This patch updates the code that deals with conditions from predicate
info to make use of constant ranges.
For ssa_copy instructions inserted by PredicateInfo, we have 2 ranges:
1. The range of the original value.
2. The range imposed by the linked condition.
1. is known, 2. can be determined using makeAllowedICmpRegion. The
intersection of those ranges is the range for the copy.
With this patch, we get a nice increase in the number of instructions
eliminated by both SCCP and IPSCCP for some benchmarks:
For MultiSource, SPEC2000 & SPEC2006:
Tests: 237
Same hash: 170 (filtered out)
Remaining: 67
Metric: sccp.NumInstRemoved
Program base patch diff
test-suite...Source/Benchmarks/sim/sim.test 10.00 71.00 610.0%
test-suite...CFP2000/177.mesa/177.mesa.test 361.00 1626.00 350.4%
test-suite...encode/alacconvert-encode.test 141.00 602.00 327.0%
test-suite...decode/alacconvert-decode.test 141.00 602.00 327.0%
test-suite...CI_Purple/SMG2000/smg2000.test 1639.00 4093.00 149.7%
test-suite...peg2/mpeg2dec/mpeg2decode.test 75.00 163.00 117.3%
test-suite...T2006/401.bzip2/401.bzip2.test 358.00 513.00 43.3%
test-suite...rks/FreeBench/pifft/pifft.test 11.00 15.00 36.4%
test-suite...langs-C/unix-tbl/unix-tbl.test 4.00 5.00 25.0%
test-suite...lications/sqlite3/sqlite3.test 541.00 667.00 23.3%
test-suite.../CINT2000/254.gap/254.gap.test 243.00 299.00 23.0%
test-suite...ks/Prolangs-C/agrep/agrep.test 25.00 29.00 16.0%
test-suite...marks/7zip/7zip-benchmark.test 1135.00 1304.00 14.9%
test-suite...lications/ClamAV/clamscan.test 1105.00 1268.00 14.8%
test-suite...urce/Applications/lua/lua.test 398.00 436.00 9.5%
Metric: sccp.IPNumInstRemoved
Program base patch diff
test-suite...C/CFP2000/179.art/179.art.test 1.00 3.00 200.0%
test-suite...006/447.dealII/447.dealII.test 429.00 1056.00 146.2%
test-suite...nch/fourinarow/fourinarow.test 3.00 7.00 133.3%
test-suite...CI_Purple/SMG2000/smg2000.test 818.00 1748.00 113.7%
test-suite...ks/McCat/04-bisect/bisect.test 3.00 5.00 66.7%
test-suite...CFP2000/177.mesa/177.mesa.test 165.00 255.00 54.5%
test-suite...ediabench/gsm/toast/toast.test 18.00 27.00 50.0%
test-suite...telecomm-gsm/telecomm-gsm.test 18.00 27.00 50.0%
test-suite...ks/Prolangs-C/agrep/agrep.test 24.00 35.00 45.8%
test-suite...TimberWolfMC/timberwolfmc.test 43.00 62.00 44.2%
test-suite...encode/alacconvert-encode.test 46.00 66.00 43.5%
test-suite...decode/alacconvert-decode.test 46.00 66.00 43.5%
test-suite...langs-C/unix-tbl/unix-tbl.test 12.00 17.00 41.7%
test-suite...peg2/mpeg2dec/mpeg2decode.test 31.00 41.00 32.3%
test-suite.../CINT2000/254.gap/254.gap.test 117.00 154.00 31.6%
Reviewers: efriedma, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D76611
Now that we have scalable vectors, there's a distinction that isn't
getting captured in the original SequentialType: some vectors don't have
a known element count, so counting the number of elements doesn't make
sense.
In some cases, there's a better way to express the commonality using
other methods. If we're dealing with GEPs, there's GEP methods; if we're
dealing with a ConstantDataSequential, we can query its element type
directly.
In the relatively few remaining cases, I just decided to write out
the type checks. We're talking about relatively few places, and I think
the abstraction doesn't really carry its weight. (See thread "[RFC]
Refactor class hierarchy of VectorType in the IR" on llvmdev.)
