Unbreak JIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50173 91177308-0d34-0410-b5e6-96231b3b80d8

Disable stack realignment for these tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50172 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test becase ABI stack alignment dropped to 'normal' value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50171 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test, instruction count is valid only if stack is not realigned

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50170 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50169 91177308-0d34-0410-b5e6-96231b3b80d8

Only allow increase of max alignment value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50168 91177308-0d34-0410-b5e6-96231b3b80d8

Be over-conservative: scan for all used virtual registers and calculate maximal stack alignment in assumption, that there will be spill of vector register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50167 91177308-0d34-0410-b5e6-96231b3b80d8

Add X86 Maximal Stack Alignment Calculator Pass before RA

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50166 91177308-0d34-0410-b5e6-96231b3b80d8

Add facility for pre-RA passes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50165 91177308-0d34-0410-b5e6-96231b3b80d8

Use precomputed value, if any

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50164 91177308-0d34-0410-b5e6-96231b3b80d8

Do proper book-keeping of offsets and prologue/epilogue code for stack realignment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50163 91177308-0d34-0410-b5e6-96231b3b80d8

If stack realignment is used - incoming args will use EBP as base register and locals - ESP

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50162 91177308-0d34-0410-b5e6-96231b3b80d8

Eastimate required stack alignment early, so we can decide, whether we will need frame pointer or not

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50161 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50160 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50159 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50158 91177308-0d34-0410-b5e6-96231b3b80d8

Make stack alignment options global for all targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50157 91177308-0d34-0410-b5e6-96231b3b80d8

Provide option for enabling-disabling stack realignment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50156 91177308-0d34-0410-b5e6-96231b3b80d8

Disable stack realignment for functions with dynamic-sized alloca's

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50155 91177308-0d34-0410-b5e6-96231b3b80d8

Provide ABI-correct stack alignment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50154 91177308-0d34-0410-b5e6-96231b3b80d8

Provide convenient helpers for some operations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50153 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50152 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50151 91177308-0d34-0410-b5e6-96231b3b80d8

Use the Create static method for some of these objects. They no longer
have public constructors. This should fix PR2246.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50145 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code for propagation of constant arguments into
callees.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50142 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a number of bugs in ipconstantprop, simplify the code, fit in 80 cols,
fix read after free bug (PR2238).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50141 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite multiple return value handling in SCCP.  Before, the -sccp pass
would turn every getresult instruction into undef.  This helps with
rdar://5778210

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50140 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50139 91177308-0d34-0410-b5e6-96231b3b80d8

Validate that the result of a function type is valid using shared
logic with vmcore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50138 91177308-0d34-0410-b5e6-96231b3b80d8

Enforce that multiple return values have to have at least one result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50137 91177308-0d34-0410-b5e6-96231b3b80d8

remove this testcase.  It isn't testing loop rotate, it is testing all
of -std-compile-opts and is now failing because other passes are generating
IR that looks different to input of loop rotate.  Devang, please
introduce a testcase that only runs loop rotate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50136 91177308-0d34-0410-b5e6-96231b3b80d8

returning an empty multiple return list is not valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50135 91177308-0d34-0410-b5e6-96231b3b80d8

fix description of 'ret' to be more correct with multiple return
values.  Clarify that it is impossible to 'multiply return' a
struct with zero elements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50131 91177308-0d34-0410-b5e6-96231b3b80d8

Allow undef as an operand to getresult.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50130 91177308-0d34-0410-b5e6-96231b3b80d8

Verify that the operand of a getresult instruction is a
call/invoke or undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50129 91177308-0d34-0410-b5e6-96231b3b80d8

make this test more interesting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50128 91177308-0d34-0410-b5e6-96231b3b80d8

distill down the essense of this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50125 91177308-0d34-0410-b5e6-96231b3b80d8

new test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50123 91177308-0d34-0410-b5e6-96231b3b80d8

Do not change the type of a ByVal argument to a
type of a different size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50121 91177308-0d34-0410-b5e6-96231b3b80d8

Don't do: "(X & 4) >> 1 == 2  --> (X & 4) == 4" if there are more than one uses of the shift result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50118 91177308-0d34-0410-b5e6-96231b3b80d8

Start doing the significantly useful part of jump threading: handle cases
where a comparison has a phi input and that phi is a constant.  For example,
stuff like:

