Philip Reames [Sat, 31 Dec 2016 02:33:22 +0000 (02:33 +0000)]
[SmallPtrSet] Introduce a find primitive and rewrite count/erase in terms of it
This was originally motivated by a compile time problem I've since figured out how to solve differently, but the cleanup seemed useful. We had the same logic - which essentially implemented find - in several places. By commoning them out, I can implement find and allow erase to be inlined at the call sites if profitable.
Philip Reames [Fri, 30 Dec 2016 18:00:55 +0000 (18:00 +0000)]
[CVP] Adjust iteration order to reduce the amount of work required
CVP doesn't care about the order of blocks visited, but by using a pre-order traversal over the graph we can a) not visit unreachable blocks and b) optimize as we go so that analysis of later blocks produce slightly more precise results.
I noticed this via inspection and don't have a concrete example which points to the issue.
Mehdi Amini [Fri, 30 Dec 2016 01:15:50 +0000 (01:15 +0000)]
Fix test change in r290736: restore index generation
I remove one extra line, but because annoyingly llvm-lit does not
clean the output directory before running the test, it didn't fail
locally (the file was present from a previous run).
Dehao Chen [Fri, 30 Dec 2016 00:50:28 +0000 (00:50 +0000)]
Use continuous boosting factor for complete unroll.
Summary:
The current loop complete unroll algorithm checks if unrolling complete will reduce the runtime by a certain percentage. If yes, it will apply a fixed boosting factor to the threshold (by discounting cost). The problem for this approach is that the threshold abruptly. This patch makes the boosting factor a function of runtime reduction percentage, capped by a fixed threshold. In this way, the threshold changes continuously.
The patch also simplified the code by reducing one parameter in UP.
The patch only affects code-gen of two speccpu2006 benchmark:
445.gobmk binary size decreases 0.08%, no performance change.
464.h264ref binary size increases 0.24%, no performance change.
[LICM] Remove unneeded tracking of whether changes were made. NFC.
"Changed" doesn't actually change within the loop, so there's
no reason to keep track of it - we always return false during
analysis and true after the transformation is made.
David Majnemer [Fri, 30 Dec 2016 00:28:58 +0000 (00:28 +0000)]
[InstCombine] More thoroughly canonicalize the position of zexts
We correctly canonicalized (add (sext x), (sext y)) to (sext (add x, y))
where possible. However, we didn't perform the same canonicalization
for zexts or for muls.
[LICM] Compute exit blocks for promotion eagerly. NFC.
This moves the exit block and insertion point computation to be eager,
instead of after seeing the first scalar we can promote.
The cost is relatively small (the computation happens anyway, see discussion
on D28147), and the code is easier to follow, and can bail out earlier
if there's a catchswitch present.
[LICM] Don't try to promote in loops where we have no chance to promote. NFC.
We would check whether we have a prehader *or* dedicated exit blocks,
and go into the promotion loop. Then, for each alias set we'd check
if we have a preheader *and* dedicated exit blocks, and bail if not.
[LICM] Only recompute LCSSA when we actually promoted something.
We want to recompute LCSSA only when we actually promoted a value.
This means we only need to look at changes made by promotion when
deciding whether to recompute it or not, not at regular sinking/hoisting.
(This was what the code was documented as doing, just not what it did)
Daniel Berlin [Thu, 29 Dec 2016 22:15:12 +0000 (22:15 +0000)]
NewGVN: Fix PR 31491 by ensuring that we touch the right instructions. Change to one based numbering so we can assert we don't cause the same bug again.
Justin Lebar [Thu, 29 Dec 2016 19:59:38 +0000 (19:59 +0000)]
[ADT] Rewrite IntrusiveRefCntPtr's comments. NFC
Edit for voice, and also add examples. In particular, add an
explanation for why you might want to specialize IntrusiveRefCntPtrInfo,
which is not obvious.
Justin Lebar [Thu, 29 Dec 2016 19:59:26 +0000 (19:59 +0000)]
[ADT] Delete RefCountedBaseVPTR.
Summary:
This class is unnecessary.
Its comment indicated that it was a compile error to allocate an
instance of a class that inherits from RefCountedBaseVPTR on the stack.
This may have been true at one point, but it's not today.