Differential Revision: https://reviews.llvm.org/D75661
Summary:
In some cases, ASan may insert instrumentation before function arguments
have been stored into their allocas. This causes two issues:
1) The argument value must be spilled until it can be stored into the
reserved alloca, wasting a stack slot.
2) Until the store occurs in a later basic block, the debug location
will point to the wrong frame offset, and backtraces will show an
uninitialized value.
The proposed solution is to move instructions which initialize allocas
for arguments up into the entry block, before the position where ASan
starts inserting its instrumentation.
For the motivating test case, before the patch we see:
```
| 0033: movq %rdi, 0x68(%rbx) | | DW_TAG_formal_parameter |
| ... | | DW_AT_name ("a") |
| 00d1: movq 0x68(%rbx), %rsi | | DW_AT_location (RBX+0x90) |
| 00d5: movq %rsi, 0x90(%rbx) | | ^ not correct ... |
```
and after the patch we see:
```
| 002f: movq %rdi, 0x70(%rbx) | | DW_TAG_formal_parameter |
| | | DW_AT_name ("a") |
| | | DW_AT_location (RBX+0x70) |
```
rdar://61122691
Reviewers: aprantl, eugenis
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77182
Summary:
It can be helpful to test behaviour w.r.t locations without having DEBUG_VALUE
around. In particular, because DEBUG_VALUE has the potential to change CodeGen
behaviour (e.g. hasOneUse() vs hasOneNonDbgUse()) while locations generally
don't.
Reviewers: aprantl, bogner
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77438
The patch introduces the system to distinctively store the information
needed for the Control Flow Graph as well as the instrumentary needed for
the follow-up changes: BlockFrequencyInfo and BranchProbabilityInfo.
The patch is a part of sequence of three patches, related to graphs Heat Coloring.
Reviewers: rcorcs, apilipenko, davidxl, sfertile, fedor.sergeev, eraman, bollu
Differential Revision: https://reviews.llvm.org/D76820
This patch moves calls to their own recipe, to simplify the transition
to VPUser for operands of VPWidenRecipe, as discussed in D76992.
Subsequently additional information can be added to the recipe rather
than computing it during the execute step.
Reviewers: rengolin, Ayal, gilr, hsaito
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D77467
Summary:
In D77454 we explain that `LoadInst` and `StoreInst` always have their alignment defined.
This allows to work backward here and to infer that `getNewAlignment` does not need to return `0` in case of failure.
Returning `1` also works since it needs to be greater than the Load/Store alignment which is a least `1`.
This is patch is part of a series to introduce an Alignment type.
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2019-July/133851.html
See this patch for the introduction of the type: https://reviews.llvm.org/D64790
Reviewers: courbet
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77538
Summary:
This is a roll forward of D77394 minus AlignmentFromAssumptions (which needs to be addressed separately)
Differences from D77394:
- DebugStr() now prints the alignment value or `None` and no more `Align(x)` or `MaybeAlign(x)`
- This is to keep Warning message consistent (CodeGen/SystemZ/alloca-04.ll)
- Removed a few unneeded headers from Alignment (since it's included everywhere it's better to keep the dependencies to a minimum)
Reviewers: courbet
Subscribers: sdardis, hiraditya, jrtc27, atanasyan, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77537
This patch adds a -matrix-default-layout option which can be used to
set the default matrix layout to row-major or column-major (default).
The initial patch updates codegen for loads, stores, binary operators
and matrix multiply.
Reviewers: anemet, Gerolf, andrew.w.kaylor, LuoYuanke
Reviewed By: anemet
Differential Revision: https://reviews.llvm.org/D76325
This patch adds initial fusion for load/multiply/store chains of matrix
operations.
The patch contains roughly two parts:
1. Code generation for a fused load/multiply/store chain (LowerMatrixMultiplyFused).
First, we ensure that both loads of the multiply operands do not alias the store.
If they do, we create new non-aliasing copies of the operands. Note that this
may introduce new basic block. Finally we process TileSize x TileSize blocks.