  Threading edge through bool from 'bb2149' to 'bb2231' with cost: 1, across block:
bb2237: ; preds = %bb2231, %bb2149
%tmp2328.rle = phi i32 [ %tmp2232, %bb2231 ], [ %tmp2232439, %bb2149 ] ; <i32> [#uses=2]
%done.0 = phi i32 [ %done.2, %bb2231 ], [ 0, %bb2149 ] ; <i32> [#uses=1]
%tmp2239 = icmp eq i32 %done.0, 0 ; <i1> [#uses=1]
br i1 %tmp2239, label %bb2231, label %bb2327

or

bb38.i298: ; preds = %bb33.i295, %bb1693
%tmp39.i296.rle = phi %struct.ibox* [ null, %bb1693 ], [ %tmp39.i296.rle1109, %bb33.i295 ] ; <%struct.ibox*> [#uses=2]
%minspan.1.i291.reg2mem.1 = phi i32 [ 32000, %bb1693 ], [ %minspan.0.i288, %bb33.i295 ] ; <i32> [#uses=1]
%tmp40.i297 = icmp eq %struct.ibox* %tmp39.i296.rle, null ; <i1> [#uses=1]
br i1 %tmp40.i297, label %implfeeds.exit311, label %bb43.i301

This triggers thousands of times in spec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50110 91177308-0d34-0410-b5e6-96231b3b80d8

Dig through multiple levels of AND to thread jumps if needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50106 91177308-0d34-0410-b5e6-96231b3b80d8

Move MemCpyOpt after GVN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50097 91177308-0d34-0410-b5e6-96231b3b80d8

Teach jump threading to thread through blocks like:

  br (and X, phi(Y, Z, false)), label L1, label L2

This triggers once on 252.eon and 6 times on 176.gcc.  Blocks
in question often look like this:

bb262: ; preds = %bb261, %bb248
%iftmp.251.0 = phi i1 [ true, %bb261 ], [ false, %bb248 ] ; <i1> [#uses=4]
%tmp270 = icmp eq %struct.rtx_def* %tmp.0.i, null ; <i1> [#uses=1]
%bothcond = or i1 %iftmp.251.0, %tmp270 ; <i1> [#uses=1]
br i1 %bothcond, label %bb288, label %bb273

In this case, it is clear that it doesn't matter if tmp.0.i is null when coming from bb261.  When coming from bb248, it is all that matters.

Another random example:

check_asm_operands.exit: ; preds = %check_asm_operands.exit.thr_comm, %bb30.i, %bb12.i, %bb6.i413
%tmp.0.i420 = phi i1 [ true, %bb6.i413 ], [ true, %bb12.i ], [ true, %bb30.i ], [ false, %check_asm_operands.exit.thr_comm ; <i1> [#uses=1]
call void @llvm.stackrestore( i8* %savedstack ) nounwind
%tmp4389 = icmp eq i32 %added_sets_1.0, 0 ; <i1> [#uses=1]
%tmp4394 = icmp eq i32 %added_sets_2.0, 0 ; <i1> [#uses=1]
%bothcond80 = and i1 %tmp4389, %tmp4394 ; <i1> [#uses=1]
%bothcond81 = and i1 %bothcond80, %tmp.0.i420 ; <i1> [#uses=1]
br i1 %bothcond81, label %bb4398, label %bb4397

Here is the case from 252.eon:

bb290.i.i: ; preds = %bb23.i57.i.i, %bb8.i39.i.i, %bb100.i.i, %bb100.i.i, %bb85.i.i110
%myEOF.1.i.i = phi i1 [ true, %bb100.i.i ], [ true, %bb100.i.i ], [ true, %bb85.i.i110 ], [ true, %bb8.i39.i.i ], [ false, %bb23.i57.i.i ] ; <i1> [#uses=2]
%i.4.i.i = phi i32 [ %i.1.i.i, %bb85.i.i110 ], [ %i.0.i.i, %bb100.i.i ], [ %i.0.i.i, %bb100.i.i ], [ %i.3.i.i, %bb8.i39.i.i ], [ %i.3.i.i, %bb23.i57.i.i ] ; <i32> [#uses=3]
%tmp292.i.i = load i8* %tmp16.i.i100, align 1 ; <i8> [#uses=1]
%tmp293.not.i.i = icmp ne i8 %tmp292.i.i, 0 ; <i1> [#uses=1]
%bothcond.i.i = and i1 %tmp293.not.i.i, %myEOF.1.i.i ; <i1> [#uses=1]
br i1 %bothcond.i.i, label %bb202.i.i, label %bb301.i.i
  Factoring out 3 common predecessors.

