Sam Parker [Tue, 11 Apr 2017 08:43:32 +0000 (08:43 +0000)]
[SelectionDAG] Check CALLSEQ_BEGIN nodes in DelayForLiveRegs
A fix for the bug reported in PR30911.
The issue arises when multiple CALLSEQ_BEGIN nodes are unscheduled as
the last node to be unscheduled will gain access to the CallResource
register. But when a node is being picked, only CALLSEQ_END nodes are
checked against the CallResource and have their chains evaluated.
This then means that other CALLSEQ_BEGIN nodes can be scheduled
before the existing call sequence has been finalised. This patch adds
a check against the FrameSetup nodes in DelayForLiveRegs to prevent
this from happening.
Module::getOrInsertFunction is using C-style vararg instead of
variadic templates.
From a user prospective, it forces the use of an annoying nullptr
to mark the end of the vararg, and there's not type checking on the
arguments. The variadic template is an obvious solution to both
issues.
Sanjoy Das [Tue, 11 Apr 2017 04:11:47 +0000 (04:11 +0000)]
[LoopUnswitch] Fix a test case
(h/t to Chandler for pointing this out)
The test in question was not at all testing what it was supposed to
test. We do not //care// about placing `!make.implicit` in inner
constant branch (since it will be folded away anyway). We care about
placing `!make.implicit` in the outer branch that switches between
either version of the loop.
Having said that, it is _correct_ to leave behind the `!make.implicit`
in the inner branch, but there is no need to do so.
Hal Finkel [Tue, 11 Apr 2017 02:03:17 +0000 (02:03 +0000)]
[PowerPC] multiply-with-overflow might use the CTR register
Check the legality of ISD::[US]MULO to see whether
Intrinsic::[us]mul_with_overflow will legalize into a function call (and, thus,
will use the CTR register). Fixes PR32485.
Hal Finkel [Tue, 11 Apr 2017 00:18:42 +0000 (00:18 +0000)]
[bugpoint] Also remove comdat's from externalized GVs
We were removing comdats from externalized functions (function declarations
can't be comdat), but were not doing the same for variable. Failure to do this
would cause bugpoint to fail ("Declaration may not be in a Comdat!").
Remove AttributeSetNode::get(AttributeList, unsigned) and sink constructor
The getter was equivalent to AttributeList::getAttributes(unsigned),
which seems like a better way to express getting the AttributeSet for a
given index. This static helper was only used in one place anyway.
The constructor doesn't benefit from inlining and doesn't need to be in
a header.
Reland "[IR] Make AttributeSetNode public, avoid temporary AttributeList copies"
This re-lands r299875.
I introduced a bug in Clang code responsible for replacing K&R, no
prototype declarations with a real function definition with a prototype.
The bug was here:
// Collect any return attributes from the call.
- if (oldAttrs.hasAttributes(llvm::AttributeList::ReturnIndex))
- newAttrs.push_back(llvm::AttributeList::get(newFn->getContext(),
- oldAttrs.getRetAttributes()));
+ newAttrs.push_back(oldAttrs.getRetAttributes());
Previously getRetAttributes() carried AttributeList::ReturnIndex in its
AttributeList. Now that we return the AttributeSetNode* directly, it no
longer carries that index, and we call this overload with a single node:
AttributeList::get(LLVMContext&, ArrayRef<AttributeSetNode*>)
That aborted with an assertion on x86_32 targets. I added an explicit
triple to the test and added CHECKs to help find issues like this in the
future sooner.
CodeGen: BlockPlacement: Don't always tail-duplicate with no other successor.
The math works out where it can actually be counter-productive. The probability
calculations correctly handle the case where the alternative is 0 probability,
rely on those calculations.
Includes a test case that demonstrates the problem.
CodeGen: BranchFolding: Merge identical blocks, even if they are short.
Merging identical blocks when it doesn't reduce fallthrough. It is common for
the blocks created from critical edge splitting to be identical. We would like
to merge these blocks whenever doing so would not reduce fallthrough.
Matt Arsenault [Mon, 10 Apr 2017 22:27:50 +0000 (22:27 +0000)]
Allow DataLayout to specify addrspace for allocas.
