Summary: Partial unrolling should have separate threshold with full unrolling.
Reviewers: efriedma, mzolotukhin
Reviewed By: efriedma, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28831
llvm-svn: 292293
unique exit block if available rather than rolling it ourselves.
This is a little disappointing because that helper doesn't do anything
clever to short-circuit the (surprisingly expensive) computation of all
exit blocks. What's worse is that the way we compute this is hopelessly,
hilariously inefficient. We're literally computing the same information
two different ways and multiple times each way:
- hasDedicatedExits computes the exit block set and then looks at the
predecessors of each
- getExitingBlocks computes the set of loop blocks which have exiting
successors
- getUniqueExitBlock(s) computes the set of non-loop blocks reached from
loop blocks (sound familiar?)
Anyways, at some point we should clean all of this up in the LoopInfo
API, but for now just simplifying the user I'm about to touch.
llvm-svn: 292282
I hope that for any code, it is changed only with good reason and only
when the author knows what they are doing...
There is of course good reason to comment here about the subtlety of the
process, and I've left that comment in tact.
llvm-svn: 292275
instead of members.
No state was being provided by the object so this seems strictly
simpler.
I've also tried to improve the name and comments for the functions to
more thoroughly document what they are doing.
llvm-svn: 292274
that we know has exactly one element when all we are going to do is get
that one element out of it.
Instead, pass around that one element.
There are more simplifications to come in this code...
llvm-svn: 292273
If a memory instruction will be vectorized, but it's pointer operand is
non-consecutive-like, the instruction is a gather or scatter operation. Its
pointer operand will be non-uniform. This should fix PR31671.
Reference: https://llvm.org/bugs/show_bug.cgi?id=31671
Differential Revision: https://reviews.llvm.org/D28819
llvm-svn: 292254
Add missing fabs(fpext) optimzation that worked with the call,
and also fixes it creating a second fpext when there were multiple
uses.
llvm-svn: 292172
It's not clear what 'First' and 'Second' mean, so use 'Inner' and 'Outer'
to match foldShiftedShift() and add comments with formulas, so it's easier
to see what's going on.
llvm-svn: 292153
Simplify a pshufb shuffle mask based on the elements of the mask that are actually demanded.
Differential Revision: https://reviews.llvm.org/D28745
llvm-svn: 292101
a function's CFG when that CFG is unchanged.
This allows transformation passes to simply claim they preserve the CFG
and analysis passes to check for the CFG being preserved to remove the
fanout of all analyses being listed in all passes.
I've gone through and removed or cleaned up as many of the comments
reminding us to do this as I could.
Differential Revision: https://reviews.llvm.org/D28627
llvm-svn: 292054
mark it as never invalidated in the new PM.
The old PM already required this to work, and after a discussion with
Hal this seems to really be the only sensible answer. The cache
gracefully degrades as the IR is mutated, and most things which do this
should already be incrementally updating the cache.
This gets rid of a bunch of logic preserving and testing the
invalidation of this analysis.
llvm-svn: 292039
cover domtree and alias analysis. These are the pretty clear analyses
that we would always want to survive this pass.
To make these survive, we also need to preserve the assumption cache.
Added a test that verifies the important bits of this preservation.
llvm-svn: 292037
Allows LLVM to optimize sequences like the following:
%add = add nuw i32 %x, 1
%cmp = icmp ugt i32 %add, %y
Into:
%cmp = icmp uge i32 %x, %y
Previously, only signed comparisons were being handled.
Decrements could also be handled, but 'sub nuw %x, 1' is currently canonicalized to
'add %x, -1' in InstCombineAddSub, losing the nuw flag. Removing that canonicalization
seems like it might have far-reaching ramifications so I kept this simple for now.
Patch by Matti Niemenmaa!
Differential Revision: https://reviews.llvm.org/D24700
llvm-svn: 291975
Summary:
This is a testcase where phi node cycling happens, and because we do
not order the leaders by domination or anything similar, the leader
keeps changing.
Using std::set for the members is too expensive, and we actually don't
need them sorted all the time, only at leader changes.
We could keep both a set and a vector, and keep them mostly sorted and
resort as necessary, or use a set and a fibheap, but all of this seems
premature.
After running some statistics, we are able to avoid the vast majority
of sorting by keeping a "next leader" field. Most congruence classes only have
leader changes once or twice during GVN.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28594
llvm-svn: 291968
Other than on COFF with incremental linking, global metadata should
not need any extra alignment.
Differential Revision: https://reviews.llvm.org/D28628
llvm-svn: 291859
Summary:
We can sometimes end up with multiple copies of a local function that
have the same GUID in the index. This happens when there are local
functions with the same name that are in different source files with the
same name (but in different directories), and they were compiled in
their own directory so had the same path at compile time.
