The bug manifests when there are two loads and two stores chained as follows in
a DAG,
(ld v3f32) -> (st f32) -> (ld v3f32) -> (st f32)
and the stores' values are extracted from the preceding vector loads.
MergeConsecutiveStores would replace the first store in the chain with the
merged vector store, which would create a cycle between the merged store node
and the last load node that appears in the chain.
This commits fixes the bug by replacing the last store in the chain instead.
Go bindings: make various DIBuilder arguments optional.
r234262 changed some code in DIBuilderBindings.cpp to use the unwrap function
to unwrap debug metadata. The problem with this is that unwrap asserts that
its argument is non-null, which is not what we want in a number of places
in DIBuilder where the argument is optional. This change makes certain
arguments optional by adding null checks in places where it is required,
fixing the llgo build.
Eliminate O(n^2) worst-case behavior in SSA construction
The code uses a priority queue and a worklist, which share the same
visited set, but the visited set is only updated when inserting into
the priority queue. Instead, switch to using separate visited sets
for the priority queue and worklist.
Adam Nemet [Wed, 8 Apr 2015 17:48:40 +0000 (17:48 +0000)]
[LoopAccesses] Allow analysis to complete in the presence of uniform stores
(Re-apply r234361 with a fix and a testcase for PR23157)
Both run-time pointer checking and the dependence analysis are capable
of dealing with uniform addresses. I.e. it's really just an orthogonal
property of the loop that the analysis computes.
Run-time pointer checking will only try to reason about SCEVAddRec
pointers or else gives up. If the uniform pointer turns out the be a
SCEVAddRec in an outer loop, the run-time checks generated will be
correct (start and end bounds would be equal).
In case of the dependence analysis, we work again with SCEVs. When
compared against a loop-dependent address of the same underlying object,
the difference of the two SCEVs won't be constant. This will result in
returning an Unknown dependence for the pair.
When compared against another uniform access, the difference would be
constant and we should return the right type of dependence
(forward/backward/etc).
The changes also adds support to query this property of the loop and
modify the vectorizer to use this.
Scott Douglass [Wed, 8 Apr 2015 17:18:28 +0000 (17:18 +0000)]
[ARM] make vminnm/vmaxnm work with ?le, ?ge and no-nans-fp-math
Because -menable-no-nans causes fcmp conditions to be rewritten
without 'o' or 'u' the recognition code in needs to cope. Also
extended it to handle 'le' and 'ge.
[mips] [IAS] Do not generate redundant move when expanding lw/sw with symbol.
Summary:
Even though there is no 2nd register operand in the "lw/sw $8, symbol" case, we still try to find one,
and we end up with $0, which makes us generate an unnecessary "addu $8, $8, $0" (a.k.a. "move $8, $8").
We can avoid this by checking if the 2nd register operand is different from $0, before generating the addu.
One could make the argument for writing it immediately after the ELF header,
but writing it in the middle of the sections like we were doing just makes
it harder for no reason.
[ARM][Debug Info] Restore emitting of .cfi_def_cfa_offset for functions without stack frame
Summary: Looks like new code from [[ http://reviews.llvm.org/rL222057 | rL222057 ]] doesn't account for early `return` in `ARMFrameLowering::emitPrologue`, which leads to loosing `.cfi_def_cfa_offset` directive for functions without stack frame.
[mips] [IAS] Remove AssemblerPredicate's from RelocPIC and RelocStatic.
Summary:
These AssemblerPredicate's are unnecessary and actually make some instructions unusable when assembling pre-MIPS32 ISAs.
For example, this was causing the IAS to reject the 'j' instruction for MIPS I-V.
Sanjoy Das [Wed, 8 Apr 2015 04:27:22 +0000 (04:27 +0000)]
[InstCombine] Refactor out OptimizeOverflowCheck. NFCI.
