 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28486 91177308-0d34-0410-b5e6-96231b3b80d8

extract element from a shuffle vector can be trivially turned into an
extractelement from the SV's source.  This implement vec_shuffle.ll:test[45]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28485 91177308-0d34-0410-b5e6-96231b3b80d8

new testcases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28484 91177308-0d34-0410-b5e6-96231b3b80d8

Some clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28483 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for the missing FP condition codes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28482 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28481 91177308-0d34-0410-b5e6-96231b3b80d8

X86 / Cygwin asm / alignment fixes.
Patch contributed by Anton Korobeynikov!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28480 91177308-0d34-0410-b5e6-96231b3b80d8

The patch corresponding to this test had to be reverted as unsafe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28479 91177308-0d34-0410-b5e6-96231b3b80d8

Revert a patch that is unsafe, due to out of range array accesses in inner
array scopes possibly accessing valid memory in outer subscripts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28478 91177308-0d34-0410-b5e6-96231b3b80d8

New test case. x86 isel was creating a cycle in the DAG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28477 91177308-0d34-0410-b5e6-96231b3b80d8

Can't trust NodeDepth when checking for possibility of load folding creating
a cycle. This increase the search space and will increase compile time (in
practice it appears to be small, e.g. 176.gcc goes from 62 sec to 65 sec)
that will be addressed later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28476 91177308-0d34-0410-b5e6-96231b3b80d8

Build breakage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28475 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build failure of povray

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28473 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Benchmarks/MallocBench/cfrac

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28471 91177308-0d34-0410-b5e6-96231b3b80d8

implement initial version of ARMAsmPrinter::printOperand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28470 91177308-0d34-0410-b5e6-96231b3b80d8

port the ARM backend to use ISD::CALL instead of LowerCallTo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28469 91177308-0d34-0410-b5e6-96231b3b80d8

create test/Regression/CodeGen/ARM/ and add a minimal test to it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28468 91177308-0d34-0410-b5e6-96231b3b80d8

-sched-commute-nodes is now on by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28467 91177308-0d34-0410-b5e6-96231b3b80d8

Turn on -sched-commute-nodes by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28465 91177308-0d34-0410-b5e6-96231b3b80d8

Fix breakage on platforms where string/cassert don't pull in int64_t.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28464 91177308-0d34-0410-b5e6-96231b3b80d8

Switch X86 over to a call-selection model where the lowering code creates
the copyto/fromregs instead of making the X86ISD::CALL selection code create
them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28463 91177308-0d34-0410-b5e6-96231b3b80d8

CALL node change (arg / sign pairs instead of just arguments).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28462 91177308-0d34-0410-b5e6-96231b3b80d8

CALL node change: now including signness of every argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28461 91177308-0d34-0410-b5e6-96231b3b80d8

CALL node change: now containing signness of each argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28460 91177308-0d34-0410-b5e6-96231b3b80d8

Assert if InflightSet is not cleared after instruction selecting a BB.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28459 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a really ugly bug. The TableGen'd isel is not freeing the "inflight set"
correctly. That is causing non-deterministic behavior (and possibly preventing
some load folding from happening).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28458 91177308-0d34-0410-b5e6-96231b3b80d8

Abstract out the current optimization level into a flag that can be overridden
on the make line, to avoid bugs in native compilers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28457 91177308-0d34-0410-b5e6-96231b3b80d8

Remove flags implied by -O3

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28456 91177308-0d34-0410-b5e6-96231b3b80d8

Update testcase with recent cbe change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28455 91177308-0d34-0410-b5e6-96231b3b80d8

Clear HandleMap and ReplaceMap after instruction selection. Or it may cause
non-deterministic behavior.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28454 91177308-0d34-0410-b5e6-96231b3b80d8

For PR786:
Minor tweaks in public headers and a few .cpp files so that LLVM can build
successfully with -pedantic and projects using LLVM with -pedantic don't
get warnings from LLVM. There's still more -pedantic warnings to fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28453 91177308-0d34-0410-b5e6-96231b3b80d8