In this case make sure we import the copy in the caller's module. While
it isn't a correctness problem (the renamed reference which is based on the
module IR hash will be unique since the module must have had an
externally visible function that was imported), importing the wrong copy
will result in lost performance opportunity since it won't be referenced
and inlined.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28440
llvm-svn: 291841
It was always zero. When we move a store from `initial` to its
own congruency class, we end up with a negative store count, which
is obviously wrong.
Also, while here, change StoreCount to be signed so that the assertions
actually fire.
Ack'ed by Daniel Berlin.
llvm-svn: 291725
When using profiling and ASan together (-fprofile-instr-generate -fcoverage-mapping -fsanitize=address), at least on Darwin, the section of globals that ASan emits (__asan_globals) is misaligned and starts at an odd offset. This really doesn't have anything to do with profiling, but it triggers the issue because profiling emits a string section, which can have arbitrary size. This patch changes the alignment to sizeof(GlobalStruct).
Differential Revision: https://reviews.llvm.org/D28573
llvm-svn: 291715
This means that we can use a shorter instruction sequence in the case where
the size is a power of two and on the boundary between two representations.
Differential Revision: https://reviews.llvm.org/D28421
llvm-svn: 291706
classes, and updating checking to allow for equivalence through
reachability.
(Sadly, the checking here is not perfect, and can't be made perfect,
so we'll have to disable it after we are satisfied with correctness.
Right now it is just "very unlikely" to happen.)
llvm-svn: 291698
The removed assert seems bogus - it's perfectly legal for the roots of the
vectorized subtrees to be equal even if the original scalar values aren't,
if the original scalars happen to be equivalent.
This fixes PR31599.
Differential Revision: https://reviews.llvm.org/D28539
llvm-svn: 291692
Summary:
Revert LowerTypeTests: Split the pass in two: a resolution phase and a lowering phase.
This change separates how type identifiers are resolved from how intrinsic
calls are lowered. All information required to lower an intrinsic call
is stored in a new TypeIdLowering data structure. The idea is that this
data structure can either be initialized using the module itself during
regular LTO, or using the module summary in ThinLTO backends.
Original URL: https://reviews.llvm.org/D28341
Reviewers: pcc
Subscribers: mehdi_amini, llvm-commits
Differential Revision: https://reviews.llvm.org/D28532
llvm-svn: 291684
Here's my second try at making @llvm.assume processing more efficient. My
previous attempt, which leveraged operand bundles, r289755, didn't end up
working: it did make assume processing more efficient but eliminating the
assumption cache made ephemeral value computation too expensive. This is a
more-targeted change. We'll keep the assumption cache, but extend it to keep a
map of affected values (i.e. values about which an assumption might provide
some information) to the corresponding assumption intrinsics. This allows
ValueTracking and LVI to find assumptions relevant to the value being queried
without scanning all assumptions in the function. The fact that ValueTracking
started doing O(number of assumptions in the function) work, for every
known-bits query, has become prohibitively expensive in some cases.
As discussed during the review, this is a pragmatic fix that, longer term, will
likely be replaced by a more-principled solution (perhaps based on an extended
SSA form).
Differential Revision: https://reviews.llvm.org/D28459
llvm-svn: 291671
the latter to the Transforms library.
While the loop PM uses an analysis to form the IR units, the current
plan is to have the PM itself establish and enforce both loop simplified
form and LCSSA. This would be a layering violation in the analysis
library.
Fundamentally, the idea behind the loop PM is to *transform* loops in
addition to running passes over them, so it really seemed like the most
natural place to sink this was into the transforms library.
We can't just move *everything* because we also have loop analyses that
rely on a subset of the invariants. So this patch splits the the loop
infrastructure into the analysis management that has to be part of the
analysis library, and the transform-aware pass manager.
This also required splitting the loop analyses' printer passes out to
the transforms library, which makes sense to me as running these will
transform the code into LCSSA in theory.
I haven't split the unittest though because testing one component
without the other seems nearly intractable.
Differential Revision: https://reviews.llvm.org/D28452
llvm-svn: 291662
updated instructions:
pmulld, pmullw, pmulhw, mulsd, mulps, mulpd, divss, divps, divsd, divpd, addpd and subpd.
special optimization case which replaces pmulld with pmullw\pmulhw\pshuf seq.
In case if the real operands bitwidth <= 16.
Differential Revision: https://reviews.llvm.org/D28104
llvm-svn: 291657
arguments much like the CGSCC pass manager.
This is a major redesign following the pattern establish for the CGSCC layer to
support updates to the set of loops during the traversal of the loop nest and
to support invalidation of analyses.
An additional significant burden in the loop PM is that so many passes require
access to a large number of function analyses. Manually ensuring these are
cached, available, and preserved has been a long-standing burden in LLVM even
with the help of the automatic scheduling in the old pass manager. And it made
the new pass manager extremely unweildy. With this design, we can package the
common analyses up while in a function pass and make them immediately available
to all the loop passes. While in some cases this is unnecessary, I think the
simplicity afforded is worth it.
