oota-llvm.git
9 years ago[x86] Add some tests for a common unpack pattern of vector shuffle that
Chandler Carruth [Sat, 22 Nov 2014 05:44:43 +0000 (05:44 +0000)]
[x86] Add some tests for a common unpack pattern of vector shuffle that
has a remarkably unique and efficient lowering.

While we get this some of the time already, we miss a few cases and
there wasn't a principled reason we got it. We should at least test
this. v8 already has tests for this pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222607 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Preserve nsw when folding X*(2^C) -> X << C
David Majnemer [Sat, 22 Nov 2014 04:52:55 +0000 (04:52 +0000)]
InstCombine: Preserve nsw when folding X*(2^C) -> X << C

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222606 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Preserve nsw/nuw for ((X << C2)*C1) -> (X * (C1 << C2))
David Majnemer [Sat, 22 Nov 2014 04:52:52 +0000 (04:52 +0000)]
InstCombine: Preserve nsw/nuw for ((X << C2)*C1) -> (X * (C1 << C2))

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222605 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Preserve nsw for (mul %V, -1) -> (sub 0, %V)
David Majnemer [Sat, 22 Nov 2014 04:52:38 +0000 (04:52 +0000)]
InstCombine: Preserve nsw for (mul %V, -1) -> (sub 0, %V)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222604 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[InstCombine] Re-commit of r218721 (Optimize icmp-select-icmp sequence)
Gerolf Hoflehner [Fri, 21 Nov 2014 23:36:44 +0000 (23:36 +0000)]
[InstCombine] Re-commit of  r218721 (Optimize icmp-select-icmp sequence)

Fixes the self-host fail. Note that this commit activates dominator
analysis in the combiner by default (like the original commit did).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222590 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix transformation of add with pc argument to adr for non-immediate
Joerg Sonnenberger [Fri, 21 Nov 2014 22:39:34 +0000 (22:39 +0000)]
Fix transformation of add with pc argument to adr for non-immediate
arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222587 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[asan] remove old experimental code
Kostya Serebryany [Fri, 21 Nov 2014 22:34:29 +0000 (22:34 +0000)]
[asan] remove old experimental code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222586 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add a failing test case for offset order in ds_read2 instructions
Tom Stellard [Fri, 21 Nov 2014 22:31:47 +0000 (22:31 +0000)]
R600/SI: Add a failing test case for offset order in ds_read2 instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222585 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add an s_mov_b32 to patterns which use the M0RegClass
Tom Stellard [Fri, 21 Nov 2014 22:31:46 +0000 (22:31 +0000)]
R600/SI: Add an s_mov_b32 to patterns which use the M0RegClass

We need to use a s_mov_b32 rather than a copy, so that CSE will
eliminate redundant moves to the m0 register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222584 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Emit s_mov_b32 m0, -1 before every DS instruction
Tom Stellard [Fri, 21 Nov 2014 22:31:44 +0000 (22:31 +0000)]
R600/SI: Emit s_mov_b32 m0, -1 before every DS instruction

This s_mov_b32 will write to a virtual register from the M0Reg
class and all the ds instructions now take an extra M0Reg explicit
argument.

This change is necessary to prevent issues with the scheduler
mixing together instructions that expect different values in the m0
registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222583 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add SIFoldOperands pass
Tom Stellard [Fri, 21 Nov 2014 22:06:37 +0000 (22:06 +0000)]
R600/SI: Add SIFoldOperands pass

This pass attempts to fold the source operands of mov and copy
instructions into their uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222581 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][microMIPS] This patch implements functionality in MIPS delay slot
Jozef Kolek [Fri, 21 Nov 2014 22:04:35 +0000 (22:04 +0000)]
[mips][microMIPS] This patch implements functionality in MIPS delay slot
filler such as if delay slot filler have to put NOP instruction into the
delay slot of microMIPS BEQ or BNE instruction which uses the register $0,
then instead of emitting NOP this instruction is replaced by the corresponding
microMIPS compact branch instruction, i.e. BEQZC or BNEZC.

