Tamas Berghammer [Fri, 22 May 2015 10:01:56 +0000 (10:01 +0000)]
Revert "[X86] Fix a variable name for r237977 so that it works with every compilers."
Revert "[X86] Refactor the prologue emission to prepare for shrink-wrapping."
Puyan Lotfi [Fri, 22 May 2015 08:11:26 +0000 (08:11 +0000)]
Compile time improvements to VirtRegRewriter.
This change to VirtRegRewriter::addMBBLiveIns adds live-in registers for each
MachineBasicBlock's LiveIns set without isLiveIn checks as they are being added
because doing so is expensive. After all live-in registers are added, the LiveIn
vectors are sorted and uniqued.
Shave a pointer off of `MCSymbolName` by storing `StringMapEntry<bool>*`
instead of `StringRef`. This brings `sizeof(MCSymbol)` down to 64 on
64-bit platforms, a nice round number. My profile showed memory
dropping from 914 MB down to 908 MB, roughly 0.7%. Other than memory
usage, no functionality change here.
(I'm looking at `llc` memory usage on `verify-uselistorder.lto.opt.bc`;
see r236629 for details.)
SDAG: Give SDDbgValues their own allocator (and reset it)
Previously `SDDbgValue`s used the general allocator that lives for all
of `SelectionDAG`. Instead, give them their own allocator, and reset it
whenever `SDDbgInfo::clear()` is called, plugging a spiritual leak.
This drops `SelectionDAGBuilder::visitIntrinsicCall()` off of my heap
profile (was at around 2% of `llc` for codegen of `-flto -g`). Thanks
to Pete Cooper for spotting the problem and suggesting the fix.
Cleanup how `SDDbgValue` is initialized, and rearrange the fields to
save two pointers in the struct layout. No real functionality change
though (and I doubt the memory savings would show up in a profile).
David Majnemer [Fri, 22 May 2015 03:56:46 +0000 (03:56 +0000)]
[InstSimplify] Handle some overflow intrinsics in InstSimplify
This change does a few things:
- Move some InstCombine transforms to InstSimplify
- Run SimplifyCall from within InstCombine::visitCallInst
- Teach InstSimplify to fold [us]mul_with_overflow(X, undef) to 0.
Chandler Carruth [Fri, 22 May 2015 03:07:28 +0000 (03:07 +0000)]
[Unroll] Replace a hand-wavy FIXME with a FIXME that explains the actual
problem instead of suggesting doing something that is trivial to do but
incorrect given the current design of the libraries.
Chandler Carruth [Fri, 22 May 2015 03:02:22 +0000 (03:02 +0000)]
[Unroll] Extract the logic for caching SCEV-modeled GEPs with their
simplified model for use simulating each iteration into a separate
helper function that just returns the cache.
Building this cache had nothing to do with the rest of the unroll
analysis and so this removes an unnecessary coupling, etc. It should
also make it easier to think about the concept of providing fast cached
access to basic SCEV models as an orthogonal concept to the overall
unroll simulation.
I'd really like to see this kind of caching logic folded into SCEV
itself, it seems weird for us to provide it at this layer rather than
making repeated queries into SCEV fast all on their own.
Philip Reames [Fri, 22 May 2015 02:14:05 +0000 (02:14 +0000)]
[LICM] Sinking doesn't involve the preheader
PR23608 pointed out that using the preheader to gain a context instruction isn't always legal because a loop might not have a preheader. When looking into that, I realized that using the preheader to determine legality for sinking is questionable at best. Given no test covers that case and the original commit didn't seem to intend it, I restructured the code to only ask context sensative queries for hoising of loads and stores. This is effectively a partial revert of 237593.
Daniel Berlin [Fri, 22 May 2015 00:13:05 +0000 (00:13 +0000)]
MergedLoadStoreMotion preserves MemoryDependenceAnalysis, it does not require it.
(It already was coded assuming it can sometimes be null, so no other changes are necessary)
Quentin Colombet [Fri, 22 May 2015 00:12:31 +0000 (00:12 +0000)]
[X86] Refactor the prologue emission to prepare for shrink-wrapping.
- Add a late pass to expand pseudo instructions (tail call and EH returns).
Instead of doing it in the prologue emission.
- Factor some static methods in X86FrameLowering to ease code sharing.
David Blaikie [Fri, 22 May 2015 00:05:05 +0000 (00:05 +0000)]
unique_ptrify LoadedObjectInfo::clone
As noted in the original review, this is unused in tree & is used by
Julia... that's problematic. This API coudl easily be deleted/modified
by accident without any validation that it remains correct.