LLVM makes several assumptions about address space 0. However,
alloca is presently constrained to always return this address space.
There's no real way to avoid using alloca, so without this
there is no way to opt out of these assumptions.
The problematic assumptions include:
- That the pointer size used for the stack is the same size as
the code size pointer, which is also the maximum sized pointer.
- That 0 is an invalid, non-dereferencable pointer value.
These are problems for AMDGPU because alloca is used to
implement the private address space, which uses a 32-bit
index as the pointer value. Other pointers are 64-bit
and behave more like LLVM's notion of generic address
space. By changing the address space used for allocas,
we can change our generic pointer type to be LLVM's generic
pointer type which does have similar properties.
[GVNHoist] Call isGuaranteedToTransferExecutionToSuccessor on each instruction
w.r.t. https://bugs.llvm.org/show_bug.cgi?id=32153
The consensus seems to be isGuaranteedToTransferExecutionToSuccessor should be called for each function.
Revert "[IR] Make AttributeSetNode public, avoid temporary AttributeList copies"
This reverts r299875. A Linux bot came back with a test failure:
http://bb.pgr.jp/builders/test-clang-i686-linux-RA/builds/741/steps/test_clang/logs/Clang%20%3A%3A%20CodeGen__2006-05-19-SingleEltReturn.c
[IR] Make AttributeSetNode public, avoid temporary AttributeList copies
Summary:
AttributeList::get(Fn|Ret|Param)Attributes no longer creates a temporary
AttributeList just to hide the AttributeSetNode type.
I've also added a factory method to create AttributeLists from a
parallel array of AttributeSetNodes. I think this simplifies
construction of AttributeLists when rewriting function prototypes.
Previously we would test if a particular index had attributes, and
conditionally add a temporary attribute list to a vector. Now the
attribute set vector is parallel to the argument vector already that
these passes already construct.
My long term vision is to wrap AttributeSetNode* inside an AttributeSet
type that holds the enum attributes, but that will come in a follow up
change.
I haven't done any performance measurements for this change because
profiling hasn't shown that any of the affected code is hot.
[llvm-pdbdump] Display padding bytes on record layout
When dumping classes, show where padding occurs, and at the end of the
class print statistics about how many bytes total of padding exist in a
class.
Since PDB doesn't specifically contain information about padding, we have
to mimic this by sort of reversing a small portion of the record layout
algorithm (e.g. looking at offsets and sizes and trying to determine
whether something is part of the same field or a new field).
Daniel Berlin [Mon, 10 Apr 2017 18:46:00 +0000 (18:46 +0000)]
MemorySSA: Make lifetime starts defs for mustaliased pointers
Summary:
While we don't want them aliasing with other pointers, there seems to
be no point in not having them clobber must-aliased'd pointers.
If some day, we split the aliasing and ordering chains, we'd make this
not aliasing but an ordering barrier (IE it doesn't affect it's
memory, but we can't hoist it above it).
Matthew Simpson [Mon, 10 Apr 2017 18:34:37 +0000 (18:34 +0000)]
[ARM/AArch64] Ensure valid vector element types for interleaved accesses
This patch refactors and strengthens the type checks performed for interleaved
accesses. The primary functional change is to ensure that the interleaved
accesses have valid element types. The added test cases previously failed
because the element type is f128.
Matt Arsenault [Mon, 10 Apr 2017 17:58:06 +0000 (17:58 +0000)]
AMDGPU: Fix crash when disassembling VOP3 mac
The unused dummy src2_modifiers is missing, so it crashes
when trying to print it.
I tried to fully remove src2_modifiers, but there are some
irritations in the places where it is converted to mad since
it starts to require modifying use lists while iterating over
them.
[Support] Add support for finding unset bits in a BitVector.
BitVector had methods for searching for the first and next
set bits, but it did not have analagous methods for finding
the first and next unset bits. This is useful when your ones
and zeros are grouped together and you want to iterate over
ranges of ones and zeros.