Differential Revision: http://reviews.llvm.org/D3566

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222580 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Mark s_mov_b32 and s_mov_b64 as rematerializable
Tom Stellard [Fri, 21 Nov 2014 22:00:16 +0000 (22:00 +0000)]
R600/SI: Mark s_mov_b32 and s_mov_b64 as rematerializable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222579 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Use hex notation for constant in test
Tom Stellard [Fri, 21 Nov 2014 22:00:13 +0000 (22:00 +0000)]
R600/SI: Use hex notation for constant in test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222578 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding sxth instruction.
Colin LeMahieu [Fri, 21 Nov 2014 21:54:59 +0000 (21:54 +0000)]
[Hexagon] Adding sxth instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222577 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding sxtb instruction. Renaming some identically named classes that...
Colin LeMahieu [Fri, 21 Nov 2014 21:35:52 +0000 (21:35 +0000)]
[Hexagon] Adding sxtb instruction.  Renaming some identically named classes that will be removed after converting referencing defs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222575 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[asan] add statistic counter to dynamic alloca instrumentation
Kostya Serebryany [Fri, 21 Nov 2014 21:25:18 +0000 (21:25 +0000)]
[asan] add statistic counter to dynamic alloca instrumentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222573 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Removing SUB_rr and replacing with A2_sub.
Colin LeMahieu [Fri, 21 Nov 2014 21:19:18 +0000 (21:19 +0000)]
[Hexagon] Removing SUB_rr and replacing with A2_sub.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222571 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove duplication of relocation names in lib/Object/ELFYAML.cpp
Tim Northover [Fri, 21 Nov 2014 20:16:09 +0000 (20:16 +0000)]
Remove duplication of relocation names in lib/Object/ELFYAML.cpp

We can now use the ELF relocation .def files to create the mapping
of relocation numbers to names and avoid having to duplicate the
list of relocations.

Patch by Will Newton.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222567 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove duplication of relocation names in lib/Object/ELF.cpp
Tim Northover [Fri, 21 Nov 2014 20:16:07 +0000 (20:16 +0000)]
Remove duplication of relocation names in lib/Object/ELF.cpp

We can now use the ELF relocation .def files to create the mapping
of relocation numbers to names and avoid having to duplicate the
list of relocations.

Patch by Will Newton.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222566 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSplit ELF relocation defintions into per-architecture .def files
Tim Northover [Fri, 21 Nov 2014 20:16:02 +0000 (20:16 +0000)]
Split ELF relocation defintions into per-architecture .def files

This should allow the list of relocations for a particular
architecture to be kept in a single header rather than duplicated
whenever we need to enumerate all the relocations.

Patch by Will Newton.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222565 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug Info: revert r222195, r222210 and r222239.
Manman Ren [Fri, 21 Nov 2014 19:55:23 +0000 (19:55 +0000)]
Debug Info: revert r222195, r222210 and r222239.

This is no longer needed after David's fix at r222377 + r222485.
rdar://18958417

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222563 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDisable header duplication at -Oz in loop-rotate pass.
Roman Divacky [Fri, 21 Nov 2014 19:53:24 +0000 (19:53 +0000)]
Disable header duplication at -Oz in loop-rotate pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222562 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug Info: add an assertion that the context field of a global variable can not
Manman Ren [Fri, 21 Nov 2014 19:47:48 +0000 (19:47 +0000)]
Debug Info: add an assertion that the context field of a global variable can not
be a DIType with identifier.

This makes sure that there is no need to use DIScopeRef for global variable's
context.

rdar://18958417

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222561 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Objective-C] Support a new special module flag that will be put into the
Manman Ren [Fri, 21 Nov 2014 19:24:55 +0000 (19:24 +0000)]
[Objective-C] Support a new special module flag that will be put into the
objc_imageinfo struct.

rdar://17954668

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222558 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLazyValueInfo: range'ify some for-loops. No functional change.
Hans Wennborg [Fri, 21 Nov 2014 19:07:46 +0000 (19:07 +0000)]
LazyValueInfo: range'ify some for-loops. No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222557 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd params() to FunctionType. NFC.
Rafael Espindola [Fri, 21 Nov 2014 19:03:35 +0000 (19:03 +0000)]
Add params() to FunctionType. NFC.