Moreover you really do not want to allocate *any* refcounted object on
the stack, vptrs or not, so if we did have a way to prevent these
objects from being stack-allocated, we'd want to apply it to regular
RefCountedBase too, obviating the need for a separate RefCountedBaseVPTR
class.
It seems that the main way RefCountedBaseVPTR provides safety is by
making its subclass's destructor virtual. This may have been helpful at
one point, but these days clang will emit an error if you define a class
with virtual functions that inherits from RefCountedBase but doesn't
have a virtual destructor.
Bryant Wong [Thu, 29 Dec 2016 19:32:34 +0000 (19:32 +0000)]
Correctly handle multi-lined RUN lines.
`utils/update_{llc_test,test}_checks` ought to be able to handle RUN commands
that span multiple lines, as shown in the example at
http://llvm.org/docs/CommandGuide/FileCheck.html#the-filecheck-check-prefix-option
Reid Kleckner [Thu, 29 Dec 2016 17:07:10 +0000 (17:07 +0000)]
Revert "[COFF] Use 32-bit jump table entries in .rdata for Win64"
This reverts commit r290694. It broke sanitizer tests on Win64. I'll
probably bring this back, but the jump tables will just live in .text
like they do for MSVC.
Sanjoy Das [Thu, 29 Dec 2016 15:47:05 +0000 (15:47 +0000)]
[TBAAVerifier] Be stricter around verifying scalar nodes
This fixes the issue exposed in PR31393, where we weren't trying
sufficiently hard to diagnose bad TBAA metadata.
This does reduce the variety in the error messages we print out, but I
think the tradeoff of verifying more, simply and quickly overrules the
need for more helpful error messags here.
Igor Laevsky [Thu, 29 Dec 2016 14:31:07 +0000 (14:31 +0000)]
Introduce element-wise atomic memcpy intrinsic
This change adds a new intrinsic which is intended to provide memcpy functionality
with additional atomicity guarantees. Please refer to the review thread
or language reference for further details.
Craig Topper [Thu, 29 Dec 2016 04:24:32 +0000 (04:24 +0000)]
[InstCombine] Use a 32-bits instead of 64-bits for storing the number of elements in VectorType for a ShuffleVector. While there getVectorNumElements to avoid an explicit cast. NFC
Craig Topper [Thu, 29 Dec 2016 03:30:17 +0000 (03:30 +0000)]
[InstCombine][X86] If the lowest element of a scalar intrinsic isn't used make sure we add it to the worklist so we can DCE it sooner.
We bypassed the intrinsic and returned the passthru operand, but we should also add the intrinsic to the worklist since its now dead. This can allow DCE to find it sooner and remove it. Similar was done for InsertElement when the inserted element isn't demanded.
Craig Topper [Thu, 29 Dec 2016 02:29:04 +0000 (02:29 +0000)]
[InstCombine] Fix some of the AVX-512 scalar arithmetic test cases to do a better job of testing what they intended to test.
The accidentally had trivially dead code. Also needed to adjust the rounding mode to not CUR_DIRECTION so the intrinsics don't get converted to native operations before going through SimplifyDemandedVectorElts.
Reid Kleckner [Thu, 29 Dec 2016 01:14:41 +0000 (01:14 +0000)]
Fix mingw build by moving the static const data member before the bitfields
Apparently GCC targeting Windows breaks bitfields on static data members:
struct Foo {
unsigned X : 16;
static const int M = 42;
unsigned Y : 16;
};
static_assert(sizeof(Foo) == 4, "asdf"); // fails
Daniel Berlin [Thu, 29 Dec 2016 01:12:36 +0000 (01:12 +0000)]
NewGVN: Sort Dominator Tree in RPO order, and use that for generating order.
Summary:
The optimal iteration order for this problem is RPO order. We want to
process as many preds of a backedge as we can before we process the
backedge.
At the same time, as we add predicate handling, we want to be able to
touch instructions that are dominated by a given block by
ranges (because a change in value numbering a predicate possibly
affects all users we dominate that are using that predicate).
If we don't do it this way, we can't do value inference over
backedges (the paper covers this in depth).
The newgvn branch currently overshoots the last part, and guarantees
that it will touch *at least* the right set of instructions, but it
does touch more. This is because the bitvector instruction ranges are
currently generated in RPO order (so we take the max and the min of
the ranges of dominated blocks, which means there are some in the
middle we didn't have to touch that we did).