That is: load tiles from the input operands, multiply and store them.
2. Identify fusion candidates & matrix instructions.
As a first step, collect all instructions with shape info and fusion candidates
(currently @llvm.matrix.multiply calls). Next, try to fuse candidates and
collect instructions eliminated by fusion. Finally iterate over all matrix
instructions, skip the ones eliminated by fusion and lower the rest as usual.
Reviewers: anemet, Gerolf, hfinkel, andrew.w.kaylor, LuoYuanke
Reviewed By: anemet
Differential Revision: https://reviews.llvm.org/D75566
The new and old pass managers (PassManagerBuilder.cpp and
PassBuilder.cpp) are exposed to an `extern` declaration of
`attributor-disable` option which will guard the addition of the
attributor passes to the pass pipelines.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D76871
This patch builds upon D76140 by updating metadata on pointer typed
loads in inlined functions, when the load is the return value, and the
callsite contains return attributes which can be updated as metadata on
the load.
Added test cases show this for nonnull, dereferenceable,
dereferenceable_or_null
Reviewed-By: jdoerfert
Differential Revision: https://reviews.llvm.org/D76792
As discussed in D76983, that patch can turn a chain of insert/extract
with scalar trunc ops into bitcast+extract and existing instcombine
vector transforms end up creating a shuffle out of that (see the
PhaseOrdering test for an example). Currently, that process requires
at least this sequence: -instcombine -early-cse -instcombine.
Before D76983, the sequence of insert/extract would reach the SLP
vectorizer and become a vector trunc there.
Based on a small sampling of public targets/types, converting the
shuffle to a trunc is better for codegen in most cases (and a
regression of that form is the reason this was noticed). The trunc is
clearly better for IR-level analysis as well.
This means that we can induce "spontaneous vectorization" without
invoking any explicit vectorizer passes (at least a vector cast op
may be created out of scalar casts), but that seems to be the right
choice given that we started with a chain of insert/extract, and the
backend would expand back to that chain if a target does not support
the op.
Differential Revision: https://reviews.llvm.org/D77299
Query AAValueSimplify on pointers in memory accessing instructions to take
advantage of the constant propagation (or any other value simplification) of such values.
After 49d00824bb, VPWidenRecipe only stores a single instruction.
tryToWiden can simply return the widen recipe, like other helpers in
VPRecipeBuilder.
eraseInstFromFunction() adds the operands of the erased instructions,
as those might now be dead as well. However, this is limited to
instructions with less than 8 operands.
This check doesn't make a lot of sense to me. As the instruction
gets removed afterwards, I don't see a potential for anything
overly pathological happening here (as we can only add those
operands to the worklist once). The impact on CTMark is in
the noise. We also have the same code in instruction sinking
and don't limit the operand count there.
Differential Revision: https://reviews.llvm.org/D77325
Extracting to the same index that we are going to insert back into
allows forming select ("blend") shuffles and enables further transforms.
Admittedly, this is a quick-fix for a more general problem that I'm
hoping to solve by adding transforms for patterns that start with an
insertelement.
But this might resolve some regressions known to be caused by the
extract-extract transform (although I have not gotten more details on
those yet).
In the motivating case from PR34724:
https://bugs.llvm.org/show_bug.cgi?id=34724
The combination of subsequent instcombine and codegen transforms gets us this improvement:
vmovshdup %xmm0, %xmm2 ## xmm2 = xmm0[1,1,3,3]
vhaddps %xmm1, %xmm1, %xmm4
vmovshdup %xmm1, %xmm3 ## xmm3 = xmm1[1,1,3,3]
vaddps %xmm0, %xmm2, %xmm0
vaddps %xmm1, %xmm3, %xmm1
vshufps $200, %xmm4, %xmm0, %xmm0 ## xmm0 = xmm0[0,2],xmm4[0,3]
vinsertps $177, %xmm1, %xmm0, %xmm0 ## xmm0 = zero,xmm0[1,2],xmm1[2]
-->
vmovshdup %xmm0, %xmm2 ## xmm2 = xmm0[1,1,3,3]
vhaddps %xmm1, %xmm1, %xmm1
vaddps %xmm0, %xmm2, %xmm0
vshufps $200, %xmm1, %xmm0, %xmm0 ## xmm0 = xmm0[0,2],xmm1[0,3]
Differential Revision: https://reviews.llvm.org/D76623
As discussed in post-commit review in https://reviews.llvm.org/D73501
if the goal of this is to help vectorizer, then we should actually
be teaching vectorizer to do this, because right now this rewrite
is still budget-limited, which isn't what we'd want.
