In order to always import the same copy of a linkonce function,
even when encountering it with different thresholds (a higher one then a
lower one), keep track of the summary we decided to import.
This ensures that the backend only gets a single definition to import
for each GUID, so that it doesn't need to choose one.
Move the largest threshold the GUID was considered for import into the
current module out of the ImportMap (which is part of a larger map
maintained across the whole index), and into a new map just maintained
for the current module we are computing imports for. This saves some
memory since we no longer have the thresholds maintained across the
whole index (and throughout the in-process backends when doing a normal
non-distributed ThinLTO build), at the cost of some additional
information being maintained for each invocation of ComputeImportForModule
(the selected summary pointer for each import).
There is an additional map lookup for each callee being considered for
importing, however, this was able to subsume a map lookup in the
Worklist iteration that invokes computeImportForFunction. We also are
able to avoid calling selectCallee if we already failed to import at the
same or higher threshold.
I compared the run time and peak memory for the SPEC2006 471.omnetpp
benchmark (running in-process ThinLTO backends), as well as for a large
internal benchmark with a distributed ThinLTO build (so just looking at
the thin link time/memory). Across a number of runs with and without
this change there was no significant change in the time and memory.
(I tried a few other variations of the change but they also didn't
improve time or peak memory).
Reviewers: davidxl
Subscribers: mehdi_amini, inglorion, llvm-commits
Differential Revision: https://reviews.llvm.org/D48670
llvm-svn: 337050
Summary:
Currently LowerTypeTests emits jumptable entries for all live external
and address-taken functions; however, we could limit the number of
functions that we emit entries for significantly.
For Cross-DSO CFI, we continue to emit jumptable entries for all
exported definitions. In the non-Cross-DSO CFI case, we only need to
emit jumptable entries for live functions that are address-taken in live
functions. This ignores exported functions and functions that are only
address taken in dead functions. This change uses ThinLTO summary data
(now emitted for all modules during ThinLTO builds) to determine
address-taken and liveness info.
The logic for emitting jumptable entries is more conservative in the
regular LTO case because we don't have summary data in the case of
monolithic LTO builds; however, once summaries are emitted for all LTO
builds we can unify the Thin/monolithic LTO logic to only use summaries
to determine the liveness of address taking functions.
This change is a partial fix for PR37474. It reduces the build size for
nacl_helper by ~2-3%, the reduction is due to nacl_helper compiling in
lots of unused code and unused functions that are address taken in dead
functions no longer being being considered live due to emitted jumptable
references. The reduction for chromium is ~0.1-0.2%.
Reviewers: pcc, eugenis, javed.absar
Reviewed By: pcc
Subscribers: aheejin, dexonsmith, dschuff, mehdi_amini, eraman, steven_wu, llvm-commits, kcc
Differential Revision: https://reviews.llvm.org/D47652
llvm-svn: 337038
This reverts commit r336419: use-after-free on CallGraph::FunctionMap elements
due to the use of a stale iterator in CGPassManager::runOnModule.
The iterator may be invalidated if a pass removes a function, ex.:
llvm::LegacyInlinerBase::inlineCalls
inlineCallsImpl
llvm::CallGraph::removeFunctionFromModule
llvm-svn: 337018
We currently only support binary instructions in the alternate opcode shuffles.
This patch is an initial attempt at adding cast instructions as well, this raises several issues that we probably want to address as we continue to generalize the alternate mechanism:
1 - Duplication of cost determination - we should probably add scalar/vector costs helper functions and get BoUpSLP::getEntryCost to use them instead of determining costs directly.
2 - Support alternate instructions with the same opcode (e.g. casts with different src types) - alternate vectorization of calls with different IntrinsicIDs will require this.
3 - Allow alternates to be a different instruction type - mixing binary/cast/call etc.
4 - Allow passthrough of unsupported alternate instructions - related to PR30787/D28907 'copyable' elements.
Reapplied with fix to only accept 2 different casts if they come from the same source type (PR38154).
Differential Revision: https://reviews.llvm.org/D49135
llvm-svn: 336989
This bug was created by rL335258 because we used to always call instsimplify
after trying the associative folds. After that change it became possible
for subsequent folds to encounter unsimplified code (and potentially assert
because of it).
Instead of carrying changed state through instcombine, we can just return
immediately. This allows instsimplify to run, so we can continue assuming
that easy folds have already occurred.
llvm-svn: 336965
Summary:
This patch is crucial for proving equality laundered/stripped
pointers. eg:
bool foo(A *a) {
return a == std::launder(a);
}
Clang with -fstrict-vtable-pointers will emit something like:
define dso_local zeroext i1 @_Z3fooP1A(%struct.A* %a) {
entry:
%c = bitcast %struct.A* %a to i8*
%call = tail call i8* @llvm.launder.invariant.group.p0i8(i8* %c)
%0 = bitcast %struct.A* %a to i8*
%1 = tail call i8* @llvm.strip.invariant.group.p0i8(i8* %0)
%2 = tail call i8* @llvm.strip.invariant.group.p0i8(i8* %call)
%cmp = icmp eq i8* %1, %2
ret i1 %cmp
}
and because %2 can be replaced with @llvm.strip.invariant.group(%0)
and that %2 and %1 will produce the same value (because strip is readnone)
we can replace compare with true.
Reviewers: rsmith, hfinkel, majnemer, amharc, kuhar
Subscribers: llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D47423
llvm-svn: 336963
While that fold is clearly not happening [anymore],
we do now have separate test cases for these cases,
so we should be ok to slightly adjust these tests
to not potentially loose test coverage.
As suggested by Hiroshi Yamauchi in
https://reviews.llvm.org/D49179#1159345
llvm-svn: 336912
This converts them to what clang is now using for codegen. Unfortunately, there seem to be a few kinks to work out still. I'll try to address with follow up patches.
llvm-svn: 336871
This commit suppresses turning loops like this into "(bitwidth - ctlz(input))".
unsigned foo(unsigned input) {
unsigned num = 0;
do {
++num;
input >>= 1;
} while (input != 0);
return num;
}
The loop version returns a value of 1 for both an input of 0 and an input of 1. Converting to a naive ctlz does not preserve that.