This does not (yet) address loop simplified form or LCSSA form, but those are
the next things on my radar and I have a clear plan for them.
While the patch is very large, most of it is either mechanically updating loop
passes to the new API or the new testing for the loop PM. The code for it is
reasonably compact.
I have not yet updated all of the loop passes to correctly leverage the update
mechanisms demonstrated in the unittests. I'll do that in follow-up patches
along with improved FileCheck tests for those passes that ensure things work in
more realistic scenarios. In many cases, there isn't much we can do with these
until the loop simplified form and LCSSA form are in place.
Differential Revision: https://reviews.llvm.org/D28292
llvm-svn: 291651
These are interesting again because the user may not be aware that this
is a common reason preventing LICM.
A const is removed from an instruction pointer declaration in order to
pass it to ORE.
Differential Revision: https://reviews.llvm.org/D27940
llvm-svn: 291649
This patch reverts r291588: [PGO] Turn off comdat renaming in IR PGO by default,
as we are seeing some hash mismatches in our internal tests.
llvm-svn: 291621
Some of the callers are artificially limiting this transform to integer types;
this should make it easier to incrementally remove that restriction.
llvm-svn: 291620
Summary:
This fixes Transforms/LoopUnroll/runtime-loop3.ll which failed with
EXTENSIVE_DEBUG, because the cloned basic blocks were not added to the
correct sub-loops in LoopUnrollRuntime.cpp.
Reviewers: dexonsmith, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28482
llvm-svn: 291619
Move the code to update LoopInfo for cloned basic blocks to
addClonedBlockToLoopInfo, as suggested in
https://reviews.llvm.org/D28482.
llvm-svn: 291614
Bail out instead of asserting when we encounter this situation,
which can actually happen.
The reason the test uses the new PM is that the "bad" phi, incidentally, gets
cleaned up by LoopSimplify. But LICM can create this kind of phi and preserve
loop simplify form, so the cleanup has no chance to run.
This fixes PR31190.
We may want to solve this in a less conservative manner, since this phi is
actually uniform within the inner loop (or we may want LICM to output a cleaner
promotion to begin with).
Differential Revision: https://reviews.llvm.org/D28490
llvm-svn: 291589
Summary:
In IR PGO we append the function hash to comdat functions to avoid the
potential hash mismatch. This turns out not legal in some cases: if the comdat
function is address-taken and used in comparison. Renaming changes the semantic.
This patch turns off comdat renaming by default.
To alleviate the hash mismatch issue, we now rename the profile variable
for comdat functions. Profile allows co-existing multiple versions of profiles
with different hash value. The inlined copy will always has the correct profile
counter. The out-of-line copy might not have the correct count. But we will
not have the bogus mismatch warning.
Reviewers: davidxl
Subscribers: llvm-commits, xur
Differential Revision: https://reviews.llvm.org/D28416
llvm-svn: 291588
In some cases StructurizeCfg updates root node, but dominator info
remains unchanges, it causes crash when expensive checks are enabled.
To cope with this problem a new method was added to DominatorTreeBase
that allows adding new root nodes, it is called in StructurizeCfg to
put dominator tree in sync.
This change fixes PR27488.
Differential Revision: https://reviews.llvm.org/D28114
llvm-svn: 291530
This patch delays the fix-up step for external induction variable users until
after the dominator tree has been properly updated. This should fix PR30742.
The SCEVExpander in InductionDescriptor::transform can generate code in the
wrong location if the dominator tree is not up-to-date. We should work towards
keeping the dominator tree up-to-date throughout the transformation.
Reference: https://llvm.org/bugs/show_bug.cgi?id=30742
Differential Revision: https://reviews.llvm.org/D28168
llvm-svn: 291462
fabs(x * x) is not generally safe to assume x is positive if x is a NaN.
This is also less general than it could be, so this will be replaced
with a transformation on the intrinsic.
llvm-svn: 291359
Also move command line handling out of the pass constructor and into
a separate function.
Differential Revision: https://reviews.llvm.org/D28422
llvm-svn: 291323
Summary: LLVM's non-standard notion of phi nodes means we can't both try to substitute for undef in phi nodes *and* use phi nodes as leaders all the time. This changes NewGVN to use the same semantics as SimplifyPHINode to decide which phi nodes are equivalent.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28312
llvm-svn: 291308
Summary:
r285871 introduced an assert that was overly aggressive in the case
of a same-named local in different same-named files (in different
directories), where the source name and therefore the GUID ended up
the same because the files were compiled in their own directory without
any leading path. Change the handling in the promotion logic to get
the summary for the version in that module.
This also exposed an issue where we are not always importing the
right copy, which is a performance not correctness issue (because
the renaming is based on the module hash which must be different,
see the bug report for details). I will fix that as a follow-on.