For PR786:
Remove a spurious ;

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28452 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use -fomit-frame-pointer on darwin, it breaks stacktrace collection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28451 91177308-0d34-0410-b5e6-96231b3b80d8

Patch for a new instcombine xform, patch contributed by Nick Lewycky!
This implements Transforms/InstCombine/2006-05-10-InvalidIndexUndef.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28450 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for a new instcombine xform, patch contributed by Nick Lewycky!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28449 91177308-0d34-0410-b5e6-96231b3b80d8

Don't make zero-sized static arrays

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28448 91177308-0d34-0410-b5e6-96231b3b80d8

Patches to make the LLVM sources more -pedantic clean.  Patch provided
by Anton Korobeynikov!  This is a step towards closing PR786.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28447 91177308-0d34-0410-b5e6-96231b3b80d8

One of these xforms is only safe with unsafe math enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28446 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeGen/Generic/vector.ll:test_div with altivec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28445 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28444 91177308-0d34-0410-b5e6-96231b3b80d8

Handle SETO* like we handle SET*, restoring behavior after Evan's setcc
change.  This fixes PowerPC/fnegsel.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28443 91177308-0d34-0410-b5e6-96231b3b80d8

Print struct return functions and calls as actually returning the hidden
argument struct pointer, enabling ABI compatibility for the CBE with
platforms with strange struct-return ABIs.  This fixes 252.eon and
CoyoteBench/fftbench on Darwin/X86 among other things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28442 91177308-0d34-0410-b5e6-96231b3b80d8

Fix file header comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28441 91177308-0d34-0410-b5e6-96231b3b80d8

Better way to check for vararg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28440 91177308-0d34-0410-b5e6-96231b3b80d8

Remove PreprocessCCCArguments and PreprocessFastCCArguments now that
FORMAL_ARGUMENTS nodes include a token operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28439 91177308-0d34-0410-b5e6-96231b3b80d8

Implement an annoying part of the Darwin/X86 abi: the callee of a struct
return argument pops the hidden struct pointer if present, not the caller.

For example, in this testcase:

struct X { int D, E, F, G; };
struct X bar() {
  struct X a;
  a.D = 0;
  a.E = 1;
  a.F = 2;
  a.G = 3;
  return a;
}
void foo(struct X *P) {
  *P = bar();
}

We used to emit:

_foo:
        subl $28, %esp
        movl 32(%esp), %eax
        movl %eax, (%esp)
        call _bar
        addl $28, %esp
        ret
_bar:
        movl 4(%esp), %eax
        movl $0, (%eax)
        movl $1, 4(%eax)
        movl $2, 8(%eax)
        movl $3, 12(%eax)
        ret

This is correct on Linux/X86 but not Darwin/X86.  With this patch, we now
emit:

_foo:
        subl $28, %esp
        movl 32(%esp), %eax
        movl %eax, (%esp)
        call _bar
***     addl $24, %esp
        ret
_bar:
        movl 4(%esp), %eax
        movl $0, (%eax)
        movl $1, 4(%eax)
        movl $2, 8(%eax)
        movl $3, 12(%eax)
***     ret $4

For the record, GCC emits (which is functionally equivalent to our new code):

_bar:
        movl    4(%esp), %eax
        movl    $3, 12(%eax)
        movl    $2, 8(%eax)
        movl    $1, 4(%eax)
        movl    $0, (%eax)
        ret     $4
_foo:
        pushl   %esi
        subl    $40, %esp
        movl    48(%esp), %esi
        leal    16(%esp), %eax
        movl    %eax, (%esp)
        call    _bar
        subl    $4, %esp
        movl    16(%esp), %eax
        movl    %eax, (%esi)
        movl    20(%esp), %eax
        movl    %eax, 4(%esi)
        movl    24(%esp), %eax
        movl    %eax, 8(%esi)
        movl    28(%esp), %eax
        movl    %eax, 12(%esi)
        addl    $40, %esp
        popl    %esi
        ret