We can do better by sorting the dominator tree, and then just using
dominator tree order.
As a preliminary, the dominator tree has some RPO guarantees, but not
enough. It guarantees that for a given node, your idom must come
before you in the RPO ordering. It guarantees no relative RPO ordering
for siblings. We add siblings in whatever order they appear in the module.
So that is what we fix.
We sort the children array of the domtree into RPO order, and then use
the dominator tree for ordering, instead of RPO, since the dominator
tree is now a valid RPO ordering.
Note: This would help any other pass that iterates a forward problem
in dominator tree order. Most of them are single pass. It will still
maximize whatever result they compute. We could also build the
dominator tree in this order, but our incremental updates would still
put it out of sort order, and recomputing the sort order is almost as
hard as general incremental updates of the domtree.
Also note that the sorting does not affect any tests, etc. Nothing
depends on domtree order, including the verifier, the equals
functions for domtree nodes, etc.
How much could this matter, you ask?
Here are the current numbers.
This is generated by running NewGVN over all files in LLVM.
Note that once we propagate equalities, the differences go up by an
order of magnitude or two (IE instead of 29, the max ends up in the
thousands, since the worst case we add a factor of N, where N is the
number of branch predicates). So while it doesn't look that stark for
the default ordering, it gets *much much* worse. There are also
programs in the wild where the difference is already pretty stark
(2 iterations vs hundreds).
RPO ordering:
759040 Number of iterations is 1
112908 Number of iterations is 2
Default dominator tree ordering:
755081 Number of iterations is 1
116234 Number of iterations is 2
603 Number of iterations is 3
27 Number of iterations is 4
2 Number of iterations is 5
1 Number of iterations is 7
Dominator tree sorted:
759040 Number of iterations is 1
112908 Number of iterations is 2
<yay!>
Really bad ordering (sort domtree siblings in postorder. not quite the
worst possible, but yeah):
754008 Number of iterations is 1
21 Number of iterations is 10
8 Number of iterations is 11
6 Number of iterations is 12
5 Number of iterations is 13
2 Number of iterations is 14
2 Number of iterations is 15
3 Number of iterations is 16
1 Number of iterations is 17
2 Number of iterations is 18
96642 Number of iterations is 2
1 Number of iterations is 20
2 Number of iterations is 21
1 Number of iterations is 22
1 Number of iterations is 29
17266 Number of iterations is 3
2598 Number of iterations is 4
798 Number of iterations is 5
273 Number of iterations is 6
186 Number of iterations is 7
80 Number of iterations is 8
42 Number of iterations is 9
Reid Kleckner [Thu, 29 Dec 2016 00:55:51 +0000 (00:55 +0000)]
Add a static_assert about the sizeof(GlobalValue)
I added one for Value back in r262045, and I'm starting to think we
should have these for any class with bitfields whose memory efficiency
really matters.
Justin Lebar [Thu, 29 Dec 2016 00:30:46 +0000 (00:30 +0000)]
[GlobalValue] Move HasLLVMReservedName into existing bitfield. NFC
Summary:
Follow-up to r290691, where I introduced HasLLVMReservedName. rnk
pointed out that that patch added an extra word to GlobalValue on MSVC,
because it doesn't pack bitfields with different types.
This patch moves HasLLVMReservedName into the existing bitfield, where
we appear to have plenty of bits to spare.
Reid Kleckner [Thu, 29 Dec 2016 00:12:39 +0000 (00:12 +0000)]
[COFF] Use 32-bit jump table entries in .rdata for Win64
Summary:
We were already using 32-bit jump table entries, but this was a
consequence of the default PIC model on Win64, and not an intentional
design decision. This patch ensures that we always use 32-bit label
difference jump table entries on Win64 regardless of the PIC model. This
is a good idea because it saves executable size and object file size.
Moving the jump tables to .rdata cleans up the disassembled object code
and reduces the available ROP targets, but it requires adding one more
RIP-relative lea to the code. COFF doesn't have relocations to express
the difference between two arbitrary symbols, so we can't use the jump
table label in the label difference like we do elsewhere.