While at it, also use makeArrayRef in elements().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222556 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't repeat class/function/variable names in comments. NFC.
Sanjay Patel [Fri, 21 Nov 2014 18:58:38 +0000 (18:58 +0000)]
Don't repeat class/function/variable names in comments. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222555 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLazyValueInfo: fix some typos and indentation, etc. NFC.
Hans Wennborg [Fri, 21 Nov 2014 18:58:23 +0000 (18:58 +0000)]
LazyValueInfo: fix some typos and indentation, etc. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222554 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd and use a helper elements() to StructType. NFC.
Rafael Espindola [Fri, 21 Nov 2014 18:53:05 +0000 (18:53 +0000)]
Add and use a helper elements() to StructType. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222553 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAllow multiple -debug-only args
Matthias Braun [Fri, 21 Nov 2014 18:06:09 +0000 (18:06 +0000)]
Allow multiple -debug-only args

Debug output is shown if any of the -debug-only arguments match.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222547 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLess space; NFC
Sanjay Patel [Fri, 21 Nov 2014 18:05:59 +0000 (18:05 +0000)]
Less space; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222546 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix formatting. NFC.
Rafael Espindola [Fri, 21 Nov 2014 18:05:55 +0000 (18:05 +0000)]
Fix formatting. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222545 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a feature flag for slow 32-byte unaligned memory accesses [x86].
Sanjay Patel [Fri, 21 Nov 2014 17:40:04 +0000 (17:40 +0000)]
Add a feature flag for slow 32-byte unaligned memory accesses [x86].

This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen
for Sandy Bridge and Ivy Bridge. There is no functionality change intended for
those chips. Previously, the absence of AVX2 was being used as a proxy to detect
this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2
that do not have the 32-byte unaligned access slowdown.

Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ).

Differential Revision: http://reviews.llvm.org/D6355

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222544 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Allow FDE references outside the +/-2GB range supported by PC relative offset...
Duncan P. N. Exon Smith [Fri, 21 Nov 2014 17:21:18 +0000 (17:21 +0000)]
Revert "Allow FDE references outside the +/-2GB range supported by PC relative offsets for code models other than small/medium. For JIT application, memory layout is less controlled and can result in truncations otherwise."

This reverts commit r222538.

It's causing test failures for CFI, at least on Darwin:

http://lab.llvm.org:8080/green/job/clang-stage1-cmake-RA-incremental/1189/
http://lab.llvm.org:8080/green/job/clang-stage1-configure-RA_check/1391/

Note that the previous incremental build was on r222537, and the CFI
tests weren't failing:
http://lab.llvm.org:8080/green/job/clang-stage1-cmake-RA-incremental/1188/

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222542 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Restructure the checking patterns for v16 and v32 avx2 vector
Chandler Carruth [Fri, 21 Nov 2014 14:53:03 +0000 (14:53 +0000)]
[x86] Restructure the checking patterns for v16 and v32 avx2 vector
shuffle lowering to allow much better blend matching.

Specifically, with the new structure the code seems clearer to me and we
correctly can hit the cases where merging two 128-bit lanes is a clear
win and can be shuffled cheaply afterward.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222539 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAllow FDE references outside the +/-2GB range supported by PC relative
Joerg Sonnenberger [Fri, 21 Nov 2014 14:42:43 +0000 (14:42 +0000)]
Allow FDE references outside the +/-2GB range supported by PC relative
offsets for code models other than small/medium. For JIT application,
memory layout is less controlled and can result in truncations
otherwise.

Patch from Akos Kiss.

Differential Revision: http://reviews.llvm.org/D6079

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222538 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Make the previous logic significantly less conservative and get
Chandler Carruth [Fri, 21 Nov 2014 14:33:24 +0000 (14:33 +0000)]
[x86] Make the previous logic significantly less conservative and get
a bunch more improvements.

Non-lane-crossing is fine, the key is that lane merging only makes sense
for single-input shuffles. Not sure why I got so turned around here. The
code all works, I was just using the wrong model for it.

This only updates v4 and v8 lowering. The v16 and v32 lowering requires
restructuring the entire check sequence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222537 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[DAG] Teach how to turn a build_vector into a shuffle if some of the operands are...
Andrea Di Biagio [Fri, 21 Nov 2014 14:32:06 +0000 (14:32 +0000)]
[DAG] Teach how to turn a build_vector into a shuffle if some of the operands are zero.

Before this patch, the DAGCombiner only tried to convert build_vector dag nodes
into shuffles if all operands were either extract_vector_elt or undef.

This patch improves that logic and teaches the DAGCombiner how to deal with
build_vector dag nodes where one or more operands are zero. A build_vector
dag node with some zero operands is turned into a shuffle only if the resulting
shuffle mask is legal for the target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222536 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Teach the x86 vector shuffle lowering to detect mergable 128-bit
Chandler Carruth [Fri, 21 Nov 2014 13:56:05 +0000 (13:56 +0000)]
[x86] Teach the x86 vector shuffle lowering to detect mergable 128-bit
lanes.