Adrian McCarthy [Mon, 10 Apr 2017 16:43:09 +0000 (16:43 +0000)]
Improves pretty printing of variable types in llvm-pdbdump
* Adds support for pointers to arrays, which was missing
* Adds some tests
* Improves consistency of const and volatile qualifiers
* Eliminates non-composable special case code for arrays and function by using
a more general recursive approach
* Has a hack for getting the calling convention into the right spot for
pointer-to-functions
Given the rapid changes happenning in llvm-pdbdump, this may be difficult to
merge.
[SelectionDAG] TargetLowering::SimplifyDemandedBits how to properly calculate KnownZero bits for ISD::SETCC and ISD::AssertZExt
Summary:
For SETCC we aren't calculating the KnownZero bits at all. I've copied the code from computeKnownZero over for this.
For AssertZExt we were only setting KnownZero for bits that were demanded. But the upper bits are zero whether they were demanded or not.
I'm interested in fixing this because my belief is the first part of the ISD::AND handling code in SimplifyDemandedBits largely exists because of these two bugs. In that code we go to computeKnownBits for the LHS and optimize a RHS constant. Because computeKnownBits handles SETCC and AssertZExt correctly we get better information sometimes than when we call SimplifyDemandedBits on the LHS later. With these two issues fixed in SimplifyDemandedBits I was able to remove that computeKnownBits call and still pass all X86 tests. I'll submit that change in a separate patch.
[InstCombine] Make sure we preserve fast math flags when folding fp instructions into phi nodes
Summary: I noticed in the select folding code that we copied fast math flags, but did not do the same for the similar handling in phi nodes. This patch fixes that to do the same thing as select
[InstCombine] Make the (A|B)^B -> A & ~B transform code consistent with the very similar (A&B)^B -> ~A & B code. This should be NFC except for the addition of hasOneUse check.
I think this code is still overly complicated and should use matchers, but first I wanted to make it consistent.
General usability improvements to generic PDB library.
1. Added some asserts to make sure concrete symbol types don't
get constructed with RawSymbols that have an incompatible
SymTag enum value.
2. Added new forwarding macros that auto-define an Id/Sym method
pair whenever there is a method that returns a SymIndexId.
Previously we would just provide one method that returned only
the SymIndexId and it was up to the caller to use the Session
object to get a pointer to the symbol. Now we automatically
get both the method that returns the Id, as well as a method
that returns the pointer directly with just one macro.
3. Added some methods for dumping straight to stdout that can
be used from inside the debugger for diagnostics during a
debug session.
4. Added a clone() method and a cast<T>() method to PDBSymbol
that can shorten some usage patterns.
[SCCP] Resolve indirect branch target when possible.
Summary:
Resolve indirect branch target when possible.
This potentially eliminates more basicblocks and result in better evaluation for phi and other things.
[InstCombine] remove dead cases from icmp pair switches; NFCI
"PredicatesFoldable" returns false for signed/unsigned mismatched pairs,
so these cases should never exist. We'll default to 'unreachable' on those
predicate combos instead.
Most of what's left in these switches belongs in InstSimplify (and may
already be there), so there's probably more that can be done to reduce
this code.
Hal Finkel [Sun, 9 Apr 2017 12:57:50 +0000 (12:57 +0000)]
[MemorySSA] Fix use of pointsToConstantMemory in isUseTriviallyOptimizableToLiveOnEntry
In isUseTriviallyOptimizableToLiveOnEntry, pointsToConstantMemory needs to be
called on the load's pointer operand, not on the result of the load (which
might not even be a pointer).
Coby Tayree [Sat, 8 Apr 2017 20:29:03 +0000 (20:29 +0000)]
[AsmParser]Emit an error if a macro has two (or more) parameters sharing the same name
Introducing a new error to macro parameters' parsing:
currently, llvm-mc won't complain if a macro have two (or more) named params with the same name.
this behavior is false, as there's no merit in having some params sharing a name.
now, instead of tolerate such a phenomena - emit an appropriate error.
This concludes the refinements to Falkor Machine Model.
It includes SchedPredicates for immediate zero and LSL Fast.
Forwarding logic is also modeled for vector multiply and
accumulate only.