Additionally, while the rest of the patch series was universally profitable,
this particular patch is reportedly (https://reviews.llvm.org/D73501#1905171)
exposing cost-modeling issues on ARM.
So let's just back this particular patch out. Once there's an undo transform,
this could be considered for reintegration.
This reverts commit 44edc6fd2c.
Summary:
A recent change in the instruction simplifier enables a call to a function that just returns one of its parameter to be simplified as simply loading the parameter. This exposes a bug in the inliner where double inlining may be involved which in turn may cause compiler ICE when an already-inlined callsite is reused for further inlining.
To put it simply, in the following-like C program, when the function call second(t) is inlined, its code t = third(t) will be reduced to just loading the return value of the callsite first(). This causes the inliner internal data structure to register the first() callsite for the call edge representing the third() call, therefore incurs a double inlining when both call edges are considered an inline candidate. I'm making a fix to break the inliner from reusing a callsite for new call edges.
```
void top()
{
int t = first();
second(t);
}
void second(int t)
{
t = third(t);
fourth(t);
}
void third(int t)
{
return t;
}
```
The actual failing case is much trickier than the example here and is only reproducible with the legacy inliner. The way the legacy inliner works is to process each SCC in a bottom-up order. That means in reality function first may be already inlined into top, or function third is either inlined to second or is folded into nothing. To repro the failure seen from building a large application, we need to figure out a way to confuse the inliner so that the bottom-up inlining is not fulfilled. I'm doing this by making the second call indirect so that the alias analyzer fails to figure out the right call graph edge from top to second and top can be processed before second during the bottom-up. We also need to tweak the test code so that when the inlining of top happens, the function body of second is not that optimized, by delaying the pass of function attribute deducer (i.e, which tells function third has no side effect and just returns its parameter). Since the CGSCC pass is iterative, additional calls are added to top to postpone the inlining of second to the second round right after the first function attribute deducing pass is done. I haven't been able to repro the failure with the new pass manager since the processing order of ininlined callsites is a bit different, but in theory the issue could happen there too.
Note that this fix could introduce a side effect that blocks the simplification of inlined code, specifically for a call site that can be folded to another call site. I hope this can probably be complemented by subsequent inlining or folding, as shown in the attached unit test. The ideal fix should be to separate the use of VMap. However, in reality this failing pattern shouldn't happen often. And even if it happens, there should be a good chance that the non-folded call site will be refolded by iterative inlining or subsequent simplification.
Reviewers: wenlei, davidxl, tejohnson
Reviewed By: wenlei, davidxl
Subscribers: eraman, nikic, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76248
Consider a callee function that has a call (C) within it which feeds
into the return. When we inline that callee into a callsite that has
return attributes, we can backward propagate valid attributes to the
call (C) within that inlined callee body.
This is safe to do so only if we can guarantee transfer of execution to
successor in the window of instructions between return value (i.e. the
call C) and the return instruction.
Also, this is valid only for attributes which are a property of a
callsite and not those that are not dependent on the ABI, or a property
of the call itself.
Reviewed-By: reames, jdoerfert
Differential Revision: https://reviews.llvm.org/D76140
bitcast (shuf V, MaskC) --> shuf (bitcast V), MaskC'
We do not attempt this in InstCombine because we do not want to change
types and create new shuffle ops that are potentially not lowered as
well as the original code. Here, we can check the cost model to see if
it is worthwhile.