Mehdi Amini [Wed, 28 Dec 2016 23:45:54 +0000 (23:45 +0000)]
Change Metadata Index emission in the bitcode to use 2x32 bits for the placeholder
The Bitstream reader and writer are limited to handle a "size_t" at
most, which means that we can't backpatch and read back a 64bits
value on 32 bits platform.
Justin Lebar [Wed, 28 Dec 2016 22:59:45 +0000 (22:59 +0000)]
Speed up Function::isIntrinsic() by adding a bit to GlobalValue. NFC
Summary:
Previously isIntrinsic() called getName(). This involves a hashtable
lookup, so is nontrivially expensive. And isIntrinsic() is called
frequently, particularly by dyn_cast<IntrinsicInstr>.
This patch steals a bit of IntID and uses that to store whether or not
getName() starts with "llvm."
Mehdi Amini [Wed, 28 Dec 2016 22:30:28 +0000 (22:30 +0000)]
Add an index for Module Metadata record in the bitcode
This index record the position for each metadata record in
the bitcode, so that the reader will be able to lazy-load
on demand each individual record.
We also make sure that every abbrev is emitted upfront so
that the block can be skipped while reading.
I don't plan to commit this before having the reader
counterpart, but I figured this can be reviewed mostly
independently.
Recommit r290684 (was reverted in r290686 because a test
was broken) after adding a threshold to avoid emitting
the index when unnecessary (little amount of metadata).
This optimization "hides" a limitation of the ability
to backpatch in the bitstream: we can only backpatch
safely when the position has been flushed. So if we emit
an index for one metadata, it is possible that (part of)
the offset placeholder hasn't been flushed and the backpatch
will fail.
Piotr Padlewski [Wed, 28 Dec 2016 20:36:08 +0000 (20:36 +0000)]
[NewGVN] replace emplace_back with push_back
emplace_back is not faster if it is equivalent to push_back. In this cases emplaced value had the
same type that the one stored in container. It is ugly and it might be even slower (see
Scott Meyers presentation about emplacement).
Mehdi Amini [Wed, 28 Dec 2016 19:44:19 +0000 (19:44 +0000)]
Add an index for Module Metadata record in the bitcode
Summary:
This index record the position for each metadata record in
the bitcode, so that the reader will be able to lazy-load
on demand each individual record.
We also make sure that every abbrev is emitted upfront so
that the block can be skipped while reading.
I don't plan to commit this before having the reader
counterpart, but I figured this can be reviewed mostly
independently.
Reid Kleckner [Wed, 28 Dec 2016 19:05:12 +0000 (19:05 +0000)]
[WinEH] Don't assume endFunction is called while in .text
Jump table emission can switch to .rdata before
WinException::endFunction gets called. Just remember the appropriate
text section we started in and reset back to it when we end the
function. We were already switching sections back from .xdata anyway.
Fixes the first problem in PR31488, so that now COFF switch tables can
live in .rdata if we want them to.
Chandler Carruth [Wed, 28 Dec 2016 11:07:33 +0000 (11:07 +0000)]
[PM] Introduce a devirtualization iteration layer for the new PM.
This is an orthogonal and separated layer instead of being embedded
inside the pass manager. While it adds a small amount of complexity, it
is fairly minimal and the composability and control seems worth the
cost.
The logic for this ends up being nicely isolated and targeted. It should
be easy to experiment with different iteration strategies wrapped around
the CGSCC bottom-up walk using this kind of facility.
The mechanism used to track devirtualization is the simplest one I came
up with. I think it handles most of the cases the existing iteration
machinery handles, but I haven't done a *very* in depth analysis. It
does however match the basic intended semantics, and we can tweak or
tune its exact behavior incrementally as necessary. One thing that we
may want to revisit is freshly building the value handle set on each
iteration. While I don't think this will be a significant cost (it is
strictly fewer value handles but more churn of value handes than the old
call graph), it is conceivable that we'll want a somewhat more clever
tracking mechanism. My hope is to layer that on as a follow up patch
with data supporting any implementation complexity it adds.
This code also provides for a basic count heuristic: if the number of
indirect calls decreases and the number of direct calls increases for
a given function in the SCC, we assume devirtualization is responsible.
This matches the heuristics currently used in the legacy pass manager.
Chandler Carruth [Wed, 28 Dec 2016 10:34:50 +0000 (10:34 +0000)]
[PM] Teach the CGSCC's CG update utility to more carefully invalidate
analyses when we're about to break apart an SCC.