By special casing these we can often either reduce the total number of
shuffles significantly or reduce the number of (high latency on Haswell)
AVX2 shuffles that potentially cross 128-bit lanes. Even when these
don't actually cross lanes, they have much higher latency to support
that. Doing two of them and a blend is worse than doing a single insert
across the 128-bit lanes to blend and then doing a single interleaved
shuffle.

While this seems like a narrow case, it kept cropping up on me and the
difference is *huge* as you can see in many of the test cases. I first
hit this trying to perfectly fix the interleaving shuffle patterns used
by Halide for AVX2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222533 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Remove more windows line endings that slipped into this file...
Chandler Carruth [Fri, 21 Nov 2014 12:33:46 +0000 (12:33 +0000)]
[x86] Remove more windows line endings that slipped into this file...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222528 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Add a bunch of test cases to 256-bit shuffles that exercise
Chandler Carruth [Fri, 21 Nov 2014 12:17:50 +0000 (12:17 +0000)]
[x86] Add a bunch of test cases to 256-bit shuffles that exercise
merging 128-bit subvectors and also shuffling all the elements of those
subvectors. Currently we generate pretty bad code for many of these, but
I'm testing a patch that should dramatically improve this in addition to
making the shuffle lowering robust to other changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222525 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[DAG] Refactor the shuffle combining logic in DAGCombiner. NFC.
Andrea Di Biagio [Fri, 21 Nov 2014 11:33:07 +0000 (11:33 +0000)]
[DAG] Refactor the shuffle combining logic in DAGCombiner. NFC.

This patch simplifies the logic that combines a pair of shuffle nodes into
a single shuffle if there is a legal mask. Also added comments to better
describe the algorithm. No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222522 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86] For Silvermont CPU use 16-bit division instead of 64-bit for small positive...
Alexey Volkov [Fri, 21 Nov 2014 11:19:34 +0000 (11:19 +0000)]
[X86] For Silvermont CPU use 16-bit division instead of 64-bit for small positive numbers

Differential Revision: http://reviews.llvm.org/D5938

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222521 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[asan] Add new hidden compile-time flag asan-instrument-allocas to sanitize variable...
Yury Gribov [Fri, 21 Nov 2014 10:29:50 +0000 (10:29 +0000)]
[asan] Add new hidden compile-time flag asan-instrument-allocas to sanitize variable-sized dynamic allocas. Patch by Max Ostapenko.

Reviewed at http://reviews.llvm.org/D6055

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222519 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd LLVMScalarOpts to LLVMPowerPCCodeGen.
NAKAMURA Takumi [Fri, 21 Nov 2014 09:14:45 +0000 (09:14 +0000)]
Add LLVMScalarOpts to LLVMPowerPCCodeGen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222516 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDAGCombiner: Allow the DAGCombiner to combine multiple FDIVs with the same divisor...
Hao Liu [Fri, 21 Nov 2014 06:39:58 +0000 (06:39 +0000)]
DAGCombiner: Allow the DAGCombiner to combine multiple FDIVs with the same divisor info FMULs by the reciprocal.
E.g., ( a / D; b / D ) -> ( recip = 1.0 / D; a * recip; b * recip)

A hook is added to allow the target to control whether it needs to do such combine.

Reviewed in http://reviews.llvm.org/D6334

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222510 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove a bunch of unnecessary typecasts to 'const TargetRegisterClass *'
Craig Topper [Fri, 21 Nov 2014 05:58:21 +0000 (05:58 +0000)]
Remove a bunch of unnecessary typecasts to 'const TargetRegisterClass *'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222509 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd extra new line and remove some trailing whitespace from tablegen RegisterInfo...
Craig Topper [Fri, 21 Nov 2014 05:58:14 +0000 (05:58 +0000)]
Add extra new line and remove some trailing whitespace from tablegen RegisterInfo output file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222508 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a silly bug in StreamingMemoryObject.cpp.
Rafael Espindola [Fri, 21 Nov 2014 05:15:41 +0000 (05:15 +0000)]
Fix a silly bug in StreamingMemoryObject.cpp.

