Hal Finkel [Thu, 11 Sep 2014 08:40:17 +0000 (08:40 +0000)]
[AlignmentFromAssumptions] Don't crash just because the target is 32-bit
We used to crash processing any relevant @llvm.assume on a 32-bit target
(because we'd ask SE to subtract expressions of differing types). I've copied
our 'simple.ll' test, but with the data layout from arm-linux-gnueabihf to get
some meaningful test coverage here.
David Blaikie [Wed, 10 Sep 2014 23:54:45 +0000 (23:54 +0000)]
shared_ptrify ownershp of PoolEntries in PBQP's CostPool
Leveraging both intrusive shared_ptr-ing (std::enable_shared_from_this)
and shared_ptr<T>-owning-U (to allow external users to hold
std::shared_ptr<CostT> while keeping the underlying PoolEntry alive).
The intrusiveness could be removed if we had a weak_set that implicitly
removed items from the set when their underlying data went away.
This /might/ fix an existing memory leak reported by LeakSanitizer in
r217504.
Matt Arsenault [Wed, 10 Sep 2014 23:26:16 +0000 (23:26 +0000)]
R600/SI: Report offset in correct units for st64 DS instructions
Need to convert the 64 element offset into bytes, not just the element
size like the normal case instructions.
Noticed by inspection. This can't be hit now because
st64 instructions aren't emitted during instruction selection,
and the post-RA scheduler isn't enabled.
Rafael Espindola [Wed, 10 Sep 2014 21:27:43 +0000 (21:27 +0000)]
Add doInitialization/doFinalization to DataLayoutPass.
With this a DataLayoutPass can be reused for multiple modules.
Once we have doInitialization/doFinalization, it doesn't seem necessary to pass
a Module to the constructor.
Overall this change seems in line with the idea of making DataLayout a required
part of Module. With it the only way of having a DataLayout used is to add it
to the Module.
Hal Finkel [Wed, 10 Sep 2014 21:05:52 +0000 (21:05 +0000)]
[AlignmentFromAssumptions] Don't divide by zero for unknown starting alignment
The routine that determines an alignment given some SCEV returns zero if the
answer is unknown. In a case where we could determine the increment of an
AddRec but not the starting alignment, we would compute the integer modulus by
zero (which is illegal and traps). Prevent this by returning early if either
the start or increment alignment is unknown (zero).
The increase of the interleave factor to 4 has side-effects
like performance losses eg. due to reminder loops being executed
more frequently and may increase code size. It requires more
analysis and careful heuristic tuning. Expect double digit gains
in small benchmarks like lowercase.c and losses in puzzle.c.
Summary:
Make CallingConv::ID a plain unsigned instead of enum with a
fixed set of valus. LLVM IR allows arbitraty calling conventions (you are
free to write cc12345), and loading them as enum is an undefined
behavior. This was reported by UBSan.
[AArch 64] Use a constant pool load for weak symbol references when
using static relocation model and small code model.
Summary: currently we generate GOT based relocations for weak symbol
references regardless of the underlying relocation model. This should
be change so that in static relocation model we use a constant pool
load instead.
David Majnemer [Wed, 10 Sep 2014 12:51:52 +0000 (12:51 +0000)]
Object: Add support for bigobj
This adds support for reading the "bigobj" variant of COFF produced by
cl's /bigobj and mingw's -mbig-obj.
The most significant difference that bigobj brings is more than 2**16
sections to COFF.
bigobj brings a few interesting differences with it:
- It doesn't have a Characteristics field in the file header.
- It doesn't have a SizeOfOptionalHeader field in the file header (it's
only used in executable files).
- Auxiliary symbol records have the same width as a symbol table entry.
Since symbol table entries are bigger, so are auxiliary symbol
records.
Dan Liew [Wed, 10 Sep 2014 10:18:59 +0000 (10:18 +0000)]
Attempt to fix PR20884
This fixes the generation of broken LLVMExports.cmake file by
the Autoconf/Makefile build system when --enable-shared is passed to
configure.
When --enable_shared is passed the Makefile.rules does not set the
LLVMConfigLibs variable which cmake/modules/Makefile previously relied
on. Now it runs the llvm-config command itself to get the library names.
This still isn't perfect because the generated LLVM targets refer to the
static libraries and not the shared library but that is much larger
problem to fix.
MergeFunctions: FunctionPtr has been renamed to FunctionNode.
It's supposed to store additional pass information for current function here.
That was the reason for name change.
It appears that the -filename-equivalence option for testing llvm-cov
doesn't work correctly with -show-expansions. I'm reverting this test
to get the bots green while I look into fixing that.
