Pete Cooper [Wed, 20 May 2015 19:12:14 +0000 (19:12 +0000)]
Use a SmallString buffer instead of a std::string for debug info path lookup. NFC.
This code appends the filename to the directory then looks that up in a StringMap. We should be using the existing Twine::toStringRef method instead of Twine::str() as most times we'll succeed in the lookup.
Its possible that we should also consider allowing StringMap to lookup a key using a Twine in addition to a StringRef but that would complicate the code with little known benefit above and beyond this change.
This saves 170k temporary allocations when running llc on the verify_use_list_order bitcode with debug info for x86.
Matthias Braun [Wed, 20 May 2015 18:54:02 +0000 (18:54 +0000)]
DAGCombiner: Continue combining if FoldConstantArithmetic() fails.
DAG.FoldConstantArithmetic() can fail even though both operands are
Constants if OpaqueConstants are involved. Continue trying other combine
possibilities in tis case.
James Molloy [Wed, 20 May 2015 18:41:25 +0000 (18:41 +0000)]
Reapply r237539 with a fix for the Chromium build.
Make sure if we're truncating a constant that would then be sign extended
that the sign extension of the truncated constant is the same as the
original constant.
> Canonicalize min/max expressions correctly.
>
> This patch introduces a canonical form for min/max idioms where one operand
> is extended or truncated. This often happens when the other operand is a
> constant. For example:
>
> %1 = icmp slt i32 %a, i32 0
> %2 = sext i32 %a to i64
> %3 = select i1 %1, i64 %2, i64 0
>
> Would now be canonicalized into:
>
> %1 = icmp slt i32 %a, i32 0
> %2 = select i1 %1, i32 %a, i32 0
> %3 = sext i32 %2 to i64
>
> This builds upon a patch posted by David Majenemer
> (https://www.marc.info/?l=llvm-commits&m=143008038714141&w=2). That pass
> passively stopped instcombine from ruining canonical patterns. This
> patch additionally actively makes instcombine canonicalize too.
>
> Canonicalization of expressions involving a change in type from int->fp
> or fp->int are not yet implemented.
Pawel Bylica [Wed, 20 May 2015 17:21:09 +0000 (17:21 +0000)]
Fix icmp lowering
Summary:
During icmp lowering it can happen that a constant value can be larger than expected (see the code around the change).
APInt::getMinSignedBits() must be checked again as the shift before can change the constant sign to positive.
I'm not sure it is the best fix possible though.
r237490 accidentally dropped MCSymbolData from the MCAssembler dump.
Add it back underneath the MCSymbol dump. Remove the MCSymbol dump from
MCSymbolData, since this would cause an infinite co-recursion, and
besides, that back pointer is going away.
Renato Golin [Wed, 20 May 2015 15:05:07 +0000 (15:05 +0000)]
Get Triple::getARMCPUForArch() to use TargetParser
First ARMTargetParser FIXME, conservatively changing the way we parse CPUs
in the back-end. Still not perfect, with a lot of special cases, but moving
towards a more generic solution.
Moving all logic to the target parser made some unwritten assumptions
about architectures in Clang to break. I've added a lot of architectures
required by Clang, and default to CPUs that Clang believes it should
(and I agree).
I've also added a lot of unit tests, with the correct CPU for each
architecture, and Clang seems to be working correctly, too.
It also became clear that using "unsigned ID" as the argument for the get
methods makes it hard to know what ID, so I also changed the argument names
to match the enum type names.
AVX-512: fixed algorithm of building vectors of i1 elements
fixed extract-insert i1 element,
load i1, zextload i1 should be with "and $1, %reg" to prevent loading garbage.
added a bunch of new tests.
Igor Laevsky [Wed, 20 May 2015 11:37:25 +0000 (11:37 +0000)]
[StatepointLowering] Support of the gc.relocates for invoke statepoints.
This change implements support for lowering of the gc.relocates tied to the invoke statepoint.
This is acomplished by storing frame indices of the lowered values in "StatepointRelocatedValues" map inside FunctionLoweringInfo instead of storing them in per-basic block structure StatepointLowering.
After this change StatepointLowering is used only during "LowerStatepoint" call and it is not necessary to store it as a field in SelectionDAGBuilder anymore.