Summary:
This patch adds an enum `OverflowCheckFlavor` and a function
`OptimizeOverflowCheck`. This will allow InstCombine to optimize
overflow checks without directly introducing an intermediate call to the
`llvm.$op.with.overflow` instrinsics.
This specific change is a refactoring and does not intend to change
behavior.
LiveInterval: Fix computeFromMainRange() producing adjacent segments with same valno
If two livesegments from different subranges happened to have the same
definition they could possibly end up as two adjacent segments in the
main liverange with the same value number which is not allowed. Detect
such cases and fix them in the 2nd pass of computeFromMainRange() if
necessary.
No testcase as there is only an out-of-tree target where I can sensibly
come up with one.
Tom Stellard [Wed, 8 Apr 2015 01:09:26 +0000 (01:09 +0000)]
R600/SI: Initial support for assembler and inline assembly
This is currently considered experimental, but most of the more
commonly used instructions should work.
So far only SI has been extensively tested, CI and VI probably work too,
but may be buggy. The current set of tests cases do not give complete
coverage, but I think it is sufficient for an experimental assembler.
See the documentation in R600Usage for more information.
Tim Northover [Tue, 7 Apr 2015 22:49:47 +0000 (22:49 +0000)]
AArch64: disallow "fmov sD, #-0.0" during assembly.
We weren't checking the sign of the floating point immediate before translating
it to "fmov sD, wzr". Similarly for D-regs.
Technically "movi vD.2s, #0x80, lsl #24" would work most of the time, but it's
not a blessed alias (and I don't think it should be since people expect writing
sD to zero out the high lanes, and there's no dD equivalent). So an error it is.
Adam Nemet [Tue, 7 Apr 2015 21:46:16 +0000 (21:46 +0000)]
[LoopAccesses] Allow analysis to complete in the presence of uniform stores
Both run-time pointer checking and the dependence analysis are capable
of dealing with uniform addresses. I.e. it's really just an orthogonal
property of the loop that the analysis computes.
Run-time pointer checking will only try to reason about SCEVAddRec
pointers or else gives up. If the uniform pointer turns out the be a
SCEVAddRec in an outer loop, the run-time checks generated will be
correct (start and end bounds would be equal).
In case of the dependence analysis, we work again with SCEVs. When
compared against a loop-dependent address of the same underlying object,
the difference of the two SCEVs won't be constant. This will result in
returning an Unknown dependence for the pair.
When compared against another uniform access, the difference would be
constant and we should return the right type of dependence
(forward/backward/etc).
The changes also adds support to query this property of the loop and
modify the vectorizer to use this.
Pete Cooper [Tue, 7 Apr 2015 20:43:23 +0000 (20:43 +0000)]
Add boolean to PrintStackTraceOnErrorSignal to disable crash reporting.
The current crash reporting on Mac OS is only disabled via an environment variable.
This adds a boolean (default false) which can also disable crash reporting.
The only client right now is the unittests which don't ever want crash reporting, but do want to detect killed programs.
Reduces the time to run the APFloat unittests on my machine from
[----------] 47 tests from APFloatTest (51250 ms total)
to
[----------] 47 tests from APFloatTest (765 ms total)
Ahmed Bougacha [Tue, 7 Apr 2015 20:31:16 +0000 (20:31 +0000)]
[ARM] Mark a bunch of .td Operands with type _MEMORY.
This shouldn't affect anything in-tree, as the OperandType users are
mostly smart disassemblers and such; more information is helpful there.
However, on the flip side, that + the fact that this is just hinting at
the meaning of operands makes this not really test-worthy or testable.
Revert "Try a third time to fix MSVC build after r234290"
This reverts commit r234295 (and r234294 and r234292 before it). I
removed the implicit conversion to `MDTuple*` r234326, so there's no
longer an ambiguity in `operator[]()`.
I think MSVC should accept the original code now...
AArch64: Don't lower ISD::SELECT to ISD::SELECT_CC
Instead of lowering SELECT to SELECT_CC which is further lowered later
immediately call the SELECT_CC lowering code. This is preferable
because:
- Avoids an unnecessary roundtrip through the legalization queues with
an intermediate node.
