Computing this property within the existing walk ensures that the cost is linear with the size of the block. If we did this from within isGuaranteedToExecute, it would be quadratic without some very fancy caching.
This allows us to reliably catch a hoistable instruction within a header which may throw at some point *after* our hoistable instruction. It doesn't do anything for non-header cases, but given how common single block loops are, this seems very worthwhile.
llvm-svn: 331557
A catchswitch is both a pad and a terminator, meaning it must be the
only non-phi instruction in its basic block. When we're inserting a
bitcast in the incoming basic block for a phi, if that incoming block is
a catchswitch, we should go up the dominator tree to find a valid
insertion point rather than attempting to insert before the catchswitch
(which would result in invalid IR).
Differential Revision: https://reviews.llvm.org/D46412
llvm-svn: 331548
We currently recognize this idiom where x is signed and thus the shift in an ashr.
int cnt = 0;
while (x) {
x >>= 1; // arithmetic shift right
++cnt;
}
and turn it into (bitwidth - ctlz(x)). And if there is anything else in the loop we will create a new loop that runs that many times.
If x is initially negative, the shift result will never be 0 and thus the loop is infinite. If you put something with side effects in the loop, that side effect will now only happen bitwidth times instead of an infinite number of times.
So this transform is only safe for logical shift right (which we don't currently recognize) or if we can prove that x cannot be negative before the loop.
llvm-svn: 331493
Add logic for the special case when a cmp+select can clearly be
reduced to just a bitwise logic instruction, and remove an
over-reaching chunk of general purpose bit magic. The primary goal
is to remove cases where we are not improving the IR instruction
count when doing these select transforms, and in all cases here that
is true.
In the motivating 3-way compare tests, there are further improvements
because we can combine/propagate select values (not sure if that
belongs in instcombine, but it's there for now).
DAGCombiner has folds to turn some of these selects into bit magic,
so there should be no difference in the end result in those cases.
Not all constant combinations are handled there yet, however, so it
is possible that some targets will see more cmov/csel codegen with
this change in IR canonicalization.
Ideally, we'll go further to *not* turn selects into multiple
logic/math ops in instcombine, and we'll canonicalize to selects.
But we should make sure that this step does not result in regressions
first (and if it does, we should fix those in the backend).
The general direction for this change was discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2016-September/105373.htmlhttp://lists.llvm.org/pipermail/llvm-dev/2017-July/114885.html
Alive proofs for the new bit magic:
https://rise4fun.com/Alive/XG7
Differential Revision: https://reviews.llvm.org/D46086
llvm-svn: 331486
The code fails to check that the same value is used twice. We only make sure the left hand side of the and is part of the loop recurrence. The 'x' in the subtract can be any value.
llvm-svn: 331436
This patch was temporarily reverted because it has exposed bug 37229 on
PowerPC platform. The bug is unrelated to the patch and was just a general
bug in the optimization done for PowerPC platform only. The bug was fixed
by the patch rL331410.
This patch returns the disabled commit since the bug was fixed.
llvm-svn: 331427
Summary:
Prior to this change, LLVM would in some cases emit *massive* writeout
functions with many 10s of 1000s of function calls in straight-line
code. This is a very wasteful way to represent what are fundamentally
loops and creates a number of scalability issues. Among other things,
register allocating these calls is extremely expensive. While D46127 makes this
less severe, we'll still run into scaling issues with this eventually. If not
in the compile time, just from the code size.
Now the pass builds up global data structures modeling the inputs to
these functions, and simply loops over the data structures calling the
relevant functions with those values. This ensures that the code size is
a fixed and only data size grows with larger amounts of coverage data.
A trivial change to IRBuilder is included to make it easier to build
the constants that make up the global data.
Reviewers: wmi, echristo
Subscribers: sanjoy, mcrosier, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D46357
llvm-svn: 331407
Summary:
Some of our internal testing detected a major compile time regression which I've
tracked down to:
r278938 - Revert "Reassociate: Reprocess RedoInsts after each inst".
It appears that processing long chains of reassociatable instructions causes
non-linear (potentially exponential) growth in the number of times an
instruction is revisited. For example, the included test revisits instructions
220 times in a 20-instruction test.
It appears that r278938 reversed the order instructions were visited and that
this is preventing scheduled revisits from being cancelled as a result of
visiting the instructions naturally during normal processing. However, simply
reversing the order also harmed the generated code. Upon closer inspection, it
was discovered that revisits occurred in the opposite order to the first pass
(Thanks to escha for spotting that).
This patch makes the revisit order consistent with the first pass which allows
more revisits to be cancelled. This does appear to have a small impact on the
generated code in few cases but it significantly reduces compile-time.
After this patch, our internal test that was most affected by the regression
dropped from ~2 million revisits to ~4k resulting in Reassociate having 0.46%
of the runtime it had before (99.54% improvement).
Here's the summaries reported by lnt for the LLVM test-suite with --benchmarking-only:
| metric | geomean before patch | geomean after patch | delta |
| ----- | ----- | ----- | ----- |
| compile time | 0.1956 | 0.1261 | -35.54% |
| execution time | 0.3240 | 0.3237 | - |
| code size | 7365.4459 | 7365.6079 | - |
The results have a few wins and losses on compile-time, mostly in the +/- 2.5% range. There was one outlier though:
| Performance Regressions - compile_time | Δ | Previous | Current |
| MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/CrystalMk | 9.82% | 2.0473 | 2.2483 |
Reviewers: javed.absar, dberlin
Reviewed By: dberlin
Subscribers: kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D45734
llvm-svn: 331381
and (or (lshr X, C), ...), 1 --> (X & C') != 0
I initially thought about implementing the minimal pattern in instcombine as mentioned here:
https://bugs.llvm.org/show_bug.cgi?id=37098#c6
...but we need to do better to catch the more general sequence from the motivating test
(more than 2 bits in the compare). And a test-suite run with statistics showed that this
pattern only happened 2 times currently. It would potentially happen more often if
reassociation worked better (D45842), but it's probably still not too frequent?
This is small enough that I didn't see a need to create a whole new class/file within
AggressiveInstCombine. There are likely other relatively small matchers like what was
discussed in D44266 that would slide under foldUnusualPatterns() (name suggestions welcome).
We could potentially also consolidate matchers for ctpop, bswap, etc under here.
Differential Revision: https://reviews.llvm.org/D45986
llvm-svn: 331311
As mentioned in D45986, there's a potential ordering dependency
between instcombine and aggressive-instcombine for detecting these,
so I'm adding a few tests to confirm that the expected folds occur
using -O3 (because aggressive-instcombine only runs at -O3 currently).
llvm-svn: 331308
Summary:
This fixes a build break with r331269.
test/Transforms/LoopVectorize/pr23997.ll
should be in:
test/Transforms/LoopVectorize/X86/pr23997.ll
llvm-svn: 331281
Summary:
This is a fix for PR23997.
The loop vectorizer is not preserving the inbounds property of GEPs that it creates.
This is inhibiting some optimizations. This patch preserves the inbounds property in
the case where a load/store is being fed by an inbounds GEP.
Reviewers: mkuper, javed.absar, hsaito
Reviewed By: hsaito
Subscribers: dcaballe, hsaito, llvm-commits
Differential Revision: https://reviews.llvm.org/D46191
llvm-svn: 331269
phi is on lhs of a comparison op.
For the following testcase,
L1:
%t0 = add i32 %m, 7
%t3 = icmp eq i32* %t2, null
br i1 %t3, label %L3, label %L2
L2:
%t4 = load i32, i32* %t2, align 4
br label %L3
L3:
%t5 = phi i32 [ %t0, %L1 ], [ %t4, %L2 ]
%t6 = icmp eq i32 %t0, %t5
br i1 %t6, label %L4, label %L5
We know if we go through the path L1 --> L3, %t6 should always be true. However
currently, if the rhs of the eq comparison is phi, JumpThreading fails to
evaluate %t6 to true. And we know that Instcombine cannot guarantee always
canonicalizing phi to the left hand side of the comparison operation according
to the operand priority comparison mechanism in instcombine. The patch handles
the case when rhs of the comparison op is a phi.
Differential Revision: https://reviews.llvm.org/D46275
llvm-svn: 331266
instcombine should transform the relevant cases if the OverflowingBinaryOperator/PossiblyExactOperator can be proven to be safe.
Change-Id: I7aec62a31a894e465e00eb06aed80c3ea0c9dd45
llvm-svn: 331265
This test had values that differed in only in capitalization,
and that causes problems for the auto-generating check line
script. So I changed that in rL331226, but I accidentally
forgot to change a subsequent use of a param.
llvm-svn: 331228
Summary:
As discussed in D45733, we want to do this in InstCombine.
https://rise4fun.com/Alive/LGk
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: chandlerc, xbolva00, llvm-commits
Differential Revision: https://reviews.llvm.org/D45867
llvm-svn: 331205
Summary:
Masked merge has a pattern of: `((x ^ y) & M) ^ y`.
But, there is no difference between `((x ^ y) & M) ^ y` and `((x ^ y) & ~M) ^ x`,
We should canonicalize the pattern to non-inverted mask.
https://rise4fun.com/Alive/Yol
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45664
llvm-svn: 331112
Summary:
Masked merge has a pattern of: `((x ^ y) & M) ^ y`.
But, there is no difference between `((x ^ y) & M) ^ y` and `((x ^ y) & ~M) ^ x`,
We should canonicalize the pattern to non-inverted mask.
Differential Revision: https://reviews.llvm.org/D45663
llvm-svn: 331111
The effect of doing so is not disrupting the LoopPassManager when mixing this pass with other loop passes. This should help locality of access substaintially and avoids the cost of computing PostDom.
The assumption here is that the full GuardWidening (which does use PostDom) is run as a canonicalization before loop opts and that this version is just catching cases exposed by other loop passes. (i.e. LoopPredication, IndVarSimplify, LoopUnswitch, etc..)
llvm-svn: 331094
This patch adds support for fragment expressions
TryToShrinkGlobalToBoolean() which were previously just dropped.
Thanks to Reid Kleckner for providing me a reproducer!
llvm-svn: 331086
Summary:
Currently, we
1. match `LHS` matcher to the `first` operand of binary operator,
2. and then match `RHS` matcher to the `second` operand of binary operator.
If that does not match, we swap the `LHS` and `RHS` matchers:
1. match `RHS` matcher to the `first` operand of binary operator,
2. and then match `LHS` matcher to the `second` operand of binary operator.
This works ok.
But it complicates writing of commutative matchers, where one would like to match
(`m_Value()`) the value on one side, and use (`m_Specific()`) it on the other side.
This is additionally complicated by the fact that `m_Specific()` stores the `Value *`,
not `Value **`, so it won't work at all out of the box.
The last problem is trivially solved by adding a new `m_c_Specific()` that stores the
`Value **`, not `Value *`. I'm choosing to add a new matcher, not change the existing
one because i guess all the current users are ok with existing behavior,
and this additional pointer indirection may have performance drawbacks.
Also, i'm storing pointer, not reference, because for some mysterious-to-me reason
it did not work with the reference.
The first one appears trivial, too.
Currently, we
1. match `LHS` matcher to the `first` operand of binary operator,
2. and then match `RHS` matcher to the `second` operand of binary operator.
If that does not match, we swap the ~~`LHS` and `RHS` matchers~~ **operands**:
1. match ~~`RHS`~~ **`LHS`** matcher to the ~~`first`~~ **`second`** operand of binary operator,
2. and then match ~~`LHS`~~ **`RHS`** matcher to the ~~`second`~ **`first`** operand of binary operator.
Surprisingly, `$ ninja check-llvm` still passes with this.
But i expect the bots will disagree..
The motivational unittest is included.
I'd like to use this in D45664.
Reviewers: spatel, craig.topper, arsenm, RKSimon
Reviewed By: craig.topper
Subscribers: xbolva00, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D45828
llvm-svn: 331085
As suggested in D45842
(although still not sure if we're going to advance that),
we must invalidate references to instructions that have
been recycled (operands were changed, so result is different).
llvm-svn: 331083
We currently have a hard to solve analysis problem around the order of instructions within a potentially throwing block. We can't cheaply determine whether a given instruction is before the first potential throw in the block. While we're working on that in the background, special case the first instruction within the header.
why this particular special case? Well, headers are guaranteed to execute if the loop does, and it turns out we tend to produce this form in practice.
In a follow on patch, I tend to extend LICM with an alternate approach which works for any instruction in the header before the first throw, but this is the best I can come up with other users of the analysis (such as store promotion.)
Note: I can't show the difference in the analysis result since we're ORing in the expensive instruction walk used by SCEV. Using the full walk is not suitable for a general solution.
llvm-svn: 331079
The idea is to have a pass which performs the same transformation as GuardWidening, but can be run within a loop pass manager without disrupting the pass manager structure. As demonstrated by the test case, this doesn't quite get there because of issues with post dom, but it gives a good step in the right direction. the motivation is purely to reduce compile time since we can now preserve locality during the loop walk.
This patch only includes a legacy pass. A follow up will add a new style pass as well.
llvm-svn: 331060
We currently support LCSSA PHI nodes in the outer loop exit, if their
incoming values do not come from the outer loop latch or if the
outer loop latch has a single predecessor. In that case, the outer loop latch
will be executed only if the inner loop gets executed. If we have multiple
predecessors for the outer loop latch, it may be executed even if the inner
loop does not get executed.
This is a first step to support the case described in
https://bugs.llvm.org/show_bug.cgi?id=30472
Reviewers: efriedma, karthikthecool, mcrosier
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D43237
llvm-svn: 331037
Since PTX has grown a <2 x half> datatype vectorization has become more
important. The late LoadStoreVectorizer intentionally only does loads
and stores, but now arithmetic has to be vectorized for optimal
throughput too.
This is still very limited, SLP vectorization happily creates <2 x half>
if it's a legal type but there's still a lot of register moving
happening to get that fed into a vectorized store. Overall it's a small
performance win by reducing the amount of arithmetic instructions.
I haven't really checked what the loop vectorizer does to PTX code, the
cost model there might need some more tweaks. I didn't see it causing
harm though.
Differential Revision: https://reviews.llvm.org/D46130
llvm-svn: 331035
It doesn't unwind, and the wrong marking leads to the creation of an
.eh_frame section when it isn't necessary.
Differential Revision: https://reviews.llvm.org/D46082
llvm-svn: 331008
Summary: If file stream arg is not captured and source is fopen, we could replace IO calls by unlocked IO ("_unlocked" function variants) to gain better speed,
Reviewers: efriedma, RKSimon, spatel, sanjoy, hfinkel, majnemer
Subscribers: lebedev.ri, llvm-commits
Differential Revision: https://reviews.llvm.org/D45736
llvm-svn: 331002
Summary:
Simplify integer add expression X % C0 + (( X / C0 ) % C1) * C0 to
X % (C0 * C1). This is a common pattern seen in code generated by the XLA
GPU backend.
Add test cases for this new optimization.
Patch by Bixia Zheng!
Reviewers: sanjoy
Reviewed By: sanjoy
Subscribers: efriedma, craig.topper, lebedev.ri, llvm-commits, jlebar
Differential Revision: https://reviews.llvm.org/D45976
llvm-svn: 330992
remainder expressions as operands.
Summary:
Add test cases to prepare for the new optimization that Simplifies integer add
expression X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1).
Patch by Bixia Zheng!
Reviewers: sanjoy
Reviewed By: sanjoy
Subscribers: jlebar, llvm-commits
Differential Revision: https://reviews.llvm.org/D46017
llvm-svn: 330991
Philip Reames