Fold comparisons against a constant nan, and optimize ORD/UNORD
comparisons with a constant.  This allows us to compile isnan to:

_foo:
fcmpu cr7, f1, f1
mfcr r2
rlwinm r3, r2, 0, 31, 31
blr

instead of:

LCPI1_0: ;  float
.space 4
_foo:
lis r2, ha16(LCPI1_0)
lfs f0, lo16(LCPI1_0)(r2)
fcmpu cr7, f1, f0
mfcr r2
rlwinm r3, r2, 0, 31, 31
blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45405 91177308-0d34-0410-b5e6-96231b3b80d8

this xform is implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45404 91177308-0d34-0410-b5e6-96231b3b80d8

Disable null pointer folding transforms for non-generic address spaces. This should probably be a target-specific predicate based on address space. That way for targets where this isn't applicable the predicate can be optimized away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45403 91177308-0d34-0410-b5e6-96231b3b80d8

make sure not to zap volatile stores, thanks a lot to Dale for noticing this!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45402 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen:

as:

_bar:
pushl %esi
subl $8, %esp
movl 16(%esp), %esi
call L_foo$stub
fstps (%esi)
addl $8, %esp
popl %esi
#FP_REG_KILL
ret

instead of:

_bar:
pushl %esi
subl $8, %esp
movl 16(%esp), %esi
call L_foo$stub
fstpl (%esi)
cvtsd2ss (%esi), %xmm0
movss %xmm0, (%esi)
addl $8, %esp
popl %esi
#FP_REG_KILL
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45401 91177308-0d34-0410-b5e6-96231b3b80d8

don't fold fp_round(fp_extend(load)) -> fp_round(extload)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45400 91177308-0d34-0410-b5e6-96231b3b80d8

avoid going through a stack slot to convert from fpstack to xmm reg
if we are just going to store it back anyway.  This improves things
like:
double foo();
void bar(double *P) { *P = foo(); }

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45399 91177308-0d34-0410-b5e6-96231b3b80d8

Delete a store whose input is a load from the same pointer:
  x = load p
  store x -> p

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45398 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45397 91177308-0d34-0410-b5e6-96231b3b80d8

expand note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45393 91177308-0d34-0410-b5e6-96231b3b80d8

dead calls to llvm.stacksave can be deleted, even though they
have potential side-effects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45392 91177308-0d34-0410-b5e6-96231b3b80d8

upgrade this test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45391 91177308-0d34-0410-b5e6-96231b3b80d8

Test -simplifycfg only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45389 91177308-0d34-0410-b5e6-96231b3b80d8

add a note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45388 91177308-0d34-0410-b5e6-96231b3b80d8

add a note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45387 91177308-0d34-0410-b5e6-96231b3b80d8

Add a testcase for my recent InstCombine fix, written by Nicholas.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45386 91177308-0d34-0410-b5e6-96231b3b80d8

Repair a transform that Chris noticed a bug in.  Thanks to Nicholas for pointing out my stupid mistakes when writing this patch. :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45384 91177308-0d34-0410-b5e6-96231b3b80d8

disable this instcombine xform, it miscompiles:

define i32 @main() {
entry:
%z = alloca i32 ; <i32*> [#uses=2]
store i32 0, i32* %z
%tmp = load i32* %z ; <i32> [#uses=1]
%sub = sub i32 %tmp, 1 ; <i32> [#uses=1]
%cmp = icmp ult i32 %sub, 0 ; <i1> [#uses=1]
%retval = select i1 %cmp, i32 1, i32 0 ; <i32> [#uses=1]
ret i32 %retval
}

into ret 1, instead of ret 0.