Fixes PR31561.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28411
llvm-svn: 291304
This flag is used to track global registration in Mach-O and it doesn't need to be exported and visible.
Differential Revision: https://reviews.llvm.org/D28250
llvm-svn: 291289
This is fixing a bug where Loop Vectorization is widening a load but
with a lower alignment. Hoisting the load without propagating the alignment
will allow inst-combine to later deduce a higher alignment that what the pointer
actually is.
Differential Revision: https://reviews.llvm.org/D28408
llvm-svn: 291281
order to avoid jumpy line tables. Calls are left alone because they may be inlined.
Differential Revision: https://reviews.llvm.org/D28390
llvm-svn: 291258
This change separates how type identifiers are resolved from how intrinsic
calls are lowered. All information required to lower an intrinsic call
is stored in a new TypeIdLowering data structure. The idea is that this
data structure can either be initialized using the module itself during
regular LTO, or using the module summary in ThinLTO backends.
Differential Revision: https://reviews.llvm.org/D28341
llvm-svn: 291205
Summary:
Using the linker-supplied list of "preserved" symbols, we can compute
the list of "dead" symbols, i.e. the one that are not reachable from
a "preserved" symbol transitively on the reference graph.
Right now we are using this information to mark these functions as
non-eligible for import.
The impact is two folds:
- Reduction of compile time: we don't import these functions anywhere
or import the function these symbols are calling.
- The limited number of import/export leads to better internalization.
Patch originally by Mehdi Amini.
Reviewers: mehdi_amini, pcc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D23488
llvm-svn: 291177
Promotion is always legal when a store within the loop is guaranteed to execute.
However, this is not a necessary condition - for promotion to be memory model
semantics-preserving, it is enough to have a store that dominates every exit
block. This is because if the store dominates every exit block, the fact the
exit block was executed implies the original store was executed as well.
Differential Revision: https://reviews.llvm.org/D28147
llvm-svn: 291171
Summary:
Preheader instruction's operands will always be invariant w.r.t. the loop which its the preheader
for.
Memory aliases are handled in canSinkOrHoistInst.
Reviewers: danielcdh, davidxl
Subscribers: mzolotukhin, llvm-commits
Differential Revision: https://reviews.llvm.org/D28270
llvm-svn: 291132
Summary:
This adds a new summary flag NotEligibleToImport that subsumes
several existing flags (NoRename, HasInlineAsmMaybeReferencingInternal
and IsNotViableToInline). It also subsumes the checking of references
on the summary that was being done during the thin link by
eligibleForImport() for each candidate. It is much more efficient to
do that checking once during the per-module summary build and record
it in the summary.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28169
llvm-svn: 291108
This code seems to be target dependent which may not be the same for all targets.
Passed the decision whether the given stride is complex or not to the target by sending stride information via SCEV to getAddressComputationCost instead of 'IsComplex'.
Specifically at X86 targets we dont see any significant address computation cost in case of the strided access in general.
Differential Revision: https://reviews.llvm.org/D27518
llvm-svn: 291106
Set up basic YAML I/O support for module summaries, plumb the summary into
the pass and add a few command line flags to test YAML I/O support. Bitcode
support to come separately, as will the code in LowerTypeTests that actually
uses the summary. Also add a couple of tests that pass by virtue of the pass
doing nothing with the summary (which happens to be the correct thing to do
for those tests).
Differential Revision: https://reviews.llvm.org/D28041
llvm-svn: 291069
performing partial redundancy elimination (PRE). Not doing so can cause jumpy line
tables and confusing (though correct) source attributions.
Differential Revision: https://reviews.llvm.org/D27857
llvm-svn: 291037
Summary:
This is a relatively simple scheme: we use the index emitted in the
bitcode to avoid loading all the global metadata. Instead we load
the index with their position in the bitcode so that we can load each
of them individually. Materializing the global metadata block in this
condition only triggers loading the named metadata, and the ones
referenced from there (transitively). When materializing a function,
metadata from the global block are loaded lazily as they are
referenced.
Two main current limitations are:
1) Global values other than functions are not materialized on demand,
so we need to eagerly load METADATA_GLOBAL_DECL_ATTACHMENT records
(and their transitive dependencies).
2) When we load a single metadata, we don't recurse on the operands,
instead we use a placeholder or a temporary metadata. Unfortunately
tepmorary nodes are very expensive. This is why we don't have it
always enabled and only for importing.
These two limitations can be lifted in a subsequent improvement if
needed.
With this change, the total link time of opt with ThinLTO and Debug
Info enabled is going down from 282s to 224s (~20%).
Reviewers: pcc, tejohnson, dexonsmith
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28113
llvm-svn: 291027
This reapplies r289828 (reverted in r289833 as it broke the address sanitizer). The
debugloc is now only set when the instruction is not a call, as this causes the
verifier to assert (the inliner requires an inlinable callsite to have a debug loc
if the caller and callee have debug info).