This fixes SingleSource/Benchmarks/CoyoteBench/fftbench with LLC and the
JIT, and fixes the X86-backend portion of PR729.  The CBE still needs to
be updated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28438 91177308-0d34-0410-b5e6-96231b3b80d8

-enable-unsafe-fp-math implies -enable-finite-only-fp-math

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28437 91177308-0d34-0410-b5e6-96231b3b80d8

Make class comment visible in Doxygen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28436 91177308-0d34-0410-b5e6-96231b3b80d8

Fix missing include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28435 91177308-0d34-0410-b5e6-96231b3b80d8

Added a test case for FP equality check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28434 91177308-0d34-0410-b5e6-96231b3b80d8

Incorrect SETCC CondCode used for FP comparisons.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28433 91177308-0d34-0410-b5e6-96231b3b80d8

Added option -enable-finite-only-fp-math. When on, the codegen can assume that
FP arithmetic arguments and results are never NaNs or +=Infs. This includes
ignoring parity flag (PF) when checking for FP equality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28432 91177308-0d34-0410-b5e6-96231b3b80d8

implement minimal versions of
ARMAsmPrinter::runOnMachineFunction
LowerFORMAL_ARGUMENTS
ARMInstrInfo::isMoveInstr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28431 91177308-0d34-0410-b5e6-96231b3b80d8

Describe how to add a custom test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28430 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap long lines fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28429 91177308-0d34-0410-b5e6-96231b3b80d8

A isel deficiency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28427 91177308-0d34-0410-b5e6-96231b3b80d8

For PR784:
Support Win32 platforms for llvm-gcc path. Patch by Anton Korobeynikov

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28426 91177308-0d34-0410-b5e6-96231b3b80d8

Back out indirect branch load folding hack. It broke some tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28425 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28424 91177308-0d34-0410-b5e6-96231b3b80d8

Make TargetData strings less redundant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28423 91177308-0d34-0410-b5e6-96231b3b80d8

Silence a bogus gcc warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28422 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a parsing bug that caused 7 llvm-test regressions on PPC last night.
I'm suprised it didn't cause more!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28421 91177308-0d34-0410-b5e6-96231b3b80d8

- Use of load's chain result should be redirected to load's chain operand.
  If it reads the chain result of the call, then the use, callseq_start,
  and call would form a cycle!
- Don't forget handle node replacement!
- There could also be a TokenFactor between the load and the callseq_start.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28420 91177308-0d34-0410-b5e6-96231b3b80d8

A new entry

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28419 91177308-0d34-0410-b5e6-96231b3b80d8

Missing break statements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28418 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28417 91177308-0d34-0410-b5e6-96231b3b80d8

Handle indirect call which folds a load manually. This never matches by
the TableGen generated code since the load's chain result is read by
the callseq_start node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28416 91177308-0d34-0410-b5e6-96231b3b80d8

Sparc is big-endian.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28415 91177308-0d34-0410-b5e6-96231b3b80d8

Make all of the TargetMachine subclasses use the new string TargetData methods.

This is part of the on-going work on PR 761.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28414 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for csretcc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28413 91177308-0d34-0410-b5e6-96231b3b80d8

Print csretcc calls like this:

        call csretcc void %structret( { sbyte }* %P )

instead of this:

        callcsretcc  void %structret( { sbyte }* %P )

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28412 91177308-0d34-0410-b5e6-96231b3b80d8

Fix misencoding of calling conventions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28411 91177308-0d34-0410-b5e6-96231b3b80d8

pretty print csretcc for calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28410 91177308-0d34-0410-b5e6-96231b3b80d8

CSRet allows varargs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28409 91177308-0d34-0410-b5e6-96231b3b80d8