The logic for detecting EOF was wrong and would fail if we ever requested
more than 16k past the last read position.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222505 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[PPC] Use SeparateConstOffsetFromGEP
Hal Finkel [Fri, 21 Nov 2014 04:35:51 +0000 (04:35 +0000)]
[PPC] Use SeparateConstOffsetFromGEP

This mirrors r222331, which enabled SeparateConstOffsetFromGEP on AArch64, in
the PowerPC backend. Yields, on a POWER7 machine, a 30% speedup on
SingleSource/Benchmarks/Shootout/nestedloop (this might just be from LICM,
there is a store moved out of the inner loop) and a potential speedup on
MultiSource/Benchmarks/mediabench/mpeg2/mpeg2dec/mpeg2decode. Regardless, it
makes some code look cleaner, and synchronizing the backends in this regard
seems like a generally good thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222504 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd accessor marcos to ConstantPlaceHolder, similar to those in the base class.
Richard Trieu [Fri, 21 Nov 2014 02:42:08 +0000 (02:42 +0000)]
Add accessor marcos to ConstantPlaceHolder, similar to those in the base class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222502 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoThis Reassociate change unintentionally slipped in r222499
David Majnemer [Fri, 21 Nov 2014 02:37:38 +0000 (02:37 +0000)]
This Reassociate change unintentionally slipped in r222499

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222500 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSROA: The alloca type isn't a candidate promotion type for vectors
David Majnemer [Fri, 21 Nov 2014 02:34:55 +0000 (02:34 +0000)]
SROA: The alloca type isn't a candidate promotion type for vectors

The alloca's type is irrelevant, only those types which are used in a
load or store of the exact size of the slice should be considered.

This manifested as an assertion failure when we compared the various
types: we had a size mismatch.

This fixes PR21480.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222499 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoClarify the description of the noalias attribute
Hal Finkel [Fri, 21 Nov 2014 02:22:46 +0000 (02:22 +0000)]
Clarify the description of the noalias attribute

The previous description of the noalias attribute did not accurately specify
the implemented semantics, and the terminology used differed unnecessarily
from that used by the C specification to define the semantics of restrict. For
the argument attribute, the semantics can be precisely specified in terms of
objects accessed through pointers based on the arguments, and this is now what
is done.

Saying that the semantics are 'slightly weaker' than that provided by C99
restrict is not really useful without further elaboration, so that has been
removed from the sentence.

noalias on a return value is really used to mean that the function is
malloc-like (and, in fact, we use this attribute to represent
__attribute__((malloc)) in Clang), and this is a stronger guarantee than that
provided by restrict (because it is a property of the pointed-to memory region,
not just a guarantee on object access). Clarifying this is relevant to fixing
(and was motivated by the discussion on) PR21556.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222497 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Remove JITEventListener::NotifyFreeingMachineCode. This method is dead
Lang Hames [Fri, 21 Nov 2014 01:57:09 +0000 (01:57 +0000)]
[MCJIT] Remove JITEventListener::NotifyFreeingMachineCode. This method is dead
now that the old JIT has been removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222494 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd curly braces to workaround an MSVC bug.
Zachary Turner [Fri, 21 Nov 2014 01:19:09 +0000 (01:19 +0000)]
Add curly braces to workaround an MSVC bug.

MSVC can't parse this pattern for range-based for loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222491 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86] Do not custom lower UINT_TO_FP when the target type does not
Quentin Colombet [Fri, 21 Nov 2014 00:47:19 +0000 (00:47 +0000)]
[X86] Do not custom lower UINT_TO_FP when the target type does not
match the custom lowering.

<rdar://problem/19026326>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222489 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoVerifier: Check that all instructions have their parent pointers set up
Adrian Prantl [Fri, 21 Nov 2014 00:39:43 +0000 (00:39 +0000)]
Verifier: Check that all instructions have their parent pointers set up
correctly. This helps with catching problems caused by IRBuilder abuse
such as the one fixed in CFE r222487.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222488 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix more instances of -Wsentinel on Windows with s/NULL/nullptr/
Reid Kleckner [Thu, 20 Nov 2014 23:51:47 +0000 (23:51 +0000)]
Fix more instances of -Wsentinel on Windows with s/NULL/nullptr/

Follow up to r221940, where I must not have caught em all. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222481 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd out of line virtual destructors to all LLVMTargetMachine subclasses
Reid Kleckner [Thu, 20 Nov 2014 23:37:18 +0000 (23:37 +0000)]
Add out of line virtual destructors to all LLVMTargetMachine subclasses

These recently all grew a unique_ptr<TargetLoweringObjectFile> member in
r221878.  When anyone calls a virtual method of a class, clang-cl
requires all virtual methods to be semantically valid. This includes the
implicit virtual destructor, which triggers instantiation of the
unique_ptr destructor, which fails because the type being deleted is
incomplete.