Theoretically we could do better if we checked isKnownNonZero or we could insert a select to handle the divergence. But until we have motivating cases for that, this is the easiest solution.
llvm-svn: 336864
This loop executes one iteration without checking the input value. This produces a count of 1 for an input of 0 and 1. We are turning this into 32 - ctlz(n), but that returns 0 if n is 0.
llvm-svn: 336862
Summary:
https://bugs.llvm.org/show_bug.cgi?id=38123
This pattern will be produced by Implicit Integer Truncation sanitizer,
https://reviews.llvm.org/D48958https://bugs.llvm.org/show_bug.cgi?id=21530
in unsigned case, therefore it is probably a good idea to improve it.
https://rise4fun.com/Alive/Rny
^ there are more opportunities for folds, i will follow up with them afterwards.
Caveat: this somehow exposes a missing opportunities
in `test/Transforms/InstCombine/icmp-logical.ll`
It seems, the problem is in `foldLogOpOfMaskedICmps()` in `InstCombineAndOrXor.cpp`.
But i'm not quite sure what is wrong, because it calls `getMaskedTypeForICmpPair()`,
which calls `decomposeBitTestICmp()` which should already work for these cases...
As @spatel notes in https://reviews.llvm.org/D49179#1158760,
that code is a rather complex mess, so we'll let it slide.
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: yamauchi, majnemer, t.p.northover, llvm-commits
Differential Revision: https://reviews.llvm.org/D49179
llvm-svn: 336834
isKnownNegation() is currently proposed as part of D48754,
but it could be used to make InstSimplify stronger independently
of any abs() improvements.
llvm-svn: 336822
We currently only support binary instructions in the alternate opcode shuffles.
This patch is an initial attempt at adding cast instructions as well, this raises several issues that we probably want to address as we continue to generalize the alternate mechanism:
1 - Duplication of cost determination - we should probably add scalar/vector costs helper functions and get BoUpSLP::getEntryCost to use them instead of determining costs directly.
2 - Support alternate instructions with the same opcode (e.g. casts with different src types) - alternate vectorization of calls with different IntrinsicIDs will require this.
3 - Allow alternates to be a different instruction type - mixing binary/cast/call etc.
4 - Allow passthrough of unsupported alternate instructions - related to PR30787/D28907 'copyable' elements.
Reapplied with fix to only accept 2 different casts if they come from the same source type.
Differential Revision: https://reviews.llvm.org/D49135
llvm-svn: 336812
We currently only support binary instructions in the alternate opcode shuffles.
This patch is an initial attempt at adding cast instructions as well, this raises several issues that we probably want to address as we continue to generalize the alternate mechanism:
1 - Duplication of cost determination - we should probably add scalar/vector costs helper functions and get BoUpSLP::getEntryCost to use them instead of determining costs directly.
2 - Support alternate instructions with the same opcode (e.g. casts with different src types) - alternate vectorization of calls with different IntrinsicIDs will require this.
3 - Allow alternates to be a different instruction type - mixing binary/cast/call etc.
4 - Allow passthrough of unsupported alternate instructions - related to PR30787/D28907 'copyable' elements.
Differential Revision: https://reviews.llvm.org/D49135
llvm-svn: 336804
This was originally intended with D48893, but as discussed there, we
have to make the folds safe from producing extra poison. This should
give the single binop folds the same capabilities as the existing
folds for 2-binops+shuffle.
LLVM binary opcode review: there are a total of 18 binops. There are 7
commutative binops (add, mul, and, or, xor, fadd, fmul) which we already
fold. We're able to fold 6 more opcodes with this patch (shl, lshr, ashr,
fdiv, udiv, sdiv). There are no folds for srem/urem/frem AFAIK. We don't
bother with sub/fsub with constant operand 1 because those are
canonicalized to add/fadd. 7 + 6 + 3 + 2 = 18.
llvm-svn: 336684
The case with 2 variables is more complicated than the case where
we eliminate the shuffle entirely because a shuffle with an undef
mask element creates an undef result.
I'm not aware of any current analysis/transform that recognizes that
undef propagating to a div/rem/shift, but we have to guard against
the possibility.
llvm-svn: 336668
Summary:
Fixed two cases of where PHI nodes need to be updated by lowerswitch.
When lowerswitch find out that the switch default branch is not
reachable it remove the old default and replace it with the most
popular block from the cases, but it forget to update the PHI
nodes in the default block.
The PHI nodes also need to be updated when the switch is replaced
with a single branch.
Reviewers: hans, reames, arsenm
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D47203
llvm-svn: 336659
switch unswitching.
The core problem was that the way we handled unswitching trivial exit
edges through the default successor of a switch. For some reason
I thought the right way to do this was to add a block containing
unreachable and point the default successor at this block. In
retrospect, this has an amazing number of problems.
The first issue is the one that this pass has always worked around -- we
have to *detect* such edges and avoid unswitching them again. This
seemed pretty easy really. You juts look for an edge to a block
containing unreachable. However, this pattern is woefully unsound. So
many things can break it. The amazing thing is that I found a test case
where *simple-loop-unswitch itself* breaks this! When we do
a *non-trivial* unswitch of a switch we will end up splitting this exit
edge. The result will be a default successor that is an exit and
terminates in ... a perfectly normal branch. So the first test case that
I started trying to fix is added to the nontrivial test cases. This is
a ridiculous example that did just amazing things previously. With just
unswitch, it would create 10+ copies of this stuff stamped out. But if
you combine it *just right* with a bunch of other passes (like
simplify-cfg, loop rotate, and some LICM) you can get it to do this
infinitely. Or at least, I never got it to finish. =[
This, in turn, uncovered another related issue. When we are manipulating
these switches after doing a trivial unswitch we never correctly updated
PHI nodes to reflect our edits. As soon as I started changing how these
edges were managed, it became obvious there were more issues that
I couldn't realistically leave unaddressed, so I wrote more test cases
around PHI updates here and ensured all of that works now.