David Blaikie [Tue, 9 Sep 2014 23:13:01 +0000 (23:13 +0000)]
Sink PrevCU updating into DwarfUnit::addRange to ensure consistency
So that the two operations in DwarfDebug couldn't get separated (because
I accidentally separated them in some work in progress), put them
together. While we're here, move DwarfUnit::addRange to
DwarfCompileUnit, since it's not relevant to type units.
David Blaikie [Tue, 9 Sep 2014 22:56:36 +0000 (22:56 +0000)]
Remove DwarfDebug::PrevSection, PrevCU is sufficient for handling address range holes.
PrevSection/PrevCU are used to detect holes in the address range of a CU
to ensure the DW_AT_ranges does not include those holes. When we see a
function with no debug info, though it may be in the same range as the
prior and subsequent functions, there should be a gap in the CU's
ranges. By setting PrevCU to null in that case, the range would not be
extended to cover the gap.
Handle common linkage correctly in the gold plugin.
This is the plugin version of pr20882.
This handles the case of every common symbol being in the IR. We will need some
support from gold to handle the case where some symbols are in ELF and some in
the IR.
Add a scheduling model for AMD 16H Jaguar (btver2).
This is a first pass at a scheduling model for Jaguar.
It's structured largely on the existing SandyBridge and SLM sched models.
Using this model, in addition to turning on the PostRA scheduler, results in
some perf wins on internal and 3rd party benchmarks. There's not much difference
in LLVM's test-suite benchmarking subset of tests.
[mips] Add assembler support for .set mips0 directive.
Summary:
This directive is used to reset the assembler options to their initial values.
Assembly programmers use it in conjunction with the ".set mipsX" directives.
This patch depends on the .set push/pop directive (http://reviews.llvm.org/D4821).
Diego Novillo [Tue, 9 Sep 2014 12:40:50 +0000 (12:40 +0000)]
Re-factor sample profile reader into lib/ProfileData.
Summary:
This patch moves the profile reading logic out of the Sample Profile
transformation into a generic profile reader facility in
lib/ProfileData.
The intent is to use this new reader to implement a sample profile
reader/writer that can be used to convert sample profiles from external
sources into LLVM.
This first patch introduces no functional changes. It moves the profile
reading code from lib/Transforms/SampleProfile.cpp into
lib/ProfileData/SampleProfReader.cpp.
In subsequent patches I will:
- Add a bitcode format for sample profiles to allow for more efficient
encoding of the profile.
- Add a writer for both text and bitcode format profiles.
- Add a 'convert' command to llvm-profdata to be able to convert between
the two (and serve as entry point for other sample profile formats).
Pavel Chupin [Tue, 9 Sep 2014 11:54:12 +0000 (11:54 +0000)]
[x32] Emit callq for CALLpcrel32
Summary:
In AT&T annotation for both x86_64 and x32 calls should be printed as
callq in assembly. It's only a matter of correct mnemonic, object output
is ok.
[mips] Add assembler support for .set push/pop directive.
Summary:
These directives are used to save the current assembler options (in the case of ".set push") and restore the previously saved options (in the case of ".set pop").
llvm-cov: Combine two types that were nearly identical (NFC)
llvm-cov had a SourceRange type that was nearly identical to a
CountedRegion except that it shaved off a couple of fields. There
aren't likely to be enough of these for the minor memory savings to be
worth the extra complexity here.
Bob Wilson [Tue, 9 Sep 2014 01:13:36 +0000 (01:13 +0000)]
Set trunc store action to Expand for all X86 targets.
When compiling without SSE2, isTruncStoreLegal(F64, F32) would return Legal, whereas with SSE2 it would return Expand. And since the Target doesn't seem to actually handle a truncstore for double -> float, it would just output a store of a full double in the space for a float hence overwriting other bits on the stack.
Hans Wennborg [Mon, 8 Sep 2014 20:24:10 +0000 (20:24 +0000)]
Fast-ISel: Remove dead code after falling back from selecting call instructions (PR20863)
Previously, fast-isel would not clean up after failing to select a call
instruction, because it would have called flushLocalValueMap() which moves
the insertion point, making SavedInsertPt in selectInstruction() invalid.
Fixing this by making SavedInsertPt a member variable, and having
flushLocalValueMap() update it.
This removes some redundant code at -O0, and more importantly fixes PR20863.
Be more careful in parsing Module::ModFlagBehavior value
to make sure we don't do invalid load of an enum. Share the
conversion code between llvm::Module implementation and the
verifier.
Matt Arsenault [Mon, 8 Sep 2014 15:07:33 +0000 (15:07 +0000)]
R600/SI: Fix assertion from copying a TargetGlobalAddress
Assert in scheduler from an inserted copy_to_regclass from
a constant.
This only seems to break sometimes when a constant initializer
address is forced into VGPRs in a non-entry block. No test
since the only case I've managed to hit only happens with a future
patch, and that case will also not be a problem once scalar instructions
are used in non-entry blocks.
projects / llvm / log