I've aggressively enabled this transform even if the types are the same
size and/or equal cost because moving the bitcast allows InstCombine to
make further simplifications.
In the motivating cases from PR35454:
https://bugs.llvm.org/show_bug.cgi?id=35454
...this is enough to let instcombine and the backend eliminate the
redundant shuffles, but we probably want to extend VectorCombine to
handle the inverse pattern (shuffle-of-bitcast) to get that
simplification directly in IR.
Differential Revision: https://reviews.llvm.org/D76727
These are versions of a function that regressed with:
rGf2fbdf76d8d0
That particular problem occurs with an instcombine-simplifycfg-instcombine
sequence, but we can show that it exists within instcombine only with
other variations of the pattern.
This reverts commit f2fbdf76d8.
As noted in the post-commit thread:
https://reviews.llvm.org/rGf2fbdf76d8d0
...this can obscure a min/max pattern where the components
have extra uses. We can show that the problem is independent
of this change with a slightly modified source example, so
this revert just delays/reduces the need to fix the real
problem.
We need to improve our analysis of negation or -- more
generally -- subtraction using patches like D77230 or D68408.
Summary:
Splitting Knowledge retention into Queries in Analysis and Builder into Transform/Utils
allows Queries and Transform/Utils to use Analysis.
Reviewers: jdoerfert, sstefan1
Reviewed By: jdoerfert
Subscribers: mgorny, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77171
This patch adds
- New arguments to getMinPrefetchStride() to let the target decide on a
per-loop basis if software prefetching should be done even with a stride
within the limit of the hw prefetcher.
- New TTI hook enableWritePrefetching() to let a target do write prefetching
by default (defaults to false).
- In LoopDataPrefetch:
- A search through the whole loop to gather information before emitting any
prefetches. This way the target can get information via new arguments to
getMinPrefetchStride() and emit prefetches more selectively. Collected
information includes: Does the loop have a call, how many memory
accesses, how many of them are strided, how many prefetches will cover
them. This is NFC to before as long as the target does not change its
definition of getMinPrefetchStride().
- If a previous access to the same exact address was 'read', and the
current one is 'write', make it a 'write' prefetch.
- If two accesses that are covered by the same prefetch do not dominate
each other, put the prefetch in a block that dominates both of them.
- If a ConstantMaxTripCount is less than ItersAhead, then skip the loop.
- A SystemZ implementation of getMinPrefetchStride().
Review: Ulrich Weigand, Michael Kruse
Differential Revision: https://reviews.llvm.org/D70228
As pointed out by @thakis, currently CallSiteSplitting bails out after
checking the first PHI node. We should check all PHI nodes, until we
find one where call site splitting is beneficial.
This patch also slightly simplifies the code using BasicBlock::phis().
Reviewers: davidxl, junbuml, thakis
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D77089
There was a TODO in genericValueTraversal to provide the context
instruction and due to the lack of it users that wanted one just used
something available. Unfortunately, using a fixed instruction is wrong
in the presence of PHIs so we need to update the context instruction
properly.
Reviewed By: uenoku
Differential Revision: https://reviews.llvm.org/D76870
This cannot be triggered right now, as far as I know, but it doesn't
make sense to deduce a constant range on arguments of declarations.
Exposed during testing of AAValueSimplify extensions.
Use DL & ABI information for better alignment deduction, e.g., if a type
is accessed and the ABI specifies an alignment requirement for such an
access we can use it. This is based on a patch by @lebedev.ri and
inspired by getBaseAlign in Loads.cpp.
Depends on D76673.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D76674
If we have a must-tail call the callee and caller need to have matching
ABIs. Part of that is alignment which we might modify when we deduce
alignment of arguments of either. Since we would need to keep them in
sync, which is not as simple, we simply avoid deducing alignment for
arguments of the must-tail caller or callee.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D76673
We create a lot of AbstractAttributes and they live as long as
the Attributor does. It seems reasonable to allocate them via a
BumpPtrAllocator owned by the Attributor.
Reviewed By: baziotis
Differential Revision: https://reviews.llvm.org/D76589
Florian Hahn