And this, in turn, required some adjustment to how we collect and manage
the exit successor when it is the default successor. That showed a clear
bug where we failed to include it in our search for the outer-most loop
reached by an unswitched exit edge. This was actually already tested and
the test case didn't work. I (wrongly) thought that was due to SCEV
failing to analyze the switch. In fact, it was just a simple bug in the
code that skipped the default successor. While changing this, I handled
it correctly and have updated the test to reflect that we now get
precise SCEV analysis of trip counts for the outer loop in one of these
cases.
llvm-svn: 336646
getSafeVectorConstantForBinop() was calling getBinOpIdentity() assuming
that the constant we wanted was operand 1 (RHS). That's wrong, but I
don't think we could expose a bug or even a suboptimal fold from that
because the callers have other guards for any binop that would have
been affected.
llvm-svn: 336617
Summary:
Support for this option is needed for building Linux kernel.
This is a very frequently requested feature by kernel developers.
More details : https://lkml.org/lkml/2018/4/4/601
GCC option description for -fdelete-null-pointer-checks:
This Assume that programs cannot safely dereference null pointers,
and that no code or data element resides at address zero.
-fno-delete-null-pointer-checks is the inverse of this implying that
null pointer dereferencing is not undefined.
This feature is implemented in LLVM IR in this CL as the function attribute
"null-pointer-is-valid"="true" in IR (Under review at D47894).
The CL updates several passes that assumed null pointer dereferencing is
undefined to not optimize when the "null-pointer-is-valid"="true"
attribute is present.
Reviewers: t.p.northover, efriedma, jyknight, chandlerc, rnk, srhines, void, george.burgess.iv
Reviewed By: efriedma, george.burgess.iv
Subscribers: eraman, haicheng, george.burgess.iv, drinkcat, theraven, reames, sanjoy, xbolva00, llvm-commits
Differential Revision: https://reviews.llvm.org/D47895
llvm-svn: 336613
In non-zero address spaces, we were reporting that an object at `null`
always occupies zero bytes. This is incorrect in many cases, so just
return `unknown` in those cases for now.
Differential Revision: https://reviews.llvm.org/D48860
llvm-svn: 336611
As discussed in D49047 / D48987, shift-by-undef produces poison,
so we can't use undef vector elements in that case..
Note that we need to extend this for poison-generating flags,
and there's a proposal to create poison from FMF in D47963,
llvm-svn: 336562
As noted in D48987, there are many different ways for this transform to go wrong.
In particular, the poison potential for shifts means we have to more careful with those ops.
I added tests to make that behavior visible for all of the different cases that I could find.
This is a partial fix. To make this review easier, I did not make changes for the single binop
pattern (handled in foldSelectShuffleWith1Binop()). I also left out some potential optimizations
noted with TODO comments. I'll follow-up once we're confident that things are correct here.
The goal is to correct all marked FIXME tests to either avoid the shuffle transform or do it safely.
Note that distinguishing when the shuffle mask contains undefs and using getBinOpIdentity() allows
for some improvements to div/rem patterns, so there are wins along with the missed opportunities
and fixes.
Differential Revision: https://reviews.llvm.org/D49047
llvm-svn: 336546
r335553 with the non-trivial unswitching of switches.
The code correctly updated most aspects of the CFG and analyses, but
missed some crucial aspects:
1) When multiple cases have the same successor, we unswitch that
a single time and replace the switch with a direct branch. The CFG
here is correct, but the target of this direct branch may have had
a PHI node with multiple entries in it.
2) When we still have to clone a successor of the switch into an
unswitched copy of the loop, we'll delete potentially multiple edges
entering this successor, not just one.
3) We also have to delete multiple edges entering the successors in the
original loop when they have to be retained.
4) When the "retained successor" *also* occurs as a case successor, we
just assert failed everywhere. This doesn't happen very easily
because its always valid to simply drop the case -- the retained
successor for switches is always the default successor. However, it
is likely possible through some contrivance of different loop passes,
unrolling, and simplifying for this to occur in practice and
certainly there is nothing "invalid" about the IR so this pass needs
to handle it.
5) In the case of #4, we also will replace these multiple edges with
a direct branch much like in #1 and need to collapse the entries in
any PHI nodes to a single enrty.
All of this stems from the delightful fact that the same successor can
show up in multiple parts of the switch terminator, and each of these
are considered a distinct edge for the purpose of PHI nodes (and
iterating the successors and predecessors) but not for unswitching
itself, the dominator tree, or many other things. For the record,
I intensely dislike this "feature" of the IR in large part because of
the complexity it causes in passes like this. We already have a ton of
logic building sets and handling duplicates, and we just had to add
a bunch more.
I've added a complex test case that covers all five of the above failure
modes. I've also added a variation on it where #4 and #5 occur in loop
exit, adding fun where we have an LCSSA PHI node with "multiple entries"
despite have dedicated exits. There were no additional issues found by
this, but it seems a useful corner case to cover with testing.
One thing that working on all of this code has made painfully clear for
me as well is how amazingly inefficient our PHI node representation is
(in terms of the in-memory data structures and the APIs used to update
them). This code has truly marvelous complexity bounds because every
time we remove an entry from a PHI node we do a linear scan to find it
and then a linear update to the data structure to remove it. We could in
theory batch all of the PHI node updates into a single linear walk of
the operands making this much more efficient, but the APIs fight hard
against this and the fact that we have to handle duplicates in the
peculiar manner we do (removing all but one in some cases) makes even
implementing that very tedious and annoying. Anyways, none of this is
new here or specific to loop unswitching. All code in LLVM that updates
PHI node operands suffers from these problems.
llvm-svn: 336536
Summary:
PGOMemOPSize only modifies CFG in a couple of places; thus we can preserve the DominatorTree with little effort.
When optimizing SQLite with -O3, this patch can decrease 3.8% of the numbers of nodes traversed by DFS and 5.7% of the times DominatorTreeBase::recalculation is called.
Reviewers: kuhar, davide, dmgreen
Reviewed By: dmgreen
Subscribers: mzolotukhin, vsk, llvm-commits
Differential Revision: https://reviews.llvm.org/D48914
llvm-svn: 336522
In the 'detectCTLZIdiom' function support for loops that use LSHR instruction instead of ASHR has been added.