Swaroop Sridhar [Wed, 20 May 2015 01:07:23 +0000 (01:07 +0000)]
Add a GCStrategy for CoreCLR
This change adds a new GC strategy for supporting the CoreCLR runtime.
This strategy is currently identical to Statepoint-example GC,
but is necessary for several upcoming changes specific to CoreCLR, such as:
1. Base-pointers not explicitly reported for interior pointers
2. Different format for stack-map encoding
3. Location of Safe-point polls: polls are only needed before loop-back edges and before tail-calls (not needed at function-entry)
4. Runtime specific handshake between calls to managed/unmanaged functions.
Philip Reames [Tue, 19 May 2015 23:40:11 +0000 (23:40 +0000)]
[PlaceSafepoints] Stop special casing some intrinsics
We were special casing a handful of intrinsics as not needing a safepoint before them. After running into another valid case - memset - I took a closer look and realized that almost no intrinsics need to have a safepoint poll before them. Restructure the code to make that apparent so that we stop hitting these bugs. The only intrinsics which need a safepoint poll before them are ones which can run arbitrary code.
Reid Kleckner [Tue, 19 May 2015 23:28:23 +0000 (23:28 +0000)]
Try to fix the build with MSVC 2015 by disabling sized deallocation
I can't actually test this properly because uninstalling MSVC 2015 CTP 6
and reinstalling the 2015 RC takes hours. I can only verify that this
doesn't mess up MSVC 2013 and 2015 CTP 6 builds, which is what I've
done.
Alexey Samsonov [Tue, 19 May 2015 21:54:32 +0000 (21:54 +0000)]
[DWARF parser] Make DWARF parser more robust against missing compile/type units.
DWARF standard claims that each compilation/type unit header in
.debug_info/.debug_types section must be followed by corresponding
compile/type unit DIE, possibly with its children. Two situations
are possible:
* compile/type unit DIE is missing because DWARF producer failed to
emit it.
* DWARF parser failed to parse unit DIE correctly, for instance if it
contains some unsupported attributes (see r237721, for instance).
In either of these cases, the library, and the tools that use it
(llvm-dwarfdump, llvm-symbolizer) should not crash. Insert appropriate
checks to protect against this.
Alex Lorenz [Tue, 19 May 2015 21:41:28 +0000 (21:41 +0000)]
Revert r237708 (MIR serialization) - incremental buildbots became unstable.
The incremental buildbots entered a pass-fail cycle where during the fail
cycle one of the tests from this commit fails for an unknown reason. I
have reverted this commit and will investigate the cause of this problem.
David Blaikie [Tue, 19 May 2015 21:31:18 +0000 (21:31 +0000)]
[opaque pointer type] Provide the ability to pass an explicit type when creating a CallInst through IRBuilder
A use for this will be added to Clang shortly. I haven't hit the
specific cases I want to cleanup in LLVM just yet.
A utility that avoids the need for this in direct calls to
llvm::Functions will be added too - since in that case the type can be
retrieved from the llvm::Function directly.
Alexey Samsonov [Tue, 19 May 2015 20:29:28 +0000 (20:29 +0000)]
[DWARF parser] Add basic support for DWZ DWARF multifile extensions.
This change implements basic support for DWARF alternate sections
proposal: http://www.dwarfstd.org/ShowIssue.php?issue=120604.1&type=open
LLVM tools now understand new forms: DW_FORM_GNU_ref_alt and
DW_FORM_GNU_strp_alt, which are used as references to .debug_info and
.debug_str sections respectively, stored in a separate file, and
possibly shared between different executables / shared objects.
llvm-dwarfdump and llvm-symbolizer don't yet know how to access this
alternate debug file (usually pointed by .gnu_debugaltlink section),
but they can at lease properly parse and dump regular files, which
refer to it.
This change should fix crashes of llvm-dwarfdump and llvm-symbolizer on
files produced by running "dwz" tool. Such files are already installed
on some modern Linux distributions.
Alex Lorenz [Tue, 19 May 2015 18:17:39 +0000 (18:17 +0000)]
MIR Serialization: print and parse LLVM IR using MIR format.
This commit is the initial commit for the MIR serialization project.
It creates a new library under CodeGen called 'MIR'. This new
library adds a new machine function pass that prints out the LLVM IR
using the MIR format. This pass is then added as a last pass when a
'stop-after' option is used in llc. The new library adds the initial
functionality for parsing of MIR files as well. This commit also
extends the llc tool so that it can recognize and parse MIR input files.
Reviewers: Duncan P. N. Exon Smith, Matthias Braun, Philip Reames
Wei Mi [Tue, 19 May 2015 16:09:11 +0000 (16:09 +0000)]
Remove the InstructionSimplifierPass immediately after InstructionCombiningPass.
InstructionCombiningPass was added after LoopUnrollPass in r237395. Because
InstructionCombiningPass is strictly more powerful than InstructionSimplifierPass,
remove the unnecessary InstructionSimplifierPass.
Igor Laevsky [Tue, 19 May 2015 15:59:05 +0000 (15:59 +0000)]
[RewriteStatepointsForGC] For some values (like gep's and bitcasts) it's cheaper to clone them after statepoint than to emit proper relocates for them. This change implements this logic. There is alredy similar optimization in CodeGenPrepare, but doing so during RewriteStatepointsForGC allows to capture more opprtunities such as relocates in loops and longer instruction chains.
Daniel Sanders [Tue, 19 May 2015 12:24:52 +0000 (12:24 +0000)]
[mips] Correct and improve special-case shuffle instructions.
Summary:
The documentation writes vectors highest-index first whereas LLVM-IR writes
them lowest-index first. As a result, instructions defined in terms of
left_half() and right_half() had the halves reversed.
In addition to correcting them, they have been improved to allow shuffles
that use the same operand twice or in reverse order. For example, ilvev
used to accept masks of the form:
<0, n, 2, n+2, 4, n+4, ...>
but now accepts:
<0, 0, 2, 2, 4, 4, ...>
<n, n, n+2, n+2, n+4, n+4, ...>
<0, n, 2, n+2, 4, n+4, ...>
<n, 0, n+2, 2, n+4, 4, ...>
One further improvement is that splati.[bhwd] is now the preferred instruction
for splat-like operations. The other special shuffles are no longer used
for splats. This lead to the discovery that <0, 0, ...> would not cause
splati.[hwd] to be selected and this has also been fixed.
This fixes the enc-3des test from the test-suite on Mips64r6 with MSA.
[X86] ABI change for x86-32: pass 3 vector arguments in-register instead of 4, except on Darwin.
This changes the ABI used on 32-bit x86 for passing vector arguments.
Historically, clang passes the first 4 vector arguments in-register, and additional vector arguments on the stack, regardless of platform. That is different from the behavior of gcc, icc, and msvc, all of which pass only the first 3 arguments in-register.
The 3-register convention is documented, unofficially, in Agner's calling convention guide, and, officially, in the recently released version 1.0 of the i386 psABI.
Darwin is kept as is because the OS X ABI Function Call Guide explicitly documents the current (4-register) behavior.
Matthias Braun [Tue, 19 May 2015 01:40:21 +0000 (01:40 +0000)]
SelectionDAG: Cleanup and simplify FoldConstantArithmetic
This cleans up the FoldConstantArithmetic code by factoring out the case
of two ConstantSDNodes into an own function. This avoids unnecessary
complexity for many callers who already have ConstantSDNode arguments.
This also avoids an intermeidate SmallVector datastructure and a loop
over that datastructure.
Pete Cooper [Tue, 19 May 2015 00:24:26 +0000 (00:24 +0000)]
Store intrinsic ID by value in Function instead of a string lookup. NFC.
On 64-bit targets, Function has 4-bytes of padding in its struct layout.
This uses the space for the intrinsic ID. It is set and recalculated whenever the function name is set. This is similar to the current behavior which clears the function from the intrinsic ID cache when its renamed.
The intrinsic cache itself is removed as the only purpose was to speedup calls to getIntrinsicID() which now just reading the new field in the struct.
Tim Northover [Mon, 18 May 2015 22:07:20 +0000 (22:07 +0000)]
AArch64: work around ld64 bug more aggressively.
ld64 currently mishandles internal pointer relocations (i.e.
ARM64_RELOC_UNSIGNED referred to by section & offset rather than symbol). The
existing __cfstring clause was an early discovery and workaround for this, but
the problem is wider and we should avoid such relocations wherever possible for
now.
This code should be reverted to allowing internal relocations as soon as
possible.
projects / llvm / log