Gor Nishanov [Sat, 8 Apr 2017 00:49:46 +0000 (00:49 +0000)]
[coroutines] Make CoroSplit pass deterministic
coro-split-after-phi.ll test was flaky due to non-determinism in
the coroutine frame construction that was sorting the spill
vector using a pointer to a def as a part of the key.
The sorting was intended to make sure that spills for the same def
are kept together, however, we populate the vector by processing
defs in order, so the spill entires will end up together anyways.
This change removes spill sorting and restores the determinism
in the test.
[GlobalISel]: Fix bug where we can report GISelFailure on erased instructions
The original instruction might get legalized and erased and expanded
into intermediate instructions and the intermediate instructions might
fail legalization. This end up in reporting GISelFailure on the erased
instruction.
Instead report GISelFailure on the intermediate instruction which failed
legalization.
Petr Hosek [Fri, 7 Apr 2017 20:41:58 +0000 (20:41 +0000)]
[AArch64] Allow global register asm("x18") or asm("w18") under -ffixed-x18
When using -ffixed-x18, the x18 (or w18) register can safely be used
with the "global register variable" GCC extension, but the backend
fails to recognize it.
[AMDGPU] Unroll more to eliminate phis and conditions
Increase threshold to unroll a loop which contains an "if" statement
whose condition defined by a PHI belonging to the loop. This may help
to eliminate if region and potentially even PHI itself, saving on
both divergence and registers used for the PHI.
Add a small bonus for each of such "if" statements.
Both pickling errors encountered on clang bots and Darwin compiler-rt
should now be fixed.
This has no impact on testing time on Linux, and on Windows goes from
88s to 63s for 'check'. The tests pass on Mac, but I haven't compared
execution time.
Gor Nishanov [Fri, 7 Apr 2017 14:16:49 +0000 (14:16 +0000)]
[coroutines] Insert spills of PHI instructions correctly
Summary:
Fix a bug where we were inserting a spill in between the PHIs in the beginning of the block.
Consider this fragment:
```
begin:
%phi1 = phi i32 [ 0, %entry ], [ 2, %alt ]
%phi2 = phi i32 [ 1, %entry ], [ 3, %alt ]
%sp1 = call i8 @llvm.coro.suspend(token none, i1 false)
switch i8 %sp1, label %suspend [i8 0, label %resume
i8 1, label %cleanup]
resume:
call i32 @print(i32 %phi1)
```
Unless we are spilling the argument or result of the invoke, we were always inserting the spill immediately following the instruction.
The fix adds a check that if the spilled instruction is a PHI Node, select an appropriate insert point with `getFirstInsertionPt()` that
skips all the PHI Nodes and EH pads.
Matthew Simpson [Fri, 7 Apr 2017 14:15:34 +0000 (14:15 +0000)]
Reapply r298620: [LV] Vectorize GEPs
This patch reapplies r298620. The original patch was reverted because of two
issues. First, the patch exposed a bug in InstCombine that caused the Chromium
builds to fail (PR32414). This issue was fixed in r299017. Second, the patch
introduced a bug in the vectorizer's scalars analysis that caused test suite
builds to fail on SystemZ. The scalars analysis was too aggressive and marked a
memory instruction scalar, even though it was going to be vectorized. This
issue has been fixed in the current patch and several new test cases for the
scalars analysis have been added.
Petar Jovanovic [Fri, 7 Apr 2017 13:31:36 +0000 (13:31 +0000)]
[mips][msa] Fix generation of bm(n)zi and bins[lr]i instructions
We have two cases here, the first one being the following instruction
selection from the builtin function:
bm(n)zi builtin -> vselect node -> bins[lr]i machine instruction
In case of bm(n)zi having an immediate which has either its high or low bits
set, a bins[lr] instruction can be selected through the selectVSplatMask[LR]
function. The function counts the number of bits set, and that value is
being passed to the bins[lr]i instruction as its immediate, which in turn
copies immediate modulo the size of the element in bits plus 1 as per specs,
where we get the off-by-one-error.
The other case is:
bins[lr]i -> vselect node -> bsel.v
In this case, a bsel.v instruction gets selected with a mask having one bit
less set than required.
projects / llvm / log