This supports creating ctlz from the following code.
int lzcnt(int x) {
int count = 0;
while (x > 0) {
count++;
x = x >> 1;
}
return count;
}
Patch by Olga Moldovanova
Differential Revision: https://reviews.llvm.org/D48354
llvm-svn: 336509
after trivial unswitching.
This PR illustrates that a fundamental analysis update was not performed
with the new loop unswitch. This update is also somewhat fundamental to
the core idea of the new loop unswitch -- we actually *update* the CFG
based on the unswitching. In order to do that, we need to update the
loop nest in addition to the domtree.
For some reason, when writing trivial unswitching, I thought that the
loop nest structure cannot be changed by the transformation. But the PR
helps illustrate that it clearly can. I've expanded this to a number of
different test cases that try to cover the different cases of this. When
we unswitch, we move an exit edge of a loop out of the loop. If this
exit edge changes which loop reached by an exit is the innermost loop,
it changes the parent of the loop. Essentially, this transformation may
hoist the inner loop up the nest. I've added the simple logic to handle
this reliably in the trivial unswitching case. This just requires
updating LoopInfo and rebuilding LCSSA on the impacted loops. In the
trivial case, we don't even need to handle dedicated exits because we're
only hoisting the one loop and we just split its preheader.
I've also ported all of these tests to non-trivial unswitching and
verified that the logic already there correctly handles the loop nest
updates necessary.
Differential Revision: https://reviews.llvm.org/D48851
llvm-svn: 336477
The replaceAllDbgUsesWith utility helps passes preserve debug info when
replacing one value with another.
This improves upon the existing insertReplacementDbgValues API by:
- Updating debug intrinsics in-place, while preventing use-before-def of
the replacement value.
- Falling back to salvageDebugInfo when a replacement can't be made.
- Moving the responsibiliy for rewriting llvm.dbg.* DIExpressions into
common utility code.
Along with the API change, this teaches replaceAllDbgUsesWith how to
create DIExpressions for three basic integer and pointer conversions:
- The no-op conversion. Applies when the values have the same width, or
have bit-for-bit compatible pointer representations.
- Truncation. Applies when the new value is wider than the old one.
- Zero/sign extension. Applies when the new value is narrower than the
old one.
Testing:
- check-llvm, check-clang, a stage2 `-g -O3` build of clang,
regression/unit testing.
- This resolves a number of mis-sized dbg.value diagnostics from
Debugify.
Differential Revision: https://reviews.llvm.org/D48676
llvm-svn: 336451
As discussed in D48987 and D48893, there are many different
ways to go wrong depending on the binop (and as shown here
we already do go wrong in some cases).
llvm-svn: 336450
Previously we only iterated over functions reachable from the set of
external functions in the module. But since some of the passes under
this (notably the always-inliner and coroutine lowerer) are required for
correctness, they need to run over everything.
This just adds an extra layer of iteration over the CallGraph to keep
track of which functions we've already visited and get the next batch of
SCCs.
Should fix PR38029.
llvm-svn: 336419
Better NaN handling for AMDGCN fmed3.
All operands are checked for NaN now. The checks
were moved before the canonicalization to provide
a better mapping from fclamp. Changed the behaviour
of fmed3(x,y,NaN) to return max(x,y) instead of
min(x,y) in light of this. Updated tests as a result
and added some new cases to cover the fix.
Patch by Alan Baker
llvm-svn: 336375
We have bailout hacks based on min/max in various places in instcombine
that shouldn't be necessary. The affected test was added for:
D48930
...which is a consequence of the improvement in:
D48584 (https://reviews.llvm.org/rL336172)
I'm assuming the visitTrunc bailout in this patch was added specifically
to avoid a change from SimplifyDemandedBits, so I'm just moving that
below the EvaluateInDifferentType optimization. A narrow min/max is still
a min/max.
llvm-svn: 336293
When creating `phi` instructions to resume at the scalar part of the loop,
copy the DebugLoc from the original phi over to the new one.
Differential Revision: https://reviews.llvm.org/D48769
llvm-svn: 336256
When zext is EvaluatedInDifferentType, InstCombine
drops the dbg.value intrinsic. This patch tries to
preserve said DI, by inserting the zext's old DI in the
resulting instruction. (Only for integer type for now)
Differential Revision: https://reviews.llvm.org/D48331
llvm-svn: 336254
As the test diffs show, the current users of getBinOpIdentity()
are InstCombine and Reassociate. SLP vectorizer is a candidate
for using this functionality too (D28907).
The InstCombine shuffle improvements are part of the planned
enhancements noted in D48830.
InstCombine actually has several other uses of getBinOpIdentity()
via SimplifyUsingDistributiveLaws(), but we don't call that for
any FP ops. Fixing that might be another part of removing the
custom reassociation in InstCombine that is only done for fadd+fmul.
llvm-svn: 336215
This is the last significant change suggested in PR37806:
https://bugs.llvm.org/show_bug.cgi?id=37806#c5
...though there are several follow-ups noted in the code comments
in this patch to complete this transform.
It's possible that a binop feeding a select-shuffle has been eliminated
by earlier transforms (or the code was just written like this in the 1st
place), so we'll fail to match the patterns that have 2 binops from:
D48401,
D48678,
D48662,
D48485.
In that case, we can try to materialize identity constants for the remaining
binop to fill in the "ghost" lanes of the vector (where we just want to pass
through the original values of the source operand).
I added comments to ConstantExpr::getBinOpIdentity() to show planned follow-ups.
For now, we only handle the 5 commutative integer binops (add/mul/and/or/xor).
Differential Revision: https://reviews.llvm.org/D48830
llvm-svn: 336196
Summary:
When salvaging a dbg.declare/dbg.addr we should not add
DW_OP_stack_value to the DIExpression
(see test/Transforms/InstCombine/salvage-dbg-declare.ll).