Hal Finkel [Thu, 21 May 2015 23:45:49 +0000 (23:45 +0000)]
[PPC] Correct iterator bug in PPCTLSDynamicCall
Unfortunately, I can't reduce a small test case for this (although compiling
mpfr-3.1.2 with -O2 -mcpu=a2 would fairly reliably trigger a crash), but the
problem is fairly clear (at least once you know you're looking for one). If the
TLS instruction being replaced was at the end of the block, we'd increment the
iterator past it (so it would then point to MBB.end()), and then we'd increment
it again as part of the for statement, thus overrunning the end of the list.
Don't do that.
Quentin Colombet [Thu, 21 May 2015 21:41:55 +0000 (21:41 +0000)]
[InlineSpiller] Fix rematerialization for bundles.
Prior to this patch, we could update the operand of another MI in the same
bundle.
Longer version:
Before InlineSpiller rematerializes a vreg, it iterates over operands of each MI
in a bundle, collecting all (MI, OpNo) pairs that reference that vreg.
Then if it does rematerialize, it goes through the pair list and replaces the
operands with the new (rematerialized) vreg. The problem is, it tries to
replace all of these operands in the main MI ! This works fine for single MIs.
However, if we are processing a bundle of MIs and the list contains multiple
pairs - the rematerialization will either crash trying to access a non-existing
operand of the main MI, or silently corrupt one of the existing ones. It will
also ignore other MIs in the bundle.
The obvious fix is to use the MI pointers saved in collected (MI, OpNo) pairs.
This must have been the original intent of the pair list but somehow these
pointers got lost.
Keno Fischer [Thu, 21 May 2015 21:24:32 +0000 (21:24 +0000)]
Make it easier to use DwarfContext with MCJIT
Summary:
This supersedes http://reviews.llvm.org/D4010, hopefully properly
dealing with the JIT case and also adds an actual test case.
DwarfContext was basically already usable for the JIT (and back when
we were overwriting ELF files it actually worked out of the box by
accident), but in order to resolve relocations correctly it needs
to know the load address of the section.
Rather than trying to get this out of the ObjectFile or requiring
the user to create a new ObjectFile just to get some debug info,
this adds the capability to pass in that info directly.
As part of this I separated out part of the LoadedObjectInfo struct
from RuntimeDyld, since it is now required at a higher layer.
David Blaikie [Thu, 21 May 2015 21:12:43 +0000 (21:12 +0000)]
[opaque pointer type] Allow gep_type_iterator to work with the pointee type from the GEP instruction
The raw non-instruction/constant form of this is still relying on being
able to access the pointee type from a pointer type - those will be
cleaned up later. For now, just focus on the cases where the pointee
type is easily accessible.
Alex Lorenz [Thu, 21 May 2015 20:54:45 +0000 (20:54 +0000)]
Resubmit r237708 (MIR Serialization: print and parse LLVM IR using MIR format).
This commit is a 2nd attempt at committing the initial MIR serialization patch.
The first commit (r237708) made the incremental buildbots unstable and was
reverted in r237730. The original commit didn't add a terminating null
character to the LLVM IR source which was passed to LLParser, and this
sometimes caused the test 'llvmIR.mir' to fail with a parsing error because
the LLVM IR source didn't have a null character immediately after the end
and thus LLLexer encountered some garbage characters that ultimately caused
the error.
This commit also includes the other test fixes I committed in
r237712 (llc path fix) and r237723 (remove target triple) which
also got reverted in r237730.
--Original Commit Message--
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
Bill Schmidt [Thu, 21 May 2015 20:48:49 +0000 (20:48 +0000)]
[PPC64] Handle vpkudum mask pattern correctly when vpkudum isn't available
My recent patch to add support for ISA 2.07 vector pack/unpack
instructions didn't properly check for availability of the vpkudum
instruction when recognizing it as a special vector shuffle case.
This causes us to leave the vector shuffle in place (rather than
converting it to a vector permute) so that it can be recognized later
as a vpkudum, but that pattern is invalid for processors prior to
POWER8. Thus LLVM crashes with an "unable to select" message. We
observed this since one of our buildbots is configured to generate
code for a POWER7.
This patch fixes the problem by checking for availability of the
vpkudum instruction during custom lowering of vector shuffles.
I've added a test case variant for the vpkudum pattern when the
instruction isn't available.
Adrian Prantl [Thu, 21 May 2015 20:37:30 +0000 (20:37 +0000)]
IR / debug info: Add a DWOId field to DICompileUnit,
so DWARF skeleton CUs can be expression in IR. A skeleton CU is a
(typically empty) DW_TAG_compile_unit that has a DW_AT_(GNU)_dwo_name and
a DW_AT_(GNU)_dwo_id attribute. It is used to refer to external debug info.