We can't wait until after breaking apart the SCC to invalidate things:
1) Which SCC do we then invalidate? All of them?
2) Even if we invalidate all of them, a newly created SCC may not have
a proxy that will convey the invalidation to functions!
Previously we only invalidated one of the SCCs and too late. This led to
stale analyses remaining in the cache. And because the caching strategy
actually works, they would get used and chaos would ensue.
Doing invalidation early is somewhat pessimizing though if we *know*
that the SCC structure won't change. So it turns out that the design to
make the mutation API force the caller to know the *kind* of mutation in
advance was indeed 100% correct and we didn't do enough of it. So this
change also splits two cases of switching a call edge to a ref edge into
two separate APIs so that callers can clearly test for this and take the
easy path without invalidating when appropriate. This is particularly
important in this case as we expect most inlines to be between functions
in separate SCCs and so the common case is that we don't have to so
aggressively invalidate analyses.
The LCG API change in turn needed some basic cleanups and better testing
in its unittest. No interesting functionality changed there other than
more coverage of the returned sequence of SCCs.
While this seems like an obvious improvement over the current state, I'd
like to revisit the core concept of invalidating within the CG-update
layer at all. I'm wondering if we would be better served forcing the
callers to handle the invalidation beforehand in the cases that they
can handle it. An interesting example is when we want to teach the
inliner to *update and preserve* analyses. But we can cross that bridge
when we get there.
With this patch, the new pass manager an build all of the LLVM test
suite at -O3 and everything passes. =D I haven't bootstrapped yet and
I'm sure there are still plenty of bugs, but this gives a nice baseline
so I'm going to increasingly focus on fleshing out the missing
functionality, especially the bits that are just turned off right now in
order to let us establish this baseline.
Gadi Haber [Wed, 28 Dec 2016 10:12:48 +0000 (10:12 +0000)]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible.
There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers.
The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled.
Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky
Differential Revision: https://reviews.llvm.org/D27901
Chandler Carruth [Wed, 28 Dec 2016 03:13:12 +0000 (03:13 +0000)]
[PM] Teach the inliner's call graph update to handle inserting new edges
when they are call edges at the leaf but may (transitively) be reached
via ref edges.
It turns out there is a simple rule: insert everything as a ref edge
which is a safe conservative default. Then we let the existing update
logic handle promoting some of those to call edges.
Note that it would be fairly cheap to make these call edges right away
if that is desirable by testing whether there is some existing call path
from the source to the target. It just seemed like slightly more
complexity in this code path that isn't strictly necessary. If anyone
feels strongly about handling this differently I'm happy to change it.
Chandler Carruth [Wed, 28 Dec 2016 02:24:55 +0000 (02:24 +0000)]
[PM] Disable the loop vectorizer from the new PM's pipeline as it
currenty relies on the old PM's dependency system forming LCSSA.
The new PM will require a different design for this, and for now this is
causing most of the issues I'm currently seeing in testing. I'd like to
get to a testable baseline and then work on re-enabling things one at
a time.
[libFuzzer] add an experimental flag -experimental_len_control=1 that sets max_len to 1M and tries to increases the actual max sizes of mutations very gradually (second attempt)
Replace the use of grep with FileCheck. Tidy up some of the tests. A
few of the tests have been left as weak as previously, though some have
been made more stringent.
Chandler Carruth [Tue, 27 Dec 2016 18:04:11 +0000 (18:04 +0000)]
[PM] Remove a pointless optimization.
There is no need to do this within an analysis. That method shouldn't
even be reached if this predicate holds as the actual useful
optimization is in the analysis manager itself.
Chandler Carruth [Tue, 27 Dec 2016 17:59:22 +0000 (17:59 +0000)]
[PM] Add more dedicated testing to cover the invalidation logic added to
BasicAA in r290603.
I've kept the basic testing in the new PM test file as that also covers
the AAManager invalidation logic. If/when there is a good place for
broader AA testing it could move there.
This test is somewhat unsatisfying as I can't get it to fail even with
ASan outside of explicit checks of the invalidation. Apparently we don't
yet have any test coverage of the BasicAA code paths using either the
domtree or loopinfo -- I made both of them always be null and check-llvm
passed.
projects / llvm / log