Original commit message:
Simplify CFG will try to sink the last instruction in a series of basic blocks,
creating a "common" instruction in the successor block (sinkLastInstruction).
When it does this, the debug location of the single instruction should be the
merged debug locations of the commoned instructions.
Original review: https://reviews.llvm.org/D27590
llvm-svn: 290973
We can perform the following:
(add (zext (add nuw X, C1)), C2) -> (zext (add nuw X, C1+C2))
This is only possible if C2 is negative and C2 is greater than or equal to negative C1.
llvm-svn: 290927
I wrote this patch before seeing the comment in:
https://reviews.llvm.org/D27114
...that suggests we should actually be canonicalizing the other way.
So just in case we decide this is the right way, we might as well
have a cleaner implementation.
llvm-svn: 290912
Summary:
Regardless how the loop body weight is distributed, we should preserve
total loop body weight. i.e. we should have same weight reaching the body of the loop
or its duplicates in peeled and unpeeled case.
Reviewers: mkuper, davidxl, anemet
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28179
llvm-svn: 290833
Apparently my suggestion of using ternary doesn't really work
as clang complains about incompatible types on LHS and RHS. Some
GCC versions happen to accept the code but clang behaviour is
correct here.
llvm-svn: 290822
Summary:
This avoids the very fragile code for null expressions. We could also use a denseset that tracks which things have null expressions instead, but that seems pretty fragile and premature optimization.
This resolves a number of infinite loop cases, test reductions coming.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28193
llvm-svn: 290816
Summary: Previously, we tried to fix up the equivalences during symbolic evaluation. This does not work. Now, we change the equivalences during congruence finding, where it belongs. We also initialize the equivalence table to give a maximal answer.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28192
llvm-svn: 290815
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.
llvm-svn: 290760
This is similar to the allocfn case - if an alloca is not captured, then it's
necessarily thread-local.
Differential Revision: https://reviews.llvm.org/D28170
llvm-svn: 290738
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.
Reviewers: mzolotukhin, chandlerc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D26989
llvm-svn: 290737
"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.
llvm-svn: 290735
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.
llvm-svn: 290733
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.
llvm-svn: 290729
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.
Instead, bail immediately if we don't have both.
llvm-svn: 290728
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)
Hopefully NFC.
llvm-svn: 290726
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.
llvm-svn: 290704
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
Reviewers: chandlerc, davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28129
llvm-svn: 290699
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).
llvm-svn: 290685
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.
llvm-svn: 290649
This adds a combine that canonicalizes a chain of inserts which broadcasts
a value into a single insert + a splat shufflevector.
This fixes PR31286.
Differential Revision: https://reviews.llvm.org/D27992
llvm-svn: 290641
skipping indirectly recursive inline chains.
To do this, we implicitly build an inline stack for each callsite and
check prior to inlining that doing so would not form a cycle. This uses
the exact same technique and even shares some code with the legacy PM
inliner.
This solution remains deeply unsatisfying to me because it means we
cannot actually iterate the inliner externally. Doing so would not be
able to easily detect and avoid such cycles. Some day I would very much
like to have a solution that works without this internal state to detect
cycles, but this is not that day.
llvm-svn: 290590
PMULDQ/PMULUDQ vXi64 instructions only use the even numbered v2Xi32 input elements which SimplifyDemandedVectorElts should try and use.
This builds on r290554 which added supported for 128 and 256-bit.
llvm-svn: 290582
An earlier commit added support for unmasked scalar operations. At that time isel wouldn't generate an optimal sequence for masked operations, but that has now been fixed.
llvm-svn: 290566
inside of `InlineFunction`. Prior to this, call instructions are
specifically being rewritten and replaced within the inlined region,
invalidating some of the call sites.
Several of these regions are using the same technique to walk the
inlined region so this seems clearly safe up to this point.
I've also added a short circuit to the scan for call sites based on what
other code is doing.
With this, the most common crash I've found in the new inliner code is
fixed. I've turned it on for another test case that covers this
scenario.
I'll make my way through most of the other inliner test cases
just to get some easy coverage next.
llvm-svn: 290562
removing fully-dead comdats without removing dead entries in comdats
with live members.
This factors the core logic out of the current inliner's internals to
a reusable utility and leverages that in both places. The factored out
code should also be (minorly) more efficient in cases where we have very
few dead functions or dead comdats to consider.
I've added a test case to cover this behavior of the always inliner.
This is the last significant bug in the new PM's always inliner I've
found (so far).
llvm-svn: 290557
PMULDQ/PMULUDQ vXi64 instructions only use the even numbered v2Xi32 input elements which SimplifyDemandedVectorElts should try and use.