This is just part of the ongoing saga of PR20337, which is affecting
Blink as well. Because the MSVC ABI doesn't have key functions, we end
up referencing the vtable and implicit destructor on any virtual call
through a class. We don't actually end up emitting the dtor, so it'd be
good if we could avoid this unneeded type completion work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222480 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUpdate Makefile following directory removal in r222466
Mehdi Amini [Thu, 20 Nov 2014 22:48:24 +0000 (22:48 +0000)]
Update Makefile following directory removal in r222466

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222475 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplifyCFG: Refactor GatherConstantCompares() result in a struct
Mehdi Amini [Thu, 20 Nov 2014 22:40:25 +0000 (22:40 +0000)]
SimplifyCFG: Refactor GatherConstantCompares() result in a struct

Code seems cleaner and easier to understand this way

This is basically r222416, after fixes for MSVC lack of standard
support, and a few cleaning (got rid of a warning).
Thanks Nakamura Takumi and Nico Weber for the MSVC fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222472 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] [NFC] Deleting empty directory.
Colin LeMahieu [Thu, 20 Nov 2014 22:20:40 +0000 (22:20 +0000)]
[Hexagon] [NFC] Deleting empty directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222466 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] [NFC] Merging InstPrinter directory in to MCTargetDesc since they have...
Colin LeMahieu [Thu, 20 Nov 2014 21:56:35 +0000 (21:56 +0000)]
[Hexagon] [NFC] Merging InstPrinter directory in to MCTargetDesc since they have a circular dependency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222458 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Remove JITEventListener::NotifyFunctionEmitted - this method is dead
Lang Hames [Thu, 20 Nov 2014 21:16:16 +0000 (21:16 +0000)]
[MCJIT] Remove JITEventListener::NotifyFunctionEmitted - this method is dead
now that the legacy JIT has been removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222453 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a trip-count overflow issue in LoopUnroll.
Michael Zolotukhin [Thu, 20 Nov 2014 20:19:55 +0000 (20:19 +0000)]
Fix a trip-count overflow issue in LoopUnroll.

Currently LoopUnroll generates a prologue loop before the main loop
body to execute first N%UnrollFactor iterations. Also, this loop is
used if trip-count can overflow - it's determined by a runtime check.

However, we've been mistakenly optimizing this loop to a linear code for
UnrollFactor = 2, not taking into account that it also serves as a safe
version of the loop if its trip-count overflows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222451 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCompilation test for PostOrderIterator.
Michael Ilseman [Thu, 20 Nov 2014 19:33:33 +0000 (19:33 +0000)]
Compilation test for PostOrderIterator.

If the template specialization for externally managed sets in
PostOrderIterator call too far out of sync with each other, this unit
test will fail to build. This is especially useful for developers who
may not build Clang (the only in-tree user) every time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222447 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUpdate template specialization to reflect API changes.
Michael Ilseman [Thu, 20 Nov 2014 19:33:30 +0000 (19:33 +0000)]
Update template specialization to reflect API changes.

po_iterator_storage's insertEdge was updated to reflect the API
changes from many of our insert methods in r222334, however the
template specialization for external storage was not updated. This
updates the specialization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222446 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoX86: use the correct alloca symbol for Windows Itanium
Saleem Abdulrasool [Thu, 20 Nov 2014 18:01:26 +0000 (18:01 +0000)]
X86: use the correct alloca symbol for Windows Itanium

Windows itanium targets the MSVCRT, and the stack probe symbol is provided by
MSVCRT.  This corrects the emission of stack probes on i686-windows-itanium.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222439 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake DWARFAcceleratorTable::dump() const.
Frederic Riss [Thu, 20 Nov 2014 16:21:11 +0000 (16:21 +0000)]
Make DWARFAcceleratorTable::dump() const.

As dump() methods  should be. To allow that, do not store the DWARFFormValue
objects used for the dump in the header data.

Per Alexey's suggestion!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222436 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd missing copyright headers.
Frederic Riss [Thu, 20 Nov 2014 16:21:06 +0000 (16:21 +0000)]
Add missing copyright headers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222435 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDo not create a replaceable Variables MDNode for function forward decls.
Frederic Riss [Thu, 20 Nov 2014 15:52:34 +0000 (15:52 +0000)]
Do not create a replaceable Variables MDNode for function forward decls.

These fields would need to be explicitly deleted before we RAUW the temporary
node anyway (this was done in cfe commit r222373). Instead, do not create
these useless nodes in the first place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222434 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMCJIT tests passing on ARM after r222414 fixed the relocation
Renato Golin [Thu, 20 Nov 2014 13:32:16 +0000 (13:32 +0000)]
MCJIT tests passing on ARM after r222414 fixed the relocation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222430 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert r222416, r222422, r222426: the former revision had problems and fixing them...
Timur Iskhodzhanov [Thu, 20 Nov 2014 12:36:43 +0000 (12:36 +0000)]
Revert r222416, r222422, r222426: the former revision had problems and fixing them introduced bugs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222428 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a typo
Timur Iskhodzhanov [Thu, 20 Nov 2014 11:48:58 +0000 (11:48 +0000)]
Fix a typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222426 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplifyCFG.cpp: Tweak to let msc17 compliant.
NAKAMURA Takumi [Thu, 20 Nov 2014 08:59:02 +0000 (08:59 +0000)]
SimplifyCFG.cpp: Tweak to let msc17 compliant.

  - Use LLVM_DELETED_FUNCTION.
  - Don't use member initializers.
  - Don't use initializer list.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222422 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplifyCFG: Refactor GatherConstantCompares() result in a struct
Mehdi Amini [Thu, 20 Nov 2014 06:51:02 +0000 (06:51 +0000)]
SimplifyCFG: Refactor GatherConstantCompares() result in a struct

Code seems cleaner and easier to understand this way

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222416 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ELF] Prevent ARM ELF object writer from generating deprecated relocation code R_ARM_...
Jyoti Allur [Thu, 20 Nov 2014 05:58:11 +0000 (05:58 +0000)]
[ELF] Prevent ARM ELF object writer from generating deprecated relocation code R_ARM_PLT32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222414 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a test for r221870
David Majnemer [Thu, 20 Nov 2014 05:32:10 +0000 (05:32 +0000)]
Add a test for r221870

bad-relocs.obj.coff-i386 has a relocation whose symbol index is outside
the symbol table.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222413 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a typo in a comment.
Craig Topper [Thu, 20 Nov 2014 05:22:37 +0000 (05:22 +0000)]
Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222412 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove unnecessary extra spaces from tablegen register info output.
Craig Topper [Thu, 20 Nov 2014 05:22:35 +0000 (05:22 +0000)]
Remove unnecessary extra spaces from tablegen register info output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222411 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse array_lengthof instead of sizeof(array)/sizeof(element) in a tablegen output.
Craig Topper [Thu, 20 Nov 2014 05:22:32 +0000 (05:22 +0000)]
Use array_lengthof instead of sizeof(array)/sizeof(element) in a tablegen output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222410 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove support for undocumented SpecialCaseList entries.
Alexey Samsonov [Thu, 20 Nov 2014 01:27:19 +0000 (01:27 +0000)]
Remove support for undocumented SpecialCaseList entries.

"global-init", "global-init-src" and "global-init-type" were originally
used to blacklist entities in ASan init-order checker. However, they
were never documented, and later were replaced by "=init" category.

Old blacklist entries should be converted as follows:
  * global-init:foo -> global:foo=init
  * global-init-src:bar -> src:bar=init
  * global-init-type:baz -> type:baz=init

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222401 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding A2_xor instruction with IR selection pattern and test.
Colin LeMahieu [Wed, 19 Nov 2014 23:22:23 +0000 (23:22 +0000)]
[Hexagon] Adding A2_xor instruction with IR selection pattern and test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222399 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "[Reassociate] As the expression tree is rewritten make sure the operands...
Chad Rosier [Wed, 19 Nov 2014 23:21:20 +0000 (23:21 +0000)]
Revert "[Reassociate] As the expression tree is rewritten make sure the operands are"

This reverts commit r222142.  This is causing/exposing an execution-time regression
in spec2006/gcc and coremark on AArch64/A57/Ofast.

Conflicts:

test/Transforms/Reassociate/optional-flags.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222398 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding A2_or instruction with IR selection pattern and test.
Colin LeMahieu [Wed, 19 Nov 2014 22:58:04 +0000 (22:58 +0000)]
[Hexagon] Adding A2_or instruction with IR selection pattern and test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222396 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agogold-plugin: Fix typo in error message
Duncan P. N. Exon Smith [Wed, 19 Nov 2014 22:39:21 +0000 (22:39 +0000)]
gold-plugin: Fix typo in error message

Spotted while reading code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222395 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[CMake] Update GetSVN.cmake to use LLVM version control helper scripts.
Jordan Rose [Wed, 19 Nov 2014 22:03:21 +0000 (22:03 +0000)]
[CMake] Update GetSVN.cmake to use LLVM version control helper scripts.