Consider this example
%vla = alloca i32, i64 2
call void @llvm.dbg.declare(metadata i32* %vla, metadata !1, metadata !DIExpression())
Instcombine will turn it into
%vla1 = alloca [2 x i32]
%vla1.sub = getelementptr inbounds [2 x i32], [2 x i32]* %vla, i64 0, i64 0
call void @llvm.dbg.declare(metadata [2 x i32]* %vla1.sub, metadata !19, metadata !DIExpression())
If the GEP can be eliminated, then the dbg.declare will be salvaged
and we should get
%vla1 = alloca [2 x i32]
call void @llvm.dbg.declare(metadata [2 x i32]* %vla1, metadata !19, metadata !DIExpression())
The problem was that salvageDebugInfo did not recognize dbg.declare
as being indirect (%vla1 points to the value, it does not hold the
value), so we incorrectly got
call void @llvm.dbg.declare(metadata [2 x i32]* %vla1, metadata !19, metadata !DIExpression(DW_OP_stack_value))
I also made sure that llvm::salvageDebugInfo and
DIExpression::prependOpcodes do not add DW_OP_stack_value to
the DIExpression in case no new operands are added to the
DIExpression. That way we avoid to, unneccessarily, turn a
register location expression into an implicit location expression
in some situations (see test11 in test/Transforms/LICM/sinking.ll).
Reviewers: aprantl, vsk
Reviewed By: aprantl, vsk
Subscribers: JDevlieghere, llvm-commits
Differential Revision: https://reviews.llvm.org/D48837
llvm-svn: 336191
unswitching loops.
Original patch trying to address this was sent in D47624, but that
didn't quite handle things correctly. There are two key principles used
to select whether and how to invalidate SCEV-cached information about
loops:
1) We must invalidate any info SCEV has cached before unswitching as we
may change (or destroy) the loop structure by the act of unswitching,
and make it hard to recover everything we want to invalidate within
SCEV.
2) We need to invalidate all of the loops whose CFGs are mutated by the
unswitching. Notably, this isn't the *entire* loop nest, this is
every loop contained by the outermost loop reached by an exit block
relevant to the unswitch.
And we need to do this even when doing trivial unswitching.
I've added more focused tests that directly check that SCEV starts off
with imprecise information and after unswitching (and simplifying
instructions) re-querying SCEV will produce precise information. These
tests also specifically work to check that an *outer* loop's information
becomes precise.
However, the testing here is still a bit imperfect. Crafting test cases
that reliably fail to be analyzed by SCEV before unswitching and succeed
afterward proved ... very, very hard. It took me several hours and
careful work to build these, and I'm not optimistic about necessarily
coming up with more to cover more elaborate possibilities. Fortunately,
the code pattern we are testing here in the pass is really
straightforward and reliable.
Thanks to Max Kazantsev for the initial work on this as well as the
review, and to Hal Finkel for helping me talk through approaches to test
this stuff even if it didn't come to much.
Differential Revision: https://reviews.llvm.org/D47624
llvm-svn: 336183
This patch changes order of transform in InstCombineCompares to avoid
performing transforms based on ranges which produce complex bit arithmetics
before more simple things (like folding with constants) are done. See PR37636
for the motivating example.
Differential Revision: https://reviews.llvm.org/D48584
Reviewed By: spatel, lebedev.ri
llvm-svn: 336172
Summary:
Comment on Transforms/LoopVersioning/incorrect-phi.ll: With the change
SCEV is able to prove that the loop doesn't wrap-self (due to zext i16
to i64), disabling the entire loop versioning pass. Removed the zext and
just use i64.
Reviewers: sanjoy
Subscribers: jlebar, hiraditya, javed.absar, bixia, llvm-commits
Differential Revision: https://reviews.llvm.org/D48409
llvm-svn: 336140
Summary: It is common to have the following min/max pattern during the intermediate stages of SLP since we only optimize at the end. This patch tries to catch such patterns and allow more vectorization.
%1 = extractelement <2 x i32> %a, i32 0
%2 = extractelement <2 x i32> %a, i32 1
%cond = icmp sgt i32 %1, %2
%3 = extractelement <2 x i32> %a, i32 0
%4 = extractelement <2 x i32> %a, i32 1
%select = select i1 %cond, i32 %3, i32 %4
Author: FarhanaAleen
Reviewed By: ABataev, RKSimon, spatel
Differential Revision: https://reviews.llvm.org/D47608
llvm-svn: 336130
This extends D48485 to allow another pair of binops (add/or) to be combined either
with or without a leading shuffle:
or X, C --> add X, C (when X and C have no common bits set)
Here, we need value tracking to determine that the 'or' can be reversed into an 'add',
and we've added general infrastructure to allow extending to other opcodes or moving
to where other passes could use that functionality.
Differential Revision: https://reviews.llvm.org/D48662
llvm-svn: 336128
Due to current limitations in constant analysis, we need flags
on add or mul to show propagation for the potential transform
suggested in these tests (no other binops currently report
identity constants).
llvm-svn: 336101
This version contains a fix to add values for which the state in ParamState change
to the worklist if the state in ValueState did not change. To avoid adding the
same value multiple times, mergeInValue returns true, if it added the value to
the worklist. The value is added to the worklist depending on its state in
ValueState.
Original message:
For comparisons with parameters, we can use the ParamState lattice
elements which also provide constant range information. This improves
the code for PR33253 further and gets us closer to use
ValueLatticeElement for all values.
Also, as we are using the range information in the solver directly, we
do not need tryToReplaceWithConstantRange afterwards anymore.
Reviewers: dberlin, mssimpso, davide, efriedma
Reviewed By: mssimpso
Differential Revision: https://reviews.llvm.org/D43762
llvm-svn: 336098
We were always using the opcodes of the first 2 scalars for the costs of the alternate opcode + shuffle. This made sense when we used SK_Alternate and opcodes were guaranteed to be alternating, but this fails for the more general SK_Select case.
This fix exposes an issue demonstrated by the fmul_fdiv_v4f32_const test - the SLM model has v4f32 fdiv costs which are more than twice those of the f32 scalar cost, meaning that the cost model determines that the vectorization is not performant. Unfortunately it completely ignores the fact that the fdiv by a constant will be changed into a fmul by InstCombine for a much lower cost vectorization. But at least we're seeing this now...
llvm-svn: 336095
Summary:
This patch introduce new intrinsic -
strip.invariant.group that was described in the
RFC: Devirtualization v2
Reviewers: rsmith, hfinkel, nlopes, sanjoy, amharc, kuhar
Subscribers: arsenm, nhaehnle, JDevlieghere, hiraditya, xbolva00, llvm-commits
Differential Revision: https://reviews.llvm.org/D47103
Co-authored-by: Krzysztof Pszeniczny <krzysztof.pszeniczny@gmail.com>
llvm-svn: 336073
This is a simple implementation of the unroll-and-jam classical loop
optimisation.