Christopher, please investigate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45383 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45377 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore functions with internal linkages during linking. This snipped mimics the
behaviour of LinkGlobals() function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45375 91177308-0d34-0410-b5e6-96231b3b80d8

Adding bindings for target triple and data layout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45369 91177308-0d34-0410-b5e6-96231b3b80d8

Switch the bindings to use LLVMFoldingBuilder.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45367 91177308-0d34-0410-b5e6-96231b3b80d8

Adding an uninitialized builder constructor to the Ocaml bindings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45362 91177308-0d34-0410-b5e6-96231b3b80d8

one fewer uncond branch with my codegenprepare hack for single-mbb backedges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45360 91177308-0d34-0410-b5e6-96231b3b80d8

Tests for changes made in r45356, where IPO optimizations would drop
collector algorithms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45357 91177308-0d34-0410-b5e6-96231b3b80d8

Fixing several transforms which would drop the collector attribute
when copying functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45356 91177308-0d34-0410-b5e6-96231b3b80d8

update to llvm 2.0 syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45355 91177308-0d34-0410-b5e6-96231b3b80d8

Don't break critical edges for single-bb loops, this helps with PR1877, though
it is only a partial fix.  This change is noise for most programs, but
speeds up Shootout-C++/matrix by 20%, Ptrdist/ks by 24%, smg2000 by 8%,
hexxagon by 9%, bzip2 by 9% (not sure I trust this), ackerman by 13%, etc.

OTOH, it slows down Shootout/fib2 by 40% (I'll update PR1877 with this info).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45354 91177308-0d34-0410-b5e6-96231b3b80d8

Actually parse q operation in llvm-ar

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45353 91177308-0d34-0410-b5e6-96231b3b80d8

Disabling -g for ocaml builds; it's not downwards compatible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45352 91177308-0d34-0410-b5e6-96231b3b80d8

GC poses hazards to the inliner. Consider:

    define void @f() {
            ...
            call i32 @g()
            ...
    }

    define void @g() {
            ...
    }

The hazards are:

  - @f and @g have GC, but they differ GC. Inlining is invalid. This
    may never occur.
  - @f has no GC, but @g does. g's GC must be propagated to @f.

The other scenarios are safe:

  - @f and @g have the same GC.
  - @f and @g have no GC.
  - @g has no GC.

This patch adds inliner checks for the former two scenarios.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45351 91177308-0d34-0410-b5e6-96231b3b80d8

Noting and enforcing that GC intrinsics are valid only within a
function with GC.

This will catch the error when the inliner inlines a function with
GC into a caller with no GC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45350 91177308-0d34-0410-b5e6-96231b3b80d8

Adjusting verification of "llvm.gc*" intrinsic prototypes to match
LangRef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45349 91177308-0d34-0410-b5e6-96231b3b80d8

Flesh out the Briggs implementation a little bit more, fix a few FIXMEs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45347 91177308-0d34-0410-b5e6-96231b3b80d8

add a -backedge-hack llc-beta option to codegenprepare.
When specified, don't split backedges of single-bb loops.
This helps address PR1877

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45344 91177308-0d34-0410-b5e6-96231b3b80d8

add a simple hack

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45343 91177308-0d34-0410-b5e6-96231b3b80d8

Cygwin defines uint32_t as unsigned long. Unbreak call to std::min in this case

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45342 91177308-0d34-0410-b5e6-96231b3b80d8

Added special support for stripping CRLF characters that may appear in the
output of nm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45341 91177308-0d34-0410-b5e6-96231b3b80d8

Setting GlobalDirective in TargetAsmInfo by default rather than
providing a misleading facility. It's used once in the MIPS backend
and hardcoded as "\t.globl\t" everywhere else.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45338 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of an unneeded #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45337 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some Ocaml GC errors noticed upon review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45336 91177308-0d34-0410-b5e6-96231b3b80d8

C and Ocaml bindings for ExecutionEngine (i.e., the JIT compiler).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45335 91177308-0d34-0410-b5e6-96231b3b80d8

Sketch out an implementation of Briggs' copy placement algorithm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45334 91177308-0d34-0410-b5e6-96231b3b80d8

Add GraphTraits specializations for machine dominators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45333 91177308-0d34-0410-b5e6-96231b3b80d8

Undo a wrong bit in that last patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45327 91177308-0d34-0410-b5e6-96231b3b80d8

Document a more robust implementation of performCustomLowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45326 91177308-0d34-0410-b5e6-96231b3b80d8

fix some warnings.  This code needs to be de-tabified :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45325 91177308-0d34-0410-b5e6-96231b3b80d8

fix strict-aliasing violation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45324 91177308-0d34-0410-b5e6-96231b3b80d8

initial code for forming an FGETSIGN node. This is disabled until
legalizer support goes in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45323 91177308-0d34-0410-b5e6-96231b3b80d8

improve support for fgetsign

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45322 91177308-0d34-0410-b5e6-96231b3b80d8

Tell TargetLoweringOpt whether it is running before
or after legalize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45321 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new FGETSIGN operation, which defaults to expand on all
targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45320 91177308-0d34-0410-b5e6-96231b3b80d8

Erm, really disable :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45319 91177308-0d34-0410-b5e6-96231b3b80d8

Disable, until we'll really need it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45318 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a partial application typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45317 91177308-0d34-0410-b5e6-96231b3b80d8

Use getIntrinsicID instead of looking up intrinsic prototypes. Also
fixes a bug with indirect calls. (Test case will be included with
ocaml collector patch.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45316 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak build on gcc 3.x

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45315 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak mingw build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45314 91177308-0d34-0410-b5e6-96231b3b80d8

Preliminary PIC JIT support for X86 (32-bit) / Darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45313 91177308-0d34-0410-b5e6-96231b3b80d8

Oops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45312 91177308-0d34-0410-b5e6-96231b3b80d8

implement InstCombine/shift-trunc-shift.ll.  This allows
us to compile:
#include <math.h>
int t1(double d) { return signbit(d); }

into:

_t1:
movd %xmm0, %rax
shrq $63, %rax
ret

instead of:

_t1:
movd %xmm0, %rax
shrq $32, %rax
shrl $31, %eax
ret

on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45311 91177308-0d34-0410-b5e6-96231b3b80d8

Note what still needs doing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45310 91177308-0d34-0410-b5e6-96231b3b80d8

Remove critical edge breaking.  It won't be necessary as long as we are very careful when inserting copies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45309 91177308-0d34-0410-b5e6-96231b3b80d8

Ignoring generated files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45308 91177308-0d34-0410-b5e6-96231b3b80d8

Fix JIT code emission of X86::MovePCtoStack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45307 91177308-0d34-0410-b5e6-96231b3b80d8

If succ has succ itself as one of the predecessors then do
not merge current bb and succ even if bb's terminator is
unconditional branch to succ.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45305 91177308-0d34-0410-b5e6-96231b3b80d8

Allow JIT with non-static relocation model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45304 91177308-0d34-0410-b5e6-96231b3b80d8

Fix silly typo in the FP CEP handling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45300 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a brain fart by our beloved leader (the content
of this patch is the last line).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45289 91177308-0d34-0410-b5e6-96231b3b80d8

Get the verifier to check attributes on calls as well
as on functions.  Make it verify invokes and not just
ordinary calls.  As a (desired) side-effect, it is no
longer legal to have call attributes on arguments that
are being passed to the varargs part of a varargs
function (llvm-as drops them on the floor anyway).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45286 91177308-0d34-0410-b5e6-96231b3b80d8

Make DAE not wipe out attributes on calls, and not drop
return attributes on the floor.  In the case of a call
to a varargs function where the varargs arguments are
being removed, any call attributes on those arguments
need to be dropped.  I didn't do this because I plan to
make it illegal to have such attributes (see next patch).
With this change, compiling the gcc filter2 eh test at -O0
and then running opt -std-compile-opts on it results in
a correctly working program (compiling at -O1 or higher
results in the test failing due to a problem with how we
output eh info into the IR).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45285 91177308-0d34-0410-b5e6-96231b3b80d8