Differential Revision: https://reviews.llvm.org/D28119
llvm-svn: 290554
Mostly use a bit more idiomatic C++ where we can,
so we can combine some things later.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28111
llvm-svn: 290550
Summary:
I only do this for unmasked cases for now because isel is failing to fold the mask. I'll try to fix that soon.
I'll do the same thing for packed add/sub/mul/div in a future patch.
Reviewers: delena, RKSimon, zvi, craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27879
llvm-svn: 290535
Summary:
This patch adds support for converting the masked vpermv intrinsics into shufflevector instructions if the indices are constants.
We also need to wrap a select instruction around the shuffle to take care of the masking part. InstCombine will take care of optimizing the select if the mask is constant so I didn't bother checking for that.
Reviewers: zvi, delena, spatel, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27825
llvm-svn: 290530
whether functions are removed, and fix the new PM's always inliner to
actually pass this test.
Without this, the new PM's always inliner leaves all the functions
kicking around which won't work out very well given the semantics of
always inline.
Doing this really highlights how frustrating the current alwaysinline
semantic contract is though -- why can we put it on *external*
functions, etc?
Also I've added a number of tricky and interesting test cases for
removing functions with the always inliner. There is one remaining case
not handled -- fully removing comdats -- and I've left a FIXME about
this.
llvm-svn: 290457
The pass creates some state which expects to be cleaned up by
a later instance of the same pass. opt-bisect happens to expose
this not ideal design because calling skipLoop() will result in
this state not being cleaned up at times and an assertion firing
in `doFinalization()`. Chandler tells me the new pass manager will
give us options to avoid these design traps, but until it's not ready,
we need a workaround for the current pass infrastructure. Fix provided
by Andy Kaylor, see the review for a complete discussion.
Differential Revision: https://reviews.llvm.org/D25848
llvm-svn: 290427
Use a dummy private function with inline asm calls instead of module
level asm blocks for CFI jumptables.
The main advantage is that now jumptable codegen can be affected by
the function attributes (like target_cpu on ARM). Module level asm
gets the default subtarget based on the target triple, which is often
not good enough.
This change also uses asm constraints/arguments to reference
jumptable targets and aliases directly. We no longer do asm name
mangling in an IR pass.
Differential Revision: https://reviews.llvm.org/D28012
llvm-svn: 290384
The code have been developed by Daniel Berlin over the years, and
the new implementation goal is that of addressing shortcomings of
the current GVN infrastructure, i.e. long compile time for large
testcases, lack of phi predication, no load/store value numbering
etc...
The current code just implements the "core" GVN algorithm, although
other pieces (load coercion, phi handling, predicate system) are
already implemented in a branch out of tree. Once the core is stable,
we'll start adding pieces on top of the base framework.
The test currently living in test/Transform/NewGVN are a copy
of the ones in GVN, with proper `XFAIL` (missing features in NewGVN).
A flag will be added in a future commit to enable NewGVN, so that
interested parties can exercise this code easily.
Differential Revision: https://reviews.llvm.org/D26224
llvm-svn: 290346
from the old pass manager in the new one.
I'm not trying to support (initially) the numerous options that are
currently available to customize the pass pipeline. If we end up really
wanting them, we can add them later, but I suspect many are no longer
interesting. The simplicity of omitting them will help a lot as we sort
out what the pipeline should look like in the new PM.
I've also documented to the best of my ability *why* each pass or group
of passes is used so that reading the pipeline is more helpful. In many
cases I think we have some questionable choices of ordering and I've
left FIXME comments in place so we know what to come back and revisit
going forward. But for now, I've left it as similar to the current
pipeline as I could.
Lastly, I've had to comment out several places where passes are not
ported to the new pass manager or where the loop pass infrastructure is
not yet ready. I did at least fix a few bugs in the loop pass
infrastructure uncovered by running the full pipeline, but I didn't want
to go too far in this patch -- I'll come back and re-enable these as the
infrastructure comes online. But I'd like to keep the comments in place
because I don't want to lose track of which passes need to be enabled
and where they go.
One thing that seemed like a significant API improvement was to require
that we don't build pipelines for O0. It seems to have no real benefit.
I've also switched back to returning pass managers by value as at this
API layer it feels much more natural to me for composition. But if
others disagree, I'm happy to go back to an output parameter.
I'm not 100% happy with the testing strategy currently, but it seems at
least OK. I may come back and try to refactor or otherwise improve this
in subsequent patches but I wanted to at least get a good starting point
in place.
Differential Revision: https://reviews.llvm.org/D28042
llvm-svn: 290325
In r267672, where the loop distribution pragma was introduced, I tried
it hard to keep the old behavior for opt: when opt is invoked
with -loop-distribute, it should distribute the loop (it's off by
default when ran via the optimization pipeline).