On the path from any blocks other than bb23.i57.i.i, the load and compare
are dead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50096 91177308-0d34-0410-b5e6-96231b3b80d8

Added command line option -no-lazy to disable JIT lazy compilation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50095 91177308-0d34-0410-b5e6-96231b3b80d8

refactor some code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50094 91177308-0d34-0410-b5e6-96231b3b80d8

add a basic testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50093 91177308-0d34-0410-b5e6-96231b3b80d8

Whoops! Undo r50087, unbreak the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50088 91177308-0d34-0410-b5e6-96231b3b80d8

Reverse r47989. Part of removing 'unwinds to' support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50087 91177308-0d34-0410-b5e6-96231b3b80d8

Start removing 'unwinds to' support from mainline in preparation for 2.3.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50086 91177308-0d34-0410-b5e6-96231b3b80d8

Add MemCpyOpt to LTO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50081 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50080 91177308-0d34-0410-b5e6-96231b3b80d8

optimize "p != gep p, ..." better.  This allows us to compile
getelementptr-seteq.ll into:

define i1 @test(i64 %X, %S* %P) {
%C = icmp eq i64 %X, -1 ; <i1> [#uses=1]
ret i1 %C
}

instead of:

define i1 @test(i64 %X, %S* %P) {
%A.idx.mask = and i64 %X, 4611686018427387903 ; <i64> [#uses=1]
%C = icmp eq i64 %A.idx.mask, 4611686018427387903 ; <i1> [#uses=1]
ret i1 %C
}

And fixes the second half of PR2235.  This speeds up the insertion sort
case by 45%, from 1.12s to 0.77s.  In practice, this will significantly
speed up for loops structured like:

for (double *P = Base + N; P != Base; --P)
  ...

Which happens frequently for C++ iterators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50079 91177308-0d34-0410-b5e6-96231b3b80d8

more fallout from Nicholas' asmprinter patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50078 91177308-0d34-0410-b5e6-96231b3b80d8

Implement an x86-64 ABI detail of passing structs by hidden first
argument. The x86-64 ABI requires the incoming value of %rdi to
be copied to %rax on exit from a function that is returning a
large C struct.

Also, add a README-X86-64 entry detailing the missed optimization
opportunity and proposing an alternative approach.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50075 91177308-0d34-0410-b5e6-96231b3b80d8

fix some residual old API that fell thru the cracks of the conversion script, closes llvm.org/bugs/show_bug.cgi?id=2246

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50062 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an out-of-bounds access in -view-sunit-dags in the case of an
empty ScheduleDAG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50054 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the encoding of the MMX movd that moves from MMX to 64-bit GPR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50053 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the way AliasSet::print prints "may alias".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50051 91177308-0d34-0410-b5e6-96231b3b80d8

fix grammar-o, thanks to Duncan for noticing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50047 91177308-0d34-0410-b5e6-96231b3b80d8

Make these structs larger to ensure that they
are returned by struct return.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50038 91177308-0d34-0410-b5e6-96231b3b80d8

Make the struct bigger, to ensure it is returned
by struct return.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50037 91177308-0d34-0410-b5e6-96231b3b80d8

Use GVN in LTO instead of GCSE+LoadVN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50036 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unneeded #include's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50035 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor memcpyopt based on Chris' suggestions.  Consolidate several functions
and simplify code that was fallout from the separation of memcpyopt and gvn.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50034 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a "large integer implicitly truncated to unsigned type"
warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50033 91177308-0d34-0410-b5e6-96231b3b80d8

don't print dominators every time it is computed with -debug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50032 91177308-0d34-0410-b5e6-96231b3b80d8

another bug introduced in r47802 by nicholas, for no apparent reason.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50031 91177308-0d34-0410-b5e6-96231b3b80d8

Add an ugly note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50029 91177308-0d34-0410-b5e6-96231b3b80d8

Add jump threading to liblto.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50028 91177308-0d34-0410-b5e6-96231b3b80d8

run the jump threading pass in llvm-ld also

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50027 91177308-0d34-0410-b5e6-96231b3b80d8

enable jump threading pass by default.  This causes no miscompilations
and can significantly shrinkify some code, particularly from C++ land.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50025 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug introduced by Nicholas' "unwinds to" stuff: we
lost newlines between blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50024 91177308-0d34-0410-b5e6-96231b3b80d8

don't assume that the argument passed to fprintf("%s"  is a string.  This
fixes a crash in opt on 433.milc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50023 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new SplitBlockPredecessors to implement a todo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50022 91177308-0d34-0410-b5e6-96231b3b80d8

Move SplitBlockPredecessors out of loopsimplify into BasicBlockUtils.h
as a global helper function.  At the same type, switch it from taking
a vector of predecessors to an arbitrary sequential input.  This allows
us to switch LoopSimplify to use a SmallVector for various temporary
vectors that it passed into SplitBlockPredecessors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50020 91177308-0d34-0410-b5e6-96231b3b80d8

Move domtree/frontier updating earlier, allowing us to use it to update phi
nodes, removing a hack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50019 91177308-0d34-0410-b5e6-96231b3b80d8

Factor dominator tree and frontier updating into SplitBlockPredecessors
instead of doing it after every call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50018 91177308-0d34-0410-b5e6-96231b3b80d8

fit some more code in 80 cols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50016 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code, fit in 80 cols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50015 91177308-0d34-0410-b5e6-96231b3b80d8

fit in 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50014 91177308-0d34-0410-b5e6-96231b3b80d8

Be pessimistic in computing the buffer size when aligning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50008 91177308-0d34-0410-b5e6-96231b3b80d8

Don't forget to update the current operand when getting the size of an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50007 91177308-0d34-0410-b5e6-96231b3b80d8

finish the first cut of a jump threading pass implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50006 91177308-0d34-0410-b5e6-96231b3b80d8

replace a slow and verbose version of Instruction::isUsedOutsideOfBlock with
a call to Instruction::isUsedOutsideOfBlock.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50005 91177308-0d34-0410-b5e6-96231b3b80d8

add a handy helper method to instruction, useful for determining
whether it is used outside of some block.  This can be used to see
if there are any non-local references, for example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50004 91177308-0d34-0410-b5e6-96231b3b80d8

we can only thread blocks when there is a pred we can determine the succ of.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50003 91177308-0d34-0410-b5e6-96231b3b80d8

improve comments, infrastructure, and add some validity checks for threading.
Add a cost function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50002 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new Jump Threading pass, which will handle cases
such as those in PR2235.  Right now the pass is not very
effective. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50000 91177308-0d34-0410-b5e6-96231b3b80d8

hopefully resolve PR2240

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49999 91177308-0d34-0410-b5e6-96231b3b80d8

rearrange some code, simplify handling of shifts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49995 91177308-0d34-0410-b5e6-96231b3b80d8

Check we aren't trying to convert PPC long double.
This fixes the testsuite failure on ppcf128-4.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49994 91177308-0d34-0410-b5e6-96231b3b80d8

Cosmetic changes, as suggested by Evan. No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49993 91177308-0d34-0410-b5e6-96231b3b80d8

(re)fix handling of UGT.  Pointed out by Nick Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49991 91177308-0d34-0410-b5e6-96231b3b80d8

g++-4.3 build-fix: CHAR_BIT requires <climits>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49989 91177308-0d34-0410-b5e6-96231b3b80d8

Do not hold the JIT lock when materializing a function and verify if the
function has already been codegen'd. This is required by the Java class loading
mechanism which executes Java code when materializing a function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49988 91177308-0d34-0410-b5e6-96231b3b80d8

A better fix for my previous patch, MOVZQI2PQIrr just requires SSE2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49986 91177308-0d34-0410-b5e6-96231b3b80d8

Not all x86-64 machines have sse3 apparently.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49985 91177308-0d34-0410-b5e6-96231b3b80d8

Check that APFloat::convert is not trying to target
ppc long double, which doesn't work.

This may break some stuff temporarily, but I want to
avoid the propagation of code that assumes this works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49983 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49978 91177308-0d34-0410-b5e6-96231b3b80d8