- More importantly: Lowered operations get visited last leading to SELECT_CC
getting visited with legalized operands and unlegalized ones for preexisting
SELECT_CC nodes. This does not hurt the current code (hence no testcase) but
is required for another patch I am working on.
Revert "Workaround bot failure with explicit conversion to MDTuple*"
This reverts commit r234329, which insufficiently appeased older
`clang`s (apparently that wasn't the only call site). r234331 was a
more complete fix.
Remove `MDTupleTypedArrayWrapper::operator MDTuple*()`, since it causes
ambiguity (at least in some [1] compilers [2]) when using indexes to
`MDTupleTypedArrayWrapper::operator[](unsigned)` that are convertible to
(but not the same as) `unsigned`.
Daniel Jasper [Tue, 7 Apr 2015 16:42:35 +0000 (16:42 +0000)]
Cleanup register pressure calculation in MachineLICM.
There were four almost identical implementations of calculating/updating
the register pressure for a certain MachineInstr. Cleanup to have a
single implementation (well, controlled with two bool flags until this
is cleaned up more).
No functional changes intended.
Tested by verify that there are no binary changes in the entire llvm
test-suite. A new test was added separately in r234309 as it revealed a
pre-existing error in the register pressure calculation.
[TableGen] Prevent invalid code generation when emitting AssemblerPredicate conditions.
Summary:
The loop which emits AssemblerPredicate conditions also links them together by emitting a '&&'.
If the 1st predicate is not an AssemblerPredicate, while the 2nd one is, nothing gets emitted for the 1st one, but we still emit the '&&' because of the 2nd predicate.
This generated code looks like "( && Cond2)" and is invalid.
Lang Hames [Tue, 7 Apr 2015 06:27:56 +0000 (06:27 +0000)]
[RuntimeDyld] Always allocate at least 1 byte for object sections in the JIT to
ensure that section addresses are distinct.
mapSectionAddress will fail if two sections are allocated the same address,
which can happen if any section has zero size (since malloc(0) is implementation
defined). Unfortunately I've been unable to repro this with a simple test case.
DebugInfo: Remove DITypedArray<>, replace with typedefs
Replace all uses of `DITypedArray<>` with `MDTupleTypedArrayWrapper<>`
and `MDTypeRefArray`. The APIs are completely different, but the
provided functionality is the same: treat an `MDTuple` as if it's an
array of a particular element type.
To simplify this patch a bit, I've temporarily typedef'ed
`DebugNodeArray` to `DIArray` and `MDTypeRefArray` to `DITypeArray`.
I've also temporarily conditionalized the accessors to check for null --
eventually these should be changed to asserts and the callers should
check for null themselves.
IR: Stop using DIDescriptor::is*() and auto-casting
`DIDescriptor`'s subclasses allow construction from incompatible
pointers, and `DIDescriptor` defines a series of `isa<>`-like functions
(e.g., `isCompileUnit()` instead of `isa<MDCompileUnit>()`) that clients
tend to use like this:
if (DICompileUnit(N).isCompileUnit())
foo(DICompileUnit(N));
These construction patterns work together to make `DIDescriptor` behave
differently from normal pointers.
Instead, use built-in `isa<>`, `dyn_cast<>`, etc., and only build
`DIDescriptor`s from pointers that are valid for their type.
I've split this into a few commits for different parts of LLVM and clang
(to decrease the patch size and increase the chance of review).
Generally the changes I made were NFC, but in a few places I made things
stricter if it made sense from the surrounded code.
Eventually a follow-up commit will remove the API for the "old" way.
Kevin Enderby [Mon, 6 Apr 2015 22:33:43 +0000 (22:33 +0000)]
Fix failure on builder clang-cmake-mips where it was printing a 32-bit address
incorrectly because it came from an expression using S.getAddress() which always
returns a 64-bit value.
DebugInfo: Allow isa<> on DIDescriptor and subclasses
Allow LLVM-style casting on `DIDescriptor` and its subclasses so they
can behave more like raw pointers. I haven't bothered with tests since
I have a follow-up commit coming shortly that uses them extensively in
tree, and I'm hoping to kill `DIDescriptor` entirely before too long (so
they won't have time to bitrot).
Usage examples:
DIDescriptor D = foo();
if (DICompileUnit CU = dyn_cast<MDCompileUnit>(D))
return bar(CU);
else if (auto *SP = dyn_cast<MDSubprogram>(D))
return baz(SP);
return other(D);
Fast isel used to zero extends immediates to 64 bits. This normally goes
unnoticed because the value is truncated to 32 bits for output.
Two cases were it is noticed:
* We fail to use smaller encodings.
* If the original constant was smaller than i32.
In the tests using i1 constants, codegen would change to use -1, which is fine
(and matches what regular isel does) since only the lowest bit is then used.
Instead, this patch then changes the ir to use i8 constants, which looks more
like what clang produces.
DebugInfo: Reimplement DIRef<>::resolve() using TypedDebugNodeRef<>
Gut `DIRef<>::resolve()`, reimplementing it using
`TypedDebugNodeRef<>::resolve()`. Use three separate functions rather
than some sort of type traits, since the latter (i.e., mapping `DIScope`
=> `MDScope`) seems heavy-handed. I don't expect `DIRef<>` to last much
longer in tree anyway.
As a drive-by fix, make `TypedDebugNodeRef<>::resolve()` do the right
thing with `nullptr`.
The sanitizer test suite uses this idiom to disable a test. Now that we
actually check if a test ran after invoking it, we see that zero tests
ran, and complain.
Instead, ignore tests starting with DISABLED_ completely. Fixes the
sanitizer test suite failures on Windows.
DebugInfo: Drop confusing forwarding API from DILexicalBlockFile
Remove `DILexicalBlockFile::getScope()` (whose last use was removed from
clang in r234245), which illegally returned a `DILexicalBlock` despite
its scope sometimes being an `MDSubprogram`. Also remove the
`getLineNumber()` and `getColumnNumber()` methods that just forwarded to
`DILexicalBlock`'s versions, since there don't seem to be any callers.
Note that the block of code removed from `DebugInfo.cpp` was actually
dead code, since `isLexicalBlock()` (the previous branch) always returns
true when `isLexicalBlockFile()` returns true.
An earlier (broken and untested) version of this was squashed into
r234222 and reverted in r234225.
[lit] Fix running gtest type-parameterized tests on Windows
The '/' character in the test name of a type-parameterized test is not a
path separator, and should not be '\' on Windows. We were passing a test
name to --gtest_filter which found no tests, so the exit code was zero,
indicating a passed test.
This bug has been here since r84387 in 2009, when Jeff Yasskin added the
original lit support for type-paratermized tests. Somewhere along the
line some of the ValueMapTests started failing, but we can fix those
separately.
This reverts commit r234225, reapplying r234222 in spirit.
This time, just include what the commit message actually describes:
loosen the `DILexicalBlock` constructor to require a
`MDLexicalBlockBase`, since that's what `DILexicalBlock` is wrapping.
This class wraps `MDLexicalBlockBase`, so allow construction from it!
Currently doesn't cause any problems because of the explicit `MNode*`
constructor, but I'll be removing that soon enough.
David Blaikie [Mon, 6 Apr 2015 20:59:48 +0000 (20:59 +0000)]
[opaque pointer type] Avoid using PointerType::getElementType when parsing IR
A few calls are left in for error checking - but I'm commenting those
out & trying to build some IR tests (aiming for Argument Promotion to
start with). When I get any of these tests passing I may add flag to
disable the checking so I can add tests that pass with the assertion in
place.
projects / llvm / log