As MichaelZ has discovered this has the unintended consequence of
breaking a very common developer work-flow to reproduce compilations
using opt: First you print the pass pipeline of clang
with -debug-pass=Arguments and then invoking opt with the returned
arguments.
clang -debug-pass will include -loop-distribute but the pass is invoked
with default=off so nothing happens unless the loop carries the pragma.
While through opt (default=on) we will try to distribute all loops.
This changes opt's default to off as well to match clang. The tests are
modified to explicitly enable the transformation.
llvm-svn: 290235
No existing client is passing a non-null value here. This will come back
in a slightly different form as part of the type identifier summary work.
Differential Revision: https://reviews.llvm.org/D28006
llvm-svn: 290222
We're currently doing nearly the same thing for @llvm.objectsize in
three different places: two of them are missing checks for overflow,
and one of them could subtly break if InstCombine gets much smarter
about removing alloc sites. Seems like a good idea to not do that.
llvm-svn: 290214
Make it clear that TripCount is the upper bound of the iteration on which
control exits LatchBlock.
Differential Revision: https://reviews.llvm.org/D26675
llvm-svn: 290199
This doesn't implement *every* feature of the existing inliner, but
tries to implement the most important ones for building a functional
optimization pipeline and beginning to sort out bugs, regressions, and
other problems.
Notable, but intentional omissions:
- No alloca merging support. Why? Because it isn't clear we want to do
this at all. Active discussion and investigation is going on to remove
it, so for simplicity I omitted it.
- No support for trying to iterate on "internally" devirtualized calls.
Why? Because it adds what I suspect is inappropriate coupling for
little or no benefit. We will have an outer iteration system that
tracks devirtualization including that from function passes and
iterates already. We should improve that rather than approximate it
here.
- Optimization remarks. Why? Purely to make the patch smaller, no other
reason at all.
The last one I'll probably work on almost immediately. But I wanted to
skip it in the initial patch to try to focus the change as much as
possible as there is already a lot of code moving around and both of
these *could* be skipped without really disrupting the core logic.
A summary of the different things happening here:
1) Adding the usual new PM class and rigging.
2) Fixing minor underlying assumptions in the inline cost analysis or
inline logic that don't generally hold in the new PM world.
3) Adding the core pass logic which is in essence a loop over the calls
in the nodes in the call graph. This is a bit duplicated from the old
inliner, but only a handful of lines could realistically be shared.
(I tried at first, and it really didn't help anything.) All told,
this is only about 100 lines of code, and most of that is the
mechanics of wiring up analyses from the new PM world.
4) Updating the LazyCallGraph (in the new PM) based on the *newly
inlined* calls and references. This is very minimal because we cannot
form cycles.
5) When inlining removes the last use of a function, eagerly nuking the
body of the function so that any "one use remaining" inline cost
heuristics are immediately refined, and queuing these functions to be
completely deleted once inlining is complete and the call graph
updated to reflect that they have become dead.
6) After all the inlining for a particular function, updating the
LazyCallGraph and the CGSCC pass manager to reflect the
function-local simplifications that are done immediately and
internally by the inline utilties. These are the exact same
fundamental set of CG updates done by arbitrary function passes.
7) Adding a bunch of test cases to specifically target CGSCC and other
subtle aspects in the new PM world.
Many thanks to the careful review from Easwaran and Sanjoy and others!
Differential Revision: https://reviews.llvm.org/D24226
llvm-svn: 290161
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
This reapplies r289902 with additional testcase upgrades and a change
to the Bitcode record for DIGlobalVariable, that makes upgrading the
old format unambiguous also for variables without DIExpressions.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 290153
Background/motivation - I was circling back around to:
https://llvm.org/bugs/show_bug.cgi?id=28296
I made a simple patch for that and noticed some regressions, so added test cases for
those with rL281055, and this is hopefully the minimal fix for just those cases.
But as you can see from the surrounding untouched folds, we are missing commuted patterns
all over the place, and of course there are no regression tests to cover any of those cases.
We could sprinkle "m_c_" dust all over this file and catch most of the missing folds, but
then we still wouldn't have test coverage, and we'd still miss some fraction of commuted
patterns because they require adjustments to the match order.
I'm aware of the concern about the potential compile-time performance impact of adding
matches like this (currently being discussed on llvm-dev), but I don't think there's any
evidence yet to suggest that handling commutative pattern matching more thoroughly is not
a worthwhile goal of InstCombine.
Differential Revision: https://reviews.llvm.org/D24419
llvm-svn: 290067
This reverts commit 289920 (again).
I forgot to implement a Bitcode upgrade for the case where a DIGlobalVariable
has not DIExpression. Unfortunately it is not possible to safely upgrade
these variables without adding a flag to the bitcode record indicating which
version they are.
My plan of record is to roll the planned follow-up patch that adds a
unit: field to DIGlobalVariable into this patch before recomitting.
This way we only need one Bitcode upgrade for both changes (with a
version flag in the bitcode record to safely distinguish the record
formats).
Sorry for the churn!
llvm-svn: 289982
This patch reapplies r289863. The original patch was reverted because it
exposed a bug causing the loop vectorizer to crash in the Python runtime on
PPC. The underlying issue was fixed with r289958.
llvm-svn: 289975
After r288909, instructions feeding predicated instructions may be scalarized
if profitable. Since these instructions will remain scalar, we shouldn't
attempt to type-shrink them. We should only truncate vector types to their
minimal bit widths. This bug was exposed by enabling the vectorization of loops
containing conditional stores by default.
llvm-svn: 289958
stores by default
This uncovers a crasher in the loop vectorizer on PPC when building the
Python runtime. I'll send the testcase to the review thread for the
original commit.
llvm-svn: 289934
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
This reapplies r289902 with additional testcase upgrades.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 289920
Summary:
Instead of checking whether a global referenced by a function being
imported is defined in the same module, speculatively always add the
referenced globals to the module's export list. After all imports are
computed, for each module prune any not in its defined set from its
export list.
For a huge C++ app with aggressive importing thresholds, even with
D27687 we spent a lot of time invoking modulePath() from
exportGlobalInModule (modulePath() was still the 2nd hottest routine in
profile). The reason is that with comdat/linkonce the summary lists for
each GUID can be long. For the app in question, for example, we were
invoking exportGlobalInModule almost 2 million times, and we traversed
an average of 63 entries in the summary list each time.
This patch reduced the thin link time for the app by about 10% (on top
of D27687) when using aggressive importing thresholds, and about 3.5% on
average with default importing thresholds.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27755
llvm-svn: 289918
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 289902
This pass prepares a module containing type metadata for ThinLTO by splitting
it into regular and thin LTO parts if possible, and writing both parts to
a multi-module bitcode file. Modules that do not contain type metadata are
written unmodified as a single module.
All globals with type metadata are added to the regular LTO module, and
the rest are added to the thin LTO module.
Differential Revision: https://reviews.llvm.org/D27324
llvm-svn: 289899
Summary:
We were reinvoking exportGlobalInModule numerous times redundantly.
No need to re-export globals referenced by a global that was already
imported from its module. This resulted in a large speedup in the thin
link for a big application, particularly when importing aggressiveness
was cranked up.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27687
llvm-svn: 289896
Lowering to llvm.cttz() will result in constant folding anyway
if the argument to ffs is a constant. Pointed out by Eli for
fls() in D14590.
llvm-svn: 289888
The code change for D27687 accidentally got committed along with the
main change in r289843. Revert it temporarily, so that I can recommit it
along with its test as intended.
llvm-svn: 289875
Summary:
Thin link efficiency improvement. After adding an importing candidate to
the worklist we might have later added it again with a higher threshold.
Skip it when popped from the worklist if we recorded a higher threshold
than the current worklist entry, it will get processed again at the
higher threshold when that entry is popped.
This required adding the summary's GUID to the worklist, so that it can
be used to query the recorded highest threshold for it when we pop from the
worklist.
Reviewers: mehdi_amini
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27696
llvm-svn: 289867
This patch sets the default value of the "-enable-cond-stores-vec" command line
option to "true".
Differential Revision: https://reviews.llvm.org/D27814
llvm-svn: 289863
Now that a new API to merge debug locations has been committed at r289661 (see
review D26256 for more details), we can use it to "improve" the code added by
revision r280995.
Instead of nulling the debugloc of a commoned instruction, we use the 'merged'
debug location. At the moment, this is just a no functional change since
function `DILocation::getMergedLocation()` is just a stub and would always
return a null location.
Differential Revision: https://reviews.llvm.org/D27804
llvm-svn: 289862
Min/max canonicalization (r287585) exposes the fact that we're missing combines for min/max patterns.
This patch won't solve the example that was attached to that thread, so something else still needs fixing.
The line between InstCombine and InstSimplify gets blurry here because sometimes the icmp instruction that
we want to fold to already exists, but sometimes it's the swapped form of what we want.
Corresponding changes for smax/umin/umax to follow.
Differential Revision: https://reviews.llvm.org/D27531
llvm-svn: 289855
This is split out from D27696, since it turned out to be a bug fix and
not part of the NFC efficiency change.
Keep the same adjusted (possibly decayed) threshold in both the worklist
and the ImportList. Otherwise if we encountered it first along a cold
path, the callee would be added to the worklist with a lower decayed
threshold than when it is later encountered along a hot path. But the
logic uses the threshold recorded in the ImportList entry to check if
we should re-add it, and without this patch the threshold recorded there
is the same along both paths so we don't re-add it. Using the
same possibly decayed threshold in the ImportList ensures we re-add it
later with the higher non-decayed hot path threshold.
llvm-svn: 289843
Chandler Carruth