Fix unintented change from last commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45282 91177308-0d34-0410-b5e6-96231b3b80d8

Enable EH for linux/ppc32 targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45281 91177308-0d34-0410-b5e6-96231b3b80d8

New entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45280 91177308-0d34-0410-b5e6-96231b3b80d8

Add a few more missing gcc builtin's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45278 91177308-0d34-0410-b5e6-96231b3b80d8

Removed scoped_ptr, as its functionality is subsumed by OwningPtr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45274 91177308-0d34-0410-b5e6-96231b3b80d8

Fix JIT encoding for CMPSD as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45268 91177308-0d34-0410-b5e6-96231b3b80d8

Added OwningArrayPtr smart pointer class to provide an analogous class to
OwningPtr except that it works for pointers to arrays.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45266 91177308-0d34-0410-b5e6-96231b3b80d8

add new smart pointer for clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45261 91177308-0d34-0410-b5e6-96231b3b80d8

Type specification didn't match gcc's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45260 91177308-0d34-0410-b5e6-96231b3b80d8

More accurate checks for two-address constraints.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45259 91177308-0d34-0410-b5e6-96231b3b80d8

Implement review feedback, including additional transforms

(icmp slt (sub A B) 1) -> (icmp sle A B)
icmp sgt (sub A B) -1) -> (icmp sge A B)

and add testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45256 91177308-0d34-0410-b5e6-96231b3b80d8

Add m_Zero().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45255 91177308-0d34-0410-b5e6-96231b3b80d8

Remove xfail. This is fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45254 91177308-0d34-0410-b5e6-96231b3b80d8

The physical register + virtual register joining requirement was much too strict.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45253 91177308-0d34-0410-b5e6-96231b3b80d8

Bring back a burr scheduling heuristic that's still needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45252 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up previous patch: PHI uses should not prevent iv reuse if all other uses are addresses. This trades a constant multiply for one fewer iv.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45251 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code with the new m_Zero() pattern.  Make sure the select only
has a single use, and generalize it to not require N to be a constant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45250 91177308-0d34-0410-b5e6-96231b3b80d8

Revert my previous check-in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45249 91177308-0d34-0410-b5e6-96231b3b80d8

More eye-candy stuff :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45247 91177308-0d34-0410-b5e6-96231b3b80d8

Add iterators for child traversal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45246 91177308-0d34-0410-b5e6-96231b3b80d8

Updated comments to reflect what "side effects" means in this situation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45245 91177308-0d34-0410-b5e6-96231b3b80d8

More working CellSPU tests:
- vec_const.ll: Vector constant loads
- immed64.ll: i64, f64 constant loads

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45242 91177308-0d34-0410-b5e6-96231b3b80d8

Add lto version check mechanism.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45238 91177308-0d34-0410-b5e6-96231b3b80d8

Use a module to group calling convention values, too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45236 91177308-0d34-0410-b5e6-96231b3b80d8

Allow iv reuse if the user is a PHI node which is in turn used as addresses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45230 91177308-0d34-0410-b5e6-96231b3b80d8

Using modules to group enumerations in Ocaml bindings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45229 91177308-0d34-0410-b5e6-96231b3b80d8

Adding bindings for memory buffers and module providers. Switching
to exceptions rather than variants for error handling in Ocaml.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45226 91177308-0d34-0410-b5e6-96231b3b80d8

Enable EH on PPC Darwin.  This basically works; there
are a couple of issues that show up with the optimizer,
but I don't think they're really EH problems.
(llvm-gcc testsuite users note:  By default the testsuite
uses the unwinding code that's built as part of your local
llvm-gcc, which does not work.  You need to trick it into
using the installed system unwinding code to get useful
results.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45221 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU testcase, extract_elt.ll: extract vector element.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@45219 91177308-0d34-0410-b5e6-96231b3b80d8