This is a prerequisite for clang module debugging as discussed in
http://lists.cs.uiuc.edu/pipermail/cfe-dev/2014-November/040076.html.
In order to refer to external types stored in split DWARF (dwo) objects,
such as clang modules, we need to emit skeleton CUs, which identify the
dwarf object (i.e., the clang module) by filename (the SplitDebugFilename)
and a hash value, the dwo_id.
This patch only contains the IR changes. The idea is that a CUs with a
non-zero dwo_id field will be emitted together with a DW_AT_GNU_dwo_name
and DW_AT_GNU_dwo_id attribute.
Hal Finkel [Thu, 21 May 2015 20:30:23 +0000 (20:30 +0000)]
[PPC/LoopUnrollRuntime] Don't avoid high-cost trip count computation on the PPC/A2
On X86 (and similar OOO cores) unrolling is very limited, and even if the
runtime unrolling is otherwise profitable, the expense of a division to compute
the trip count could greatly outweigh the benefits. On the A2, we unroll a lot,
and the benefits of unrolling are more significant (seeing a 5x or 6x speedup
is not uncommon), so we're more able to tolerate the expense, on average, of a
division to compute the trip count.
Alex Lorenz [Thu, 21 May 2015 19:45:02 +0000 (19:45 +0000)]
YAML: Null terminate block scalar's value.
The commit null terminates the string value in the `yaml::BlockScalarNode`
class.
This change is motivated by the initial MIR serialization commit (r237708)
that I reverted in r237730 because the LLVM IR source from the block
scalar node wasn't terminated by a null character and thus the buildbots
failed on one testcase sometimes. This change enables me to recommit
the reverted commit.
Derek Schuff [Thu, 21 May 2015 19:40:19 +0000 (19:40 +0000)]
Fix StreamingMemoryObject to respect known object size.
The existing code for method StreamingMemoryObject.fetchToPos does not respect
the corresonding call to setKnownObjectSize(). As a result, it allows the
StreamingMemoryObject to read bytes past the object size.
This patch provides a test case, and code to fix the problem.
Patch by Karl Schimpf
Differential Revision: http://reviews.llvm.org/D8931
Nemanja Ivanovic [Thu, 21 May 2015 19:32:49 +0000 (19:32 +0000)]
Add support for VSX scalar single-precision arithmetic in the PPC target
http://reviews.llvm.org/D9891
Following up on the VSX single precision loads and stores added earlier, this
adds support for elementary arithmetic operations on single precision values
in VSX registers. These instructions utilize the new VSSRC register class.
Instructions added:
xsaddsp
xsdivsp
xsmulsp
xsresp
xsrsqrtesp
xssqrtsp
xssubsp
Rafael Espindola [Thu, 21 May 2015 19:20:38 +0000 (19:20 +0000)]
Move alignment from MCSectionData to MCSection.
This starts merging MCSection and MCSectionData.
There are a few issues with the current split between MCSection and
MCSectionData.
* It optimizes the the not as important case. We want the production
of .o files to be really fast, but the split puts the information used
for .o emission in a separate data structure.
* The ELF/COFF/MachO hierarchy is not represented in MCSectionData,
leading to some ad-hoc ways to represent the various flags.
* It makes it harder to remember where each item is.
The attached patch starts merging the two by moving the alignment from
MCSectionData to MCSection.
Most of the patch is actually just dropping 'const', since
MCSectionData is mutable, but MCSection was not.
AVX-512: Enabled SSE intrinsics on AVX-512.
Predicate UseAVX depricates pattern selection on AVX-512.
This predicate is necessary for DAG selection to select EVEX form.
But mapping SSE intrinsics to AVX-512 instructions is not ready yet.
So I replaced UseAVX with HasAVX for intrinsics patterns.
Renato Golin [Thu, 21 May 2015 13:52:20 +0000 (13:52 +0000)]
Make Triple::parseARMArch use ARMTargetParser
Simplifying Triple::parseARMArch, leaving all the parsing to ARMTargetParser.
This commit also adds AArch64 detection to ARMTargetParser canonicalization,
and a two RedHat arch names (v{6,7}hl, meaning hard-float / little-endian).
Adding enough unit tests to cover the basics. Clang checks fine.
Artur Pilipenko [Thu, 21 May 2015 12:51:38 +0000 (12:51 +0000)]
Fix memory-dereferenceable.ll test
One of the testcases introduced by D9365 had incorrect !dereferenceable metadata on load. It must fail but it doesn't due to incorrect order of CHECK/CHECK-NOT commands in test. Fixed both.
Simon Pilgrim [Thu, 21 May 2015 10:05:03 +0000 (10:05 +0000)]
[X86][SSE] Improve support for 128-bit vector sign extension
This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization).
It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner.
Create a low-overhead path for `EmitLabelDifference()` that emits a
emits an absolute number when (1) the output is an object stream and (2)
the two symbols are in the same data fragment.
This drops memory usage on Mach-O from 975 MB down to 919 MB (5.8%).
The only call is when `!doesDwarfUseRelocationsAcrossSections()` --
i.e., on Mach-O -- since otherwise an absolute offset from the start of
the section needs a relocation. (`EmitLabelDifference()` is cheaper on
ELF anyway, since it creates 1 fewer temp symbol, and it gets called far
less often. It's not clear to me if this is even a bottleneck there.)
(I'm looking at `llc` memory usage on `verify-uselistorder.lto.opt.bc`;
see r236629 for details.)
Ahmed Bougacha [Thu, 21 May 2015 01:43:39 +0000 (01:43 +0000)]
[MemCpyOpt] Do move the memset, but look at its dest's dependencies.
In effect a partial revert of r237858, which was a dumb shortcut.
Looking at the dependencies of the destination should be the proper
fix: if the new memset would depend on anything other than itself,
the transformation isn't correct.
MC: Simplify MCSymbolData initialization and remove MCSymbol pointer
Finally remove the `MCSymbolData::Symbol` pointer. It was still being
used to track whether `MCSymbolData` had been initialized, but this is
better tracked by the bitfield in `MCSymbol`.
The only caller of `MCSymbolData::initialize()` was `MCAssembler`, which
(other than `Symbol`) passed in all-0 values. Replace all that
indirection with a default constructor.
The main point is a cleanup (and there's more cleanup to do), but there
are also some small memory savings. I measured ~989 MB down to ~975 MB,
cutting a little over 1% off the top of `llc`.
(I'm looking at `llc` memory usage on `verify-uselistorder.lto.opt.bc`;
see r236629 for details.)
Make the back-pointer from `MCSymbolData` to `MCSymbol` private,
preparing to remove the back pointer entirely. I've already updated all
the users, although for now it's still used to indicate whether
`MCSymbol::Data` has been initialized.
MC: Remove last use of MCSymbolData::getSymbol(), NFC
Remove the last use of `MCSymbolData::getSymbol()`. There's some
*really* hairy stuff going on in `MachObjectWriter::WriteNList()` that I
want to come back to. In particular, it updates `Symbol` to point at
its aliasee (if any), but leaves `Data` behind, and it's not clear
whether everything makes sense there.
For now I've left the logic unchanged by adding `OrigSymbol` and moving
the FIXME from r237750 up a bit higher. I've filed PR23598 to track
looking into this.
Ahmed Bougacha [Wed, 20 May 2015 23:55:16 +0000 (23:55 +0000)]
[MemCpyOpt] Don't move the memset when optimizing memset+memcpy.
Fixes PR23599, another miscompile introduced by r235232: when there is
another dependency on the destination of the created memset (i.e., the
part of the original destination that the memcpy doesn't depend on)
between the memcpy and the original memset, we would insert the created
memset after the memcpy, and thus after the other dependency.
Instead, insert the created memset right after the old one.
Pete Cooper [Wed, 20 May 2015 22:51:27 +0000 (22:51 +0000)]
Don't generate comments in the DebugLocStream unless required. NFC.
The ByteStreamer here wasn't taking account of whether the asm streamer was text based and verbose. Only with that combination should we emit comments.
This change makes sure that we only actually convert a Twine to a string using Twine::str() if we need the comment. This saves about 10000 small allocations on a test case involving the verify-use_list-order bitcode going through llc with debug info.
Note, this is NFC as the comments would ultimately never be emitted unless required.
Davide Italiano [Wed, 20 May 2015 21:40:38 +0000 (21:40 +0000)]
[Target/ARM] Only enable OptimizeBarrierPass at -O1 and above.
Ideally this is going to be and LLVM IR pass (shared, among others
with AArch64), but for the time being just enable it if consumers
ask us for optimization and not unconditionally.
Alex Lorenz [Wed, 20 May 2015 20:41:27 +0000 (20:41 +0000)]
AsmParser: Require a terminating null character when creating memory buffer.
This commit modifies the memory buffer creation in the AsmParser library so
that it requires a terminating null character. The LLLexer in the AsmParser
library checks for EOF only when it sees a null character, thus it would
be best to require it when creating a memory buffer so that the memory
buffer constructor can verify that a terminating null character is indeed
present.
Reviewers: Duncan P. N. Exon Smith, Matthias Braun
projects / llvm / log