The basic idea is that we take an outer loop of the form:
for i..
ForeBlocks(i)
for j..
SubLoopBlocks(i, j)
AftBlocks(i)
Instead of doing normal inner or outer unrolling, we unroll as follows:
for i... i+=2
ForeBlocks(i)
ForeBlocks(i+1)
for j..
SubLoopBlocks(i, j)
SubLoopBlocks(i+1, j)
AftBlocks(i)
AftBlocks(i+1)
Remainder Loop
So we have unrolled the outer loop, then jammed the two inner loops into
one. This can lead to a simpler inner loop if memory accesses can be shared
between the now jammed loops.
To do this we have to prove that this is all safe, both for the memory
accesses (using dependence analysis) and that ForeBlocks(i+1) can move before
AftBlocks(i) and SubLoopBlocks(i, j).
Differential Revision: https://reviews.llvm.org/D41953
llvm-svn: 336062
Alternate opcode handling only supports binary operators, these tests demonstrate missed opportunities to vectorize some sitofp/uitofp and fptosi/fptoui style casts as well as some (successful) float bits manipulations
llvm-svn: 336060
This was discussed in D48401 as another improvement for:
https://bugs.llvm.org/show_bug.cgi?id=37806
If we have 2 different variable values, then we shuffle (select) those lanes,
shuffle (select) the constants, and then perform the binop. This eliminates a binop.
The new shuffle uses the same shuffle mask as the existing shuffle, so there's no
danger of creating a difficult shuffle.
All of the earlier constraints still apply, but we also check for extra uses to
avoid creating more instructions than we'll remove.
Additionally, we're disallowing the fold for div/rem because that could expose a
UB hole.
Differential Revision: https://reviews.llvm.org/D48678
llvm-svn: 335974
Summary:
An alternative to D48597.
Fixes [[ https://bugs.llvm.org/show_bug.cgi?id=37936 | PR37936 ]].
The problem is as follows:
1. `indvars` marks `%dec` as `NUW`.
2. `loop-instsimplify` runs `instsimplify`, which constant-folds `%dec` to -1 (D47908)
3. `loop-reduce` tries to do some further modification, but crashes
with an type assertion in cast, because `%dec` is no longer an `Instruction`,
If the runline is split into two, i.e. you first run `-indvars -loop-instsimplify`,
store that into a file, and then run `-loop-reduce`, there is no crash.
So it looks like the problem is due to `-loop-instsimplify` not discarding SCEV.
But in this case we can just not crash if it's not an `Instruction`.
This is just a local fix, unlike D48597, so there may very well be other problems.
Reviewers: mkazantsev, uabelho, sanjoy, silviu.baranga, wmi
Reviewed By: mkazantsev
Subscribers: evstupac, javed.absar, spatel, llvm-commits
Differential Revision: https://reviews.llvm.org/D48599
llvm-svn: 335950
Summary:
The InlinerFunctionImportStats will collect and dump stats regarding how
many function inlined into the module were imported by ThinLTO.
Reviewers: wmi, dexonsmith
Subscribers: mehdi_amini, inglorion, llvm-commits, eraman
Differential Revision: https://reviews.llvm.org/D48729
llvm-svn: 335914
When rewriting an alloca partition copy the DL from the
old alloca over the the new one.
Differential Revision: https://reviews.llvm.org/D48640
llvm-svn: 335904
This is an enhancement to D48401 that was discussed in:
https://bugs.llvm.org/show_bug.cgi?id=37806
We can convert a shift-left-by-constant into a multiply (we canonicalize IR in the other
direction because that's generally better of course). This allows us to remove the shuffle
as we do in the regular opcodes-are-the-same cases.
This requires a small hack to make sure we don't introduce any extra poison:
https://rise4fun.com/Alive/ZGv
Other examples of opcodes where this would work are add+sub and fadd+fsub, but we already
canonicalize those subs into adds, so there's nothing to do for those cases AFAICT. There
are planned enhancements for opcode transforms such or -> add.
Note that there's a different fold needed if we've already managed to simplify away a binop
as seen in the test based on PR37806, but we manage to get that one case here because this
fold is positioned above the demanded elements fold currently.
Differential Revision: https://reviews.llvm.org/D48485
llvm-svn: 335888
SCCP does not change the CFG, so we can mark it as preserved.
Reviewers: dberlin, efriedma, davide
Reviewed By: davide
Differential Revision: https://reviews.llvm.org/D47149
llvm-svn: 335820
If a trunc has a user in a block which is not reachable from entry,
we can safely perform trunc elimination as if this user didn't exist.
llvm-svn: 335816
I think the intrinsics named 'avx512.mask.' should refer to the previous behavior of taking a mask argument in the intrinsic instead of using a 'select' or 'and' instruction in IR to accomplish the masking. This is more consistent with the goal that eventually we will have no intrinsics that have masking builtin. When we reach that goal, we should have no intrinsics named "avx512.mask".
llvm-svn: 335744
This patch adds a custom trunc store lowering for v4i8 vector types.
Since there is not v.4b register, the v4i8 is promoted to v4i16 (v.4h)
and default action for v4i8 is to extract each element and issue 4
byte stores.
A better strategy would be to extended the promoted v4i16 to v8i16
(with undef elements) and extract and store the word lane which
represents the v4i8 subvectores. The construction:
define void @foo(<4 x i16> %x, i8* nocapture %p) {
%0 = trunc <4 x i16> %x to <4 x i8>
%1 = bitcast i8* %p to <4 x i8>*
store <4 x i8> %0, <4 x i8>* %1, align 4, !tbaa !2
ret void
}
Can be optimized from:
umov w8, v0.h[3]
umov w9, v0.h[2]
umov w10, v0.h[1]
umov w11, v0.h[0]
strb w8, [x0, #3]
strb w9, [x0, #2]
strb w10, [x0, #1]
strb w11, [x0]
ret
To:
xtn v0.8b, v0.8h
str s0, [x0]
ret
The patch also adjust the memory cost for autovectorization, so the C
code:
void foo (const int *src, int width, unsigned char *dst)
{
for (int i = 0; i < width; i++)
*dst++ = *src++;
}
can be vectorized to:
.LBB0_4: // %vector.body
// =>This Inner Loop Header: Depth=1
ldr q0, [x0], #16
subs x12, x12, #4 // =4
xtn v0.4h, v0.4s
xtn v0.8b, v0.8h
st1 { v0.s }[0], [x2], #4
b.ne .LBB0_4
Instead of byte operations.
llvm-svn: 335735
This prevents InstCombine from creating mis-sized dbg.values when
replacing a sequence of casts with a simpler cast. For example, in:
(fptrunc (floor (fpext X))) -> (floorf X)
We no longer emit dbg.value(X) (with a 32-bit float operand) to describe
(fpext X) (which is a 64-bit float).
This was diagnosed by the debugify check added in r335682.
llvm-svn: 335696
It's not possible to get the fragment size of some dbg.values. Teach the
mis-sized dbg.value diagnostic to detect this scenario and bail out.
Tested with:
$ find test/Transforms -print -exec opt -debugify-each -instcombine {} \;
llvm-svn: 335695
Summary:
When recording uses we need to rewrite after cloning a loop we need to
check if the use is not dominated by the original def. The initial
assumption was that the cloned basic block will introduce a new path and
thus the original def will only dominate the use if they are in the same
BB, but as the reproducer from PR37745 shows it's not always the case.
This fixes PR37745.
Reviewers: haicheng, Ka-Ka
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D48111
llvm-svn: 335675
Not sure why this logic seems to be repeated in 2 different places,
one called by the other.
On AMDGPU addrspace(3) globals start allocating at 0, so these
checks will be incorrect (not that real code actually tries
to compare these addresses)
llvm-svn: 335649
I'm not sure why the code here is skipping calls since
TTI does try to do something for general calls, but it
at least should allow intrinsics.
Skip intrinsics that should not be omitted as calls, which
is by far the most common case on AMDGPU.
llvm-svn: 335645
Similar to other patches in this series:
https://reviews.llvm.org/rL335512https://reviews.llvm.org/rL335527https://reviews.llvm.org/rL335597https://reviews.llvm.org/rL335616
...this is filling a gap in analysis that is exposed by an unrelated select-of-constants transform.
I didn't see a way to unify the sext cases because each div/rem opcode results in a different fold.
Note that in this case, the backend might want to convert the select into math:
Name: sext urem
%e = sext i1 %x to i32
%r = urem i32 %y, %e
=>
%c = icmp eq i32 %y, -1
%z = zext i1 %c to i32
%r = add i32 %z, %y
llvm-svn: 335622
Since D46637 we are better at handling uniform/non-uniform constant Pow2 detection; this patch tweaks the SLP argument handling to support them.
As SLP works with arrays of values I don't think we can easily use the pattern match helpers here.
Differential Revision: https://reviews.llvm.org/D48214
llvm-svn: 335621
Note: I didn't add a hasOneUse() check because the existing,
related fold doesn't have that check. I suspect that the
improved analysis and codegen make these some of the rare
canonicalization cases where we allow an increase in
instructions.
llvm-svn: 335597
changeToUnreachable may remove PHI nodes from executable blocks we found values
for and we would fail to replace them. By changing dead blocks to unreachable after
we replaced constants in all executable blocks, we ensure such PHI nodes are replaced
by their known value before.
Fixes PR37780.
Reviewers: efriedma, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D48421
llvm-svn: 335588
Summary:
This is a follow-up to r334830 and r335031.
In the valueCoversEntireFragment check we now also handle
the situation when there is a variable length array (VLA)
involved, and the length of the array has been reduced to
a constant.
The ConvertDebugDeclareToDebugValue functions that are related
to PHI nodes and load instructions now avoid inserting dbg.value
intrinsics when the value does not, for certain, cover the
variable/fragment that should be described.
In r334830 we assumed that the value always covered the entire
var/fragment and we had assertions in the code to show that
assumption. However, those asserts failed when compiling code
with VLAs, so we removed the asserts in r335031. Now when we
know that the valueCoversEntireFragment check can fail also for
PHI/Load instructions we avoid to insert the faulty dbg.value
intrinsic in such situations. Compared to the Store instruction
scenario we simply drop the dbg.value here (as the variable does
not change its value due to PHI/Load, so an earlier dbg.value
describing the variable should still be valid).
Reviewers: aprantl, vsk, efriedma
Reviewed By: aprantl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D48547
llvm-svn: 335580
Turn canonicalized subtraction back into (-1 - B) and combine it with (A + 1) into (A - B).
This is similar to the folding already done for (B ^ -1) + Const into (-1 + Const) - B.
Differential Revision: https://reviews.llvm.org/D48535
llvm-svn: 335579
unswitching of switches.
This works much like trivial unswitching of switches in that it reliably
moves the switch out of the loop. Here we potentially clone the entire
loop into each successor of the switch and re-point the cases at these
clones.
Due to the complexity of actually doing nontrivial unswitching, this
patch doesn't create a dedicated routine for handling switches -- it
would duplicate far too much code. Instead, it generalizes the existing
routine to handle both branches and switches as it largely reduces to
looping in a few places instead of doing something once. This actually
improves the results in some cases with branches due to being much more
careful about how dead regions of code are managed. With branches,
because exactly one clone is created and there are exactly two edges
considered, somewhat sloppy handling of the dead regions of code was
sufficient in most cases. But with switches, there are much more
complicated patterns of dead code and so I've had to move to a more
robust model generally. We still do as much pruning of the dead code
early as possible because that allows us to avoid even cloning the code.
This also surfaced another problem with nontrivial unswitching before
which is that we weren't as precise in reconstructing loops as we could
have been. This seems to have been mostly harmless, but resulted in
pointless LCSSA PHI nodes and other unnecessary cruft. With switches, we
have to get this *right*, and everything benefits from it.
While the testing may seem a bit light here because we only have two
real cases with actual switches, they do a surprisingly good job of
exercising numerous edge cases. Also, because we share the logic with
branches, most of the changes in this patch are reasonably well covered
by existing tests.
The new unswitch now has all of the same fundamental power as the old
one with the exception of the single unsound case of *partial* switch
unswitching -- that really is just loop specialization and not
unswitching at all. It doesn't fit into the canonicalization model in
any way. We can add a loop specialization pass that runs late based on
profile data if important test cases ever come up here.
Differential Revision: https://reviews.llvm.org/D47683
llvm-svn: 335553
This avoids creating unnecessary casts if the IP used to be a dbg info
intrinsic. Fixes PR37727.
Reviewers: vsk, aprantl, sanjoy, efriedma
Reviewed By: vsk, efriedma
Differential Revision: https://reviews.llvm.org/D47874
llvm-svn: 335513
We canonicalize to select with a zext-add and either zext-sub or sext-sub,
so this shows a pattern that's not conforming to the general trend.
llvm-svn: 335506
FDiv is replaced with multiplication by reciprocal and invariant
reciprocal is hoisted out of the loop, while multiplication remains
even if invariant.
Switch checks for all invariant operands and only invariant
denominator to fix the issue.
Differential Revision: https://reviews.llvm.org/D48447
llvm-svn: 335411
This gets rid of a bunch of weird special cases; instead, just use SCEV
rewriting for everything. In addition to being simpler, this fixes a
bug where we would use the wrong stride in certain edge cases.
The one bit I'm not quite sure about is the trip count handling,
specifically the FIXME about overflow. In general, I think we need to
widen the exit condition, but that's probably not profitable if the new
type isn't legal, so we probably need a check somewhere. That said, I
don't think I'm making the existing problem any worse.
As a followup to this, a bunch of IV-related code in root-finding could
be cleaned up; with SCEV-based rewriting, there isn't any reason to
assume a loop will have exactly one or two PHI nodes.
Differential Revision: https://reviews.llvm.org/D45191
llvm-svn: 335400
Enable tryToVectorizeList to support InstructionsState alternate opcode patterns at a root (build vector etc.) as well as further down the vectorization tree.
NOTE: This patch reduces some of the debug reporting if there are opcode mismatches - I can try to add it back if it proves a problem. But it could get rather messy trying to provide equivalent verbose debug strings via getSameOpcode etc.
Differential Revision: https://reviews.llvm.org/D48488
llvm-svn: 335364
SLP currently only accepts (F)Add/(F)Sub alternate counterpart ops to be merged into an alternate shuffle.
This patch relaxes this to accept any pair of BinaryOperator opcodes instead, assuming the target's cost model accepts the vectorization+shuffle.
Differential Revision: https://reviews.llvm.org/D48477
llvm-svn: 335349
This one shows another pattern that we'll need to match
in some cases, but the current ordering of folds allows
us to match this as 2 binops before simplification takes
place.
llvm-svn: 335347
AArch64 was only setting costs for SK_Transpose, which meant that many of the simpler shuffles (e.g. SK_Select and SK_PermuteSingleSrc for larger vector elements) was being severely overestimated by the default shuffle expansion.
This patch adds costs to help improve SLP performance and avoid a regression in reductions introduced by D48174.
I'm not very knowledgeable about AArch64 shuffle lowering so I've kept the extra costs to a minimum - someone who knows this code can add extra costs which should improve vectorization a lot more.
Differential Revision: https://reviews.llvm.org/D48172
llvm-svn: 335329
clear out deleted loops from the current queue beyond just the current
loop.
This is important because SimpleLoopUnswitch will now enqueue the same
loop to be re-processed. When it does this with the legacy PM, we don't
have a way of canceling the rest of the pipeline and so we can end up
deleting the loop before we reprocess it. =/
This change also makes it easy to support deleting other loops in the
queue to process, although I don't have any use cases for that.
Differential Revision: https://reviews.llvm.org/D48470
llvm-svn: 335317
With non-commutative binops, we could be using the same
variable value as operand 0 in 1 binop and operand 1 in
the other, so we have to check for that possibility and
bail out.
llvm-svn: 335312
Summary:
A reprise of D25849.
This crash was found through fuzzing some time ago and was documented in PR28879.
No check for load size has been added due to the following tests:
- Transforms/GVN/invariant.group.ll
- Transforms/GVN/pr10820.ll
These tests expect load sizes that are not a multiple of eight.
Thanks to @davide for the original patch.
Reviewers: nlopes, davide, RKSimon, reames, efriedma
Reviewed By: efriedma
Subscribers: davide, llvm-commits, Prazek
Differential Revision: https://reviews.llvm.org/D48330
llvm-svn: 335294
This is the simplest case from PR37806:
https://bugs.llvm.org/show_bug.cgi?id=37806
If we have a common variable operand used in a pair of binops with vector constants
that are vector selected together, then we can constant shuffle the constant vectors
to eliminate the shuffle instruction.
This has some tricky parts that are hopefully addressed in the tests and their
respective comments:
1. If the shuffle mask contains an undef element, then that lane of the result is
undef:
http://llvm.org/docs/LangRef.html#shufflevector-instruction
Therefore, we can replace the constant in that lane with an undef value except
for div/rem. With div/rem, an undef in the divisor would cause the whole op to
be undef. So I'm using the same hack as in D47686 - replace the undefs with '1'.
2. Intersect the wrapping and FMF of the original binops for the new binop. There
should be no extra poison or fast-math potential in the new binop that wasn't
possible in the original code.
3. Disregard other uses. Given that we're eliminating uses (shortening the
dependency chain), I think that's always the right IR canonicalization. But
I purposely chose the udiv test to demonstrate the scenario where both
intermediate values have other uses because that seems likely worse for
codegen with an expensive math op. This seems like a very rare possibility to
me, so I don't think it requires a backend patch first.
Differential Revision: https://reviews.llvm.org/D48401
llvm-svn: 335283
Roman Lebedev