This allows the logic to work with Git, and also uses the variable names
to match what Clang is actually looking for.

This is a re-application of r190556 and r190808. This changes the interface
of GetSVN.cmake. Clang change to follow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222391 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoADT: windows itanium targets msvcrt.dll
Saleem Abdulrasool [Wed, 19 Nov 2014 21:55:31 +0000 (21:55 +0000)]
ADT: windows itanium targets msvcrt.dll

Update the isOSMSVCRT to correctly identify that windows-itanium uses
msvcrt.dll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222389 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTry to fix MSVS build after r222384. No intended behavior change.
Nico Weber [Wed, 19 Nov 2014 21:16:11 +0000 (21:16 +0000)]
Try to fix MSVS build after r222384. No intended behavior change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222386 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix Use-of-uninitialized-value for adrp_addr and adrp_inst in MachODump.cpp
Kevin Enderby [Wed, 19 Nov 2014 20:20:16 +0000 (20:20 +0000)]
Fix Use-of-uninitialized-value for adrp_addr and adrp_inst in MachODump.cpp

Fixes PR21607

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222385 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplifyCFG: turn recursive GatherConstantCompares into iterative
Mehdi Amini [Wed, 19 Nov 2014 20:09:11 +0000 (20:09 +0000)]
SimplifyCFG: turn recursive GatherConstantCompares into iterative

A long sequence of || or && could lead to a stack explosion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222384 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRegisterCoalescer: Improve debug messages
Matthias Braun [Wed, 19 Nov 2014 19:46:17 +0000 (19:46 +0000)]
RegisterCoalescer: Improve debug messages

- Show "Considering..." message after flipping so you actually see the final
  destination vreg as destination.
- Add a message on final join, so you can grep for "Success" messages to obtain
  a list of which register got merged with which.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222382 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a print and verify pass after the RegisterCoalescer
Matthias Braun [Wed, 19 Nov 2014 19:46:15 +0000 (19:46 +0000)]
Add a print and verify pass after the RegisterCoalescer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222381 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMachineVerifier: Report register for bad liveranges
Matthias Braun [Wed, 19 Nov 2014 19:46:13 +0000 (19:46 +0000)]
MachineVerifier: Report register for bad liveranges

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222380 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIntroduce register dump helper
Matthias Braun [Wed, 19 Nov 2014 19:46:11 +0000 (19:46 +0000)]
Introduce register dump helper

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222379 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAliasSet: Simplify mergeSetIn
David Majnemer [Wed, 19 Nov 2014 19:36:18 +0000 (19:36 +0000)]
AliasSet: Simplify mergeSetIn

No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222376 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86] Improved lowering of v4x32 build_vector dag nodes.
Andrea Di Biagio [Wed, 19 Nov 2014 19:34:29 +0000 (19:34 +0000)]
[X86] Improved lowering of v4x32 build_vector dag nodes.

This patch improves the lowering of v4f32 and v4i32 build_vector dag nodes
that are known to have at least two non-zero elements.

With this patch, a build_vector that performs a blend with zero is
converted into a shuffle. This is done to let the shuffle legalizer expand
the dag node in a optimal way. For example, if we know that a build_vector
performs a blend with zero, we can try to lower it as a movq/blend instead of
always selecting an insertps.

This patch also improves the logic that lowers a build_vector into a insertps
with zero masking. See for example the extra test cases added to test sse41.ll.

Differential Revision: http://reviews.llvm.org/D6311

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222375 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ADT] Fix PR20728 - Incorrect APFloat::fusedMultiplyAdd results for x86_fp80.
Lang Hames [Wed, 19 Nov 2014 19:15:41 +0000 (19:15 +0000)]
[ADT] Fix PR20728 - Incorrect APFloat::fusedMultiplyAdd results for x86_fp80.

As detailed at http://llvm.org/PR20728, due to an internal overflow in
APFloat::multiplySignificand the APFloat::fusedMultiplyAdd method can return
incorrect results for x87DoubleExtended (x86_fp80) values. This commonly
manifests as incorrect constant folding of libm fmal calls on x86. E.g.

fmal(1.0L, 1.0L, 3.0L) == 0.0L      (should be 4.0L)

This patch fixes PR20728 by adding an extra bit to the significand for
intermediate results of APFloat::multiplySignificand, avoiding the overflow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222374 91177308-0d34-0410-b5e6-96231b3b80d8