InstCombine currently has odd rules for folding insert-extract chains to shuffles,
so we miss collapsing seemingly simple cases as shown in the tests here.
But poison makes this not quite as easy as we might have guessed. Alive2 tests to
show the subtle difference (similar to the regression tests):
https://alive2.llvm.org/ce/z/hp4hv3 (this is ok)
https://alive2.llvm.org/ce/z/ehEWaN (poison leakage)
SLP tends to create these patterns (as shown in the SLP tests), and this could
help with solving PR16739.
Differential Revision: https://reviews.llvm.org/D86460
The legacy PM alias analysis pipeline by default includes basic-aa.
When running `opt -foo-pass` under the NPM and -disable-basic-aa is not
specified, use basic-aa.
This decreases the number of check-llvm failures under NPM from 913 to 752.
Reviewed By: ychen, asbirlea
Differential Revision: https://reviews.llvm.org/D86167
This reverts commit 52b71aa8b1.
The problem was a missing lit.local.cfg file, which was causing the
test to be incorrectly run on bots that had not built the WebAssembly
target.
8cc911fa5b refactored the `getIntrinsicInstrCost` function and was
meant to be a nonfunctional change, but it accidentally changed how
costs were calculated in the SLP vectorizer, which regressed
WebAssembly codegen and resulted in a downstream bug report at
https://github.com/emscripten-core/emscripten/issues/11449.
The fix for this regression is in D85759, and this patch just
pre-commits the test from that patch to demonstrate the regressed
behavior first.
The entries in VectorizableTree are not necessarily ordered by their
position in basic blocks. Collect them and order them by dominance so
later instructions are guaranteed to be visited first. For instructions
in different basic blocks, we only scan to the beginning of the block,
so their order does not matter, as long as all instructions in a basic
block are grouped together. Using dominance ensures a deterministic order.
The modified test case contains an example where we compute a wrong
spill cost (2) without this patch, even though there is no call between
any instruction in the bundle.
This seems to have limited practical impact, .e.g on X86 with a recent
Intel Xeon CPU with -O3 -march=native -flto on MultiSource,SPEC2000,SPEC2006
there are no binary changes.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D82444
Summary:
This patch takes the indices operands of `insertelement`/`insertvalue`
into account while generation of seed elements for `findBuildAggregate()`.
This function has kept the original order of `insert`s before.
Also this patch optimizes `findBuildAggregate()` preventing it from
redundant temporary vector allocations and its multiple reversing.
Fixes llvm.org/pr44067
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83779
In vectorizeChainsInBlock we try to collect chains of PHI nodes
that have the same element type, but the code is relying upon
the implicit conversion from TypeSize -> uint64_t. For now, I have
modified the code to ignore PHI nodes with scalable types.
Differential Revision: https://reviews.llvm.org/D83542
Teaches the SLPVectorizer to use vectorized library functions for
non-intrinsic calls.
This already worked for intrinsics that have vectorized library
functions, thanks to D75878, but schedules with library functions with a
vector variant were being rejected early.
- assume that there are no load/store dependencies between lib
functions with a vector variant; this would otherwise prevent the
bundle from becoming "ready"
- check during legalization that the vector variant can be used
- fix-up where we previously assumed that a call would be an intrinsic
Differential Revision: https://reviews.llvm.org/D82550
At the moment this place does not check maximum size set
by TTI and just creates a maximum possible vectors.
Differential Revision: https://reviews.llvm.org/D82227
The entries in VectorizableTree are not necessarily ordered by their
position in basic blocks. Collect them and order them by dominance so
later instructions are guaranteed to be visited first. For instructions
in different basic blocks, we only scan to the beginning of the block,
so their order does not matter, as long as all instructions in a basic
block are grouped together. Using dominance ensures a deterministic order.
The modified test case contains an example where we compute a wrong
spill cost (2) without this patch, even though there is no call between
any instruction in the bundle.
This seems to have limited practical impact, .e.g on X86 with a recent
Intel Xeon CPU with -O3 -march=native -flto on MultiSource,SPEC2000,SPEC2006
there are no binary changes.
Reviewers: craig.topper, RKSimon, xbolva00, ABataev, spatel
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D82444
D68667 introduced a tighter limit to the number of GEPs to simplify
together. The limit was based on the vector element size of the pointer,
but the pointers themselves are not actually put in vectors.
IIUC we try to vectorize the index computations here, so we should base
the limit on the vector element size of the computation of the index.
This restores the test regression on AArch64 and also restores the
vectorization for a important pattern in SPEC2006/464.h264ref on
AArch64 (@test_i16_extend). We get a large benefit from doing a single
load up front and then processing the index computations in vectors.
Note that we could probably even further improve the AArch64 codegen, if
we would do zexts to i32 instead of i64 for the sub operands and then do
a single vector sext on the result of the subtractions. AArch64 provides
dedicated vector instructions to do so. Sketch of proof in Alive:
https://alive2.llvm.org/ce/z/A4xYAB
Reviewers: craig.topper, RKSimon, xbolva00, ABataev, spatel
Reviewed By: ABataev, spatel
Differential Revision: https://reviews.llvm.org/D82418
Motivating examples are seen in the PhaseOrdering tests based on:
https://bugs.llvm.org/show_bug.cgi?id=43953#c2 - if we have
intrinsics there, some pass can fold them.
The intrinsics are still named "experimental" at this point, but
if there is no fallout from this patch, that will be a good
indicator that it is safe to finalize them.
Differential Revision: https://reviews.llvm.org/D80867
relevant aggregate build instructions only (UserCost).
Users are detected with findBuildAggregate routine and the trick is
that following SLP vectorization may end up vectorizing entire list
with smaller chunks. Cost adjustment then is applied for individual
chunks and these adjustments obviously have to be smaller than the
entire aggregate build cost.
Differential Revision: https://reviews.llvm.org/D80773
The bug is related to aggregate build cost model adjustment
that adds a bias to cost triggering vectorization of actually
unprofitable to vectorize tree.
Differential Revision: https://reviews.llvm.org/D80682
This test was failing verification because the
metadata is ill-formed. This commit is split
from D80401 because it is an independent fix
(although the test would break with that change).
Summary:
Replace any extant metadata uses of a dying instruction with undef to
preserve debug info accuracy. Some alternatives include:
- Treat Instruction like any other Value, and point its extant metadata
uses to an empty ValueAsMetadata node. This makes extant dbg.value uses
trivially dead (i.e. fair game for deletion in many passes), leading to
stale dbg.values being in effect for too long.
- Call salvageDebugInfoOrMarkUndef. Not needed to make instruction removal
correct. OTOH results in wasted work in some common cases (e.g. when all
instructions in a BasicBlock are deleted).
This came up while discussing some basic cases in
https://reviews.llvm.org/D80052.
Reviewers: jmorse, TWeaver, aprantl, dexonsmith, jdoerfert
Subscribers: jholewinski, qcolombet, hiraditya, jfb, sstefan1, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D80264
This is D77454, except for stores. All the infrastructure work was done
for loads, so the remaining changes necessary are relatively small.
Differential Revision: https://reviews.llvm.org/D79968
For IR generated by a compiler, this is really simple: you just take the
datalayout from the beginning of the file, and apply it to all the IR
later in the file. For optimization testcases that don't care about the
datalayout, this is also really simple: we just use the default
datalayout.
The complexity here comes from the fact that some LLVM tools allow
overriding the datalayout: some tools have an explicit flag for this,
some tools will infer a datalayout based on the code generation target.
Supporting this properly required plumbing through a bunch of new
machinery: we want to allow overriding the datalayout after the
datalayout is parsed from the file, but before we use any information
from it. Therefore, IR/bitcode parsing now has a callback to allow tools
to compute the datalayout at the appropriate time.
Not sure if I covered all the LLVM tools that want to use the callback.
(clang? lli? Misc IR manipulation tools like llvm-link?). But this is at
least enough for all the LLVM regression tests, and IR without a
datalayout is not something frontends should generate.
This change had some sort of weird effects for certain CodeGen
regression tests: if the datalayout is overridden with a datalayout with
a different program or stack address space, we now parse IR based on the
overridden datalayout, instead of the one written in the file (or the
default one, if none is specified). This broke a few AVR tests, and one
AMDGPU test.
Outside the CodeGen tests I mentioned, the test changes are all just
fixing CHECK lines and moving around datalayout lines in weird places.
Differential Revision: https://reviews.llvm.org/D78403
Fix the assumption that all bitcasts of the same type sizes are free.
We now only assume that bitcasts between ints and ptrs of the same
size are free. This allows TTImpl to just call the concrete
implementation of getCastInstrCost.
Differential Revision: https://reviews.llvm.org/D78918
The original patch (rG86dfbc676ebe) exposed an existing bug:
we could wrongly cast a constant expression to BinaryOperator
because the pattern matching allows that. This adds a check
for that case, and there's a reduced test case to verify no
crashing.
Original commit message:
This builds on the or-reduction bailout that was added with D67841.
We still do not have IR-level load combining, although that could
be a target-specific enhancement for -vector-combiner.
The heuristic is narrowly defined to catch the motivating case from
PR39538:
https://bugs.llvm.org/show_bug.cgi?id=39538
...while preserving existing functionality.
That is, there's an unmodified test of pure load/zext/store that is
not seen in this patch at llvm/test/Transforms/SLPVectorizer/X86/cast.ll.
That's the reason for the logic difference to require the 'or'
instructions. The chances that vectorization would actually help a
memory-bound sequence like that seem small, but it looks nicer with:
vpmovzxwd (%rsi), %xmm0
vmovdqu %xmm0, (%rdi)
rather than:
movzwl (%rsi), %eax
movl %eax, (%rdi)
...
In the motivating test, we avoid creating a vector mess that is
unrecoverable in the backend, and SDAG forms the expected bswap
instructions after load combining:
movzbl (%rdi), %eax
vmovd %eax, %xmm0
movzbl 1(%rdi), %eax
vmovd %eax, %xmm1
movzbl 2(%rdi), %eax
vpinsrb $4, 4(%rdi), %xmm0, %xmm0
vpinsrb $8, 8(%rdi), %xmm0, %xmm0
vpinsrb $12, 12(%rdi), %xmm0, %xmm0
vmovd %eax, %xmm2
movzbl 3(%rdi), %eax
vpinsrb $1, 5(%rdi), %xmm1, %xmm1
vpinsrb $2, 9(%rdi), %xmm1, %xmm1
vpinsrb $3, 13(%rdi), %xmm1, %xmm1
vpslld $24, %xmm0, %xmm0
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpslld $16, %xmm1, %xmm1
vpor %xmm0, %xmm1, %xmm0
vpinsrb $1, 6(%rdi), %xmm2, %xmm1
vmovd %eax, %xmm2
vpinsrb $2, 10(%rdi), %xmm1, %xmm1
vpinsrb $3, 14(%rdi), %xmm1, %xmm1
vpinsrb $1, 7(%rdi), %xmm2, %xmm2
vpinsrb $2, 11(%rdi), %xmm2, %xmm2
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpinsrb $3, 15(%rdi), %xmm2, %xmm2
vpslld $8, %xmm1, %xmm1
vpmovzxbd %xmm2, %xmm2 # xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero
vpor %xmm2, %xmm1, %xmm1
vpor %xmm1, %xmm0, %xmm0
vmovdqu %xmm0, (%rsi)
movl (%rdi), %eax
movl 4(%rdi), %ecx
movl 8(%rdi), %edx
movbel %eax, (%rsi)
movbel %ecx, 4(%rsi)
movl 12(%rdi), %ecx
movbel %edx, 8(%rsi)
movbel %ecx, 12(%rsi)
Differential Revision: https://reviews.llvm.org/D78997
This builds on the or-reduction bailout that was added with D67841.
We still do not have IR-level load combining, although that could
be a target-specific enhancement for -vector-combiner.
The heuristic is narrowly defined to catch the motivating case from
PR39538:
https://bugs.llvm.org/show_bug.cgi?id=39538
...while preserving existing functionality.
That is, there's an unmodified test of pure load/zext/store that is
not seen in this patch at llvm/test/Transforms/SLPVectorizer/X86/cast.ll.
That's the reason for the logic difference to require the 'or'
instructions. The chances that vectorization would actually help a
memory-bound sequence like that seem small, but it looks nicer with:
vpmovzxwd (%rsi), %xmm0
vmovdqu %xmm0, (%rdi)
rather than:
movzwl (%rsi), %eax
movl %eax, (%rdi)
...
In the motivating test, we avoid creating a vector mess that is
unrecoverable in the backend, and SDAG forms the expected bswap
instructions after load combining:
movzbl (%rdi), %eax
vmovd %eax, %xmm0
movzbl 1(%rdi), %eax
vmovd %eax, %xmm1
movzbl 2(%rdi), %eax
vpinsrb $4, 4(%rdi), %xmm0, %xmm0
vpinsrb $8, 8(%rdi), %xmm0, %xmm0
vpinsrb $12, 12(%rdi), %xmm0, %xmm0
vmovd %eax, %xmm2
movzbl 3(%rdi), %eax
vpinsrb $1, 5(%rdi), %xmm1, %xmm1
vpinsrb $2, 9(%rdi), %xmm1, %xmm1
vpinsrb $3, 13(%rdi), %xmm1, %xmm1
vpslld $24, %xmm0, %xmm0
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpslld $16, %xmm1, %xmm1
vpor %xmm0, %xmm1, %xmm0
vpinsrb $1, 6(%rdi), %xmm2, %xmm1
vmovd %eax, %xmm2
vpinsrb $2, 10(%rdi), %xmm1, %xmm1
vpinsrb $3, 14(%rdi), %xmm1, %xmm1
vpinsrb $1, 7(%rdi), %xmm2, %xmm2
vpinsrb $2, 11(%rdi), %xmm2, %xmm2
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpinsrb $3, 15(%rdi), %xmm2, %xmm2
vpslld $8, %xmm1, %xmm1
vpmovzxbd %xmm2, %xmm2 # xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero
vpor %xmm2, %xmm1, %xmm1
vpor %xmm1, %xmm0, %xmm0
vmovdqu %xmm0, (%rsi)
movl (%rdi), %eax
movl 4(%rdi), %ecx
movl 8(%rdi), %edx
movbel %eax, (%rsi)
movbel %ecx, 4(%rsi)
movl 12(%rdi), %ecx
movbel %edx, 8(%rsi)
movbel %ecx, 12(%rsi)
Differential Revision: https://reviews.llvm.org/D78997
The improvements to the x86 vector insert/extract element costs in D74976 resulted in the estimated costs for vector initialization and scalarization increasing higher than should be expected. This is particularly noticeable on pre-SSE4 targets where the available of legal INSERT_VECTOR_ELT ops is more limited.
This patch does 2 things:
1 - it implements X86TTIImpl::getScalarizationOverhead to more accurately represent the typical costs of a ISD::BUILD_VECTOR pattern.
2 - it adds a DemandedElts mask to getScalarizationOverhead to permit the SLP's BoUpSLP::getGatherCost to be rewritten to use it directly instead of accumulating raw vector insertion costs.
This fixes PR45418 where a v4i8 (zext'd to v4i32) was no longer vectorizing.
A future patch should extend X86TTIImpl::getScalarizationOverhead to tweak the EXTRACT_VECTOR_ELT scalarization costs as well.
Reviewed By: @craig.topper
Differential Revision: https://reviews.llvm.org/D78216
Summary:
Currently, the internal options -vectorize-loops, -vectorize-slp, and
-interleave-loops do not have much practical effect. This is because
they are used to initialize the corresponding flags in the pass
managers, and those flags are then unconditionally overwritten when
compiling via clang or via LTO from the linkers. The only exception was
-vectorize-loops via opt because of some special hackery there.
While vectorization could still be disabled when compiling via clang,
using -fno-[slp-]vectorize, this meant that there was no way to disable
it when compiling in LTO mode via the linkers. This only affected
ThinLTO, since for regular LTO vectorization is done during the compile
step for scalability reasons. For ThinLTO it is invoked in the LTO
backends. See also the discussion on PR45434.
This patch makes it so the internal options can actually be used to
disable these optimizations. Ultimately, the best long term solution is
to mark the loops with metadata (similar to the approach used to fix
-fno-unroll-loops in D77058), but this enables a shorter term
workaround, and actually makes these internal options useful.
I constant propagated the initial values of these internal flags into
the pass manager flags (for some reasons vectorize-loops and
interleave-loops were initialized to true, while vectorize-slp was
initialized to false). As mentioned above, they are overwritten
unconditionally so this doesn't have any real impact, and these initial
values aren't particularly meaningful.
I then changed the passes to check the internl values and return without
performing the associated optimization when false (I changed the default
of -vectorize-slp to true so the options behave similarly). I was able
to remove the hackery in opt used to get -vectorize-loops=false to work,
as well as a special option there used to disable SLP vectorization.
Finally, I changed thinlto-slp-vectorize-pm.c to:
a) Only test SLP (moved the loop vectorization checking to a new test).
b) Use code that is slp vectorized when it is enabled, and check that
instead of whether the pass is enabled.
c) Test the new behavior of -vectorize-slp.
d) Test both pass managers.
The loop vectorization (and associated interleaving) testing I moved to
a new thinlto-loop-vectorize-pm.c test, with several changes:
a) Changed the flags on the interleaving testing so that it will
actually interleave, and check that.
b) Test the new behavior of -vectorize-loops and -interleave-loops.
c) Test both pass managers.
Reviewers: fhahn, wmi
Subscribers: hiraditya, steven_wu, dexonsmith, cfe-commits, davezarzycki, llvm-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D77989
This is similar to what I recently did for getArithmeticReductionCost.
I'm trying to account for the narrowing from 512->256->128 as we go.
I've also added a new helper method getMinMaxCost that tries to
handle the cases where we have native min/max instructions and
fall back to cmp+select when we don't.
Differential Revision: https://reviews.llvm.org/D76634
There seems to be a small benefit to the legalized sequence for v2f16
round with packed instructions, so allow vectorizing it by reducing
the cost.
An unintended side effect is vectorization of f32 round also
happens. The current FMA logic seems off to me, and isn't checking for
packed instructions.
This patch attempts to more accurately model the reduction of
power of 2 vectors of types we natively support. This takes into
account the narrowing of vectors that occur as we go from 512
bits to 256 bits, to 128 bits. It also takes into account the use
of wider elements in the shuffles for the first 2 steps of a
reduction from 128 bits. And uses a v8i16 shift for the final step
of vXi8 reduction.
The default implementation uses the legalized type for the arithmetic
for all levels. And uses the single source permute cost of the
legalized type for all levels. This penalizes things like
lack of v16i8 pshufb on pre-sse3 targets and the splitting and
joining that needs to be done for integer types on AVX1. We never
need v16i8 shuffle for a reduction and we only need split AVX1 ops
when type the type wide and needs to be split. I think we're still
over costing splits and joins for AVX1, but we're closer now.
I've also removed all pairwise special casing because I don't
think we ever want to generate that on X86. I've also adjusted
the add handling to more accurately account for any type splitting
that occurs before we reach a legal type.
Differential Revision: https://reviews.llvm.org/D76478
Summary:
SLPVectorizer try to vectorize list of scalar instructions of the same type,
instructions already vectorized are rejected through isValidElementType().
Without this patch, tryToVectorizeList() will first try to determine vectorization
factor of a list of Instructions before checking whether each instruction has unsupported
type or not. For instructions already vectorized for SVE, it will crash at getVectorElementSize(),
where it try to return a fixed size.
This patch make sure invalid element types are rejected before trying to get vectorization
factor. This make sure we are not trying to vectorize instructions already vectorized.
Reviewers: sdesmalen, efriedma, spatel, RKSimon, ABataev, apazos, rengolin
Reviewed By: efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76017
It seems like the SLPVectorizer is currently not aware of vector
versions of functions provided by libraries like Accelerate [1].
This patch updates SLPVectorizer to use the same infrastructure
the LoopVectorizer uses to detect vectorizable library functions.
For calls, it computes the cost of an intrinsic call (existing behavior)
and the cost of a vector function library call, if available. Like
LoopVectorizer, it assumes the cost of the vector function is simply the
cost of a call to a vector function.
[1] https://developer.apple.com/documentation/accelerate
Reviewers: ABataev, RKSimon, spatel
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D75878
This tries to improve the accuracy of extract/insert element costs by accounting for subvector extraction/insertion for >128-bit vectors and the shuffling of elements to/from the 0'th index.
It also adds INSERTPS for f32 types and PINSR/PEXTR costs for integer types (at the moment we assume the same cost as MOVD/MOVQ - which isn't always true).
Differential Revision: https://reviews.llvm.org/D74976
D74976 will handle larger vector types, but since SLM doesn't support AVX+ then we will always be extracting from 128-bit vectors so don't need to scale the cost.
The index of an ExtractElementInst is not guaranteed to be a
ConstantInt. It can be any integer value. Check explicitly for
ConstantInts.
The new test cases illustrate scenarios where we crash without
this patch. I've also added another test case to check the matching
of extractelement vector ops works.
Reviewers: RKSimon, ABataev, dtemirbulatov, vporpo
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D74758
We may calculate reassociable math ops in arbitrary order when creating a shuffle reduction,
so there's no guarantee that things like 'nsw' hold on those intermediate values. Drop all
poison-generating flags for safety.
This change is limited to shuffle reductions because I don't think we have a problem in the
general case (where we intersect flags of each scalar op that goes into a vector op), but if
there's evidence of other cases being wrong, we can extend this fix to cover those cases.
https://bugs.llvm.org/show_bug.cgi?id=44536
Differential Revision: https://reviews.llvm.org/D73727
Summary:
We don't have control/verify what will be the RHS of the division, so it might
happen to be zero, causing UB.
Reviewers: Vasilis, RKSimon, ABataev
Reviewed By: ABataev
Subscribers: vporpo, ABataev, hiraditya, llvm-commits, vdmitrie
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72740
Summary:
Make SLPVectorize to recognize homogeneous aggregates like
`{<2 x float>, <2 x float>}`, `{{float, float}, {float, float}}`,
`[2 x {float, float}]` and so on.
It's a follow-up of https://reviews.llvm.org/D70068.
Merged `findBuildVector()` and `findBuildAggregate()` to
one `findBuildAggregate()` function making it recursive
to recognize multidimensional aggregates. Aggregates required
to be homogeneous.
Reviewers: RKSimon, ABataev, dtemirbulatov, spatel, vporpo
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70587
I'm not sure what the effect of this change will be on all of the affected
tests or a larger benchmark, but it fixes the horizontal add/sub problems
noted here:
https://reviews.llvm.org/D59710?vs=227972&id=228095&whitespace=ignore-most#toc
The costs are based on reciprocal throughput numbers in Agner's tables for
PEXTR*; these appear to be very slow ops on Silvermont.
This is a small step towards the larger motivation discussed in PR43605:
https://bugs.llvm.org/show_bug.cgi?id=43605
Also, it seems likely that insert/extract is the source of perf regressions on
other CPUs (up to 30%) that were cited as part of the reason to revert D59710,
so maybe we'll extend the table-based approach to other subtargets.
Differential Revision: https://reviews.llvm.org/D70607
Summary:
Vector aggregate is homogeneous aggregate of vectors like `{ <2 x float>, <2 x float> }`.
This patch allows `findBuildAggregate()` to consider vector aggregates as
well as scalar ones. For instance, `{ <2 x float>, <2 x float> }` maps to `<4 x float>`.
Fixes vector part of llvm.org/PR42022
Reviewers: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70068
After speaking with Sanjay - seeing a number of miscompiles and working
on tracking down a testcase. None of the follow on patches seem to
have helped so far.
This reverts commit 8a0aa5310b.
After speaking with Sanjay - seeing a number of miscompiles and working
on tracking down a testcase. None of the follow on patches seem to
have helped so far.
This reverts commit 7ff57705ba.
The 1st attempt was reverted because it revealed an existing
bug where we could produce invalid IR (use of value before
definition). That should be fixed with:
rG39de82ecc9c2
The bug manifests as replacing a reduction operand with an undef
value.
The problem appears to be limited to cases where a min/max reduction
has extra uses of the compare operand to the select.
In the general case, we are tracking "ExternallyUsedValues" and
an "IgnoreList" of the reduction operations, but those may not apply
to the final compare+select in a min/max reduction.
For that, we use replaceAllUsesWith (RAUW) to ensure that the new
vectorized reduction values are transferred to all subsequent users.
Differential Revision: https://reviews.llvm.org/D70148
As discussed in D70148 (and caused a revert of the original commit):
if we insert at the select, then we can produce invalid IR because
the replacement for the compare may have uses before the select.
This reverts commit e511c4b0dff1692c267addf17dce3cebe8f97faa:
Temporarily Revert:
"[SLP] Generalization of stores vectorization."
"[SLP] Fix -Wunused-variable. NFC"
"[SLP] Vectorize jumbled stores."
after fixing the problem with compile time.
The bug manifests as replacing a reduction operand with an undef
value.
The problem appears to be limited to cases where a min/max reduction
has extra uses of the compare operand to the select.
In the general case, we are tracking "ExternallyUsedValues" and
an "IgnoreList" of the reduction operations, but those may not apply
to the final compare+select in a min/max reduction.
For that, we use replaceAllUsesWith (RAUW) to ensure that the new
vectorized reduction values are transferred to all subsequent users.
Differential Revision: https://reviews.llvm.org/D70148
Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for examples).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: xbolva00, Carrot, hiraditya, phosek, rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
We have a vector compare reduction problem seen in PR39665 comment 2:
https://bugs.llvm.org/show_bug.cgi?id=39665#c2
Or slightly reduced here:
define i1 @cmp2(<2 x double> %a0) {
%a = fcmp ogt <2 x double> %a0, <double 1.0, double 1.0>
%b = extractelement <2 x i1> %a, i32 0
%c = extractelement <2 x i1> %a, i32 1
%d = and i1 %b, %c
ret i1 %d
}
SLP would not attempt to turn this into a vector reduction because there is an
artificial lower limit on that transform. We can not completely remove that limit
without inducing regressions though, so this patch just hacks an extra attempt at
creating a 2-way reduction to the end of the analysis.
As shown in the test file, we are still not getting some of the motivating cases,
so follow-on patches will be needed to solve those cases.
Differential Revision: https://reviews.llvm.org/D59710
"[SLP] Generalization of stores vectorization."
"[SLP] Fix -Wunused-variable. NFC"
"[SLP] Vectorize jumbled stores."
As they're causing significant (10-30x) compile time regressions on
vectorizable code.
The primary cause of the compile-time regression is f228b53716.
This reverts commits:
f228b537165503455ccb21d498c9c0
Summary:
If the GEP instructions are going to be vectorized, the indices in those
GEP instructions must be of the same type. Otherwise, the compiler may
crash when trying to build the vector constant.
Reviewers: RKSimon, spatel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69627
The script uses 'TMP#' as its substitute for nameless values,
so if a test already contains 'tmp#' *named* values, then
there could be trouble. We should probably just fix the
script to avoid this problem going forward, but it's easy
enough to change a test too (and explicitly naming variables
'tmp' is always a sad choice).
The script uses 'TMP#' as its substitute for nameless values,
so if a test already contains 'tmp#' *named* values, then
there could be trouble. We should probably just fix the
script to avoid this problem going forward, but it's easy
enough to change a test too (and explicitly naming variables
'tmp' is always a sad choice).
The script uses 'TMP#' as its substitute for nameless values,
so if a test already contains 'tmp#' *named* values, then
there could be trouble. We should probably just fix the
script to avoid this problem going forward, but it's easy
enough to change a test too (and explicitly naming variables
'tmp' is always a sad choice).
Summary:
Patch adds support for vectorization of the jumbled stores. The value
operands are vectorized and then shuffled in the right order before
store.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43339
Summary:
Patch adds support for vectorization of the jumbled stores. The value
operands are vectorized and then shuffled in the right order before
store.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43339
Stores are vectorized with maximum vectorization factor of 16. Patch
tries to improve the situation and use maximal vectorization factor.
Reviewers: spatel, RKSimon, mkuper, hfinkel
Differential Revision: https://reviews.llvm.org/D43582
The 1st attempt at this modified the cost model in a bad way to avoid the vectorization,
but that caused problems for other users (the loop vectorizer) of the cost model.
I don't see an ideal solution to these 2 related, potentially large, perf regressions:
https://bugs.llvm.org/show_bug.cgi?id=42708https://bugs.llvm.org/show_bug.cgi?id=43146
We decided that load combining was unsuitable for IR because it could obscure other
optimizations in IR. So we removed the LoadCombiner pass and deferred to the backend.
Therefore, preventing SLP from destroying load combine opportunities requires that it
recognizes patterns that could be combined later, but not do the optimization itself (
it's not a vector combine anyway, so it's probably out-of-scope for SLP).
Here, we add a cost-independent bailout with a conservative pattern match for a
multi-instruction sequence that can probably be reduced later.
In the x86 tests shown (and discussed in more detail in the bug reports), SDAG combining
will produce a single instruction on these tests like:
movbe rax, qword ptr [rdi]
or:
mov rax, qword ptr [rdi]
Not some (half) vector monstrosity as we currently do using SLP:
vpmovzxbq ymm0, dword ptr [rdi + 1] # ymm0 = mem[0],zero,zero,..
vpsllvq ymm0, ymm0, ymmword ptr [rip + .LCPI0_0]
movzx eax, byte ptr [rdi]
movzx ecx, byte ptr [rdi + 5]
shl rcx, 40
movzx edx, byte ptr [rdi + 6]
shl rdx, 48
or rdx, rcx
movzx ecx, byte ptr [rdi + 7]
shl rcx, 56
or rcx, rdx
or rcx, rax
vextracti128 xmm1, ymm0, 1
vpor xmm0, xmm0, xmm1
vpshufd xmm1, xmm0, 78 # xmm1 = xmm0[2,3,0,1]
vpor xmm0, xmm0, xmm1
vmovq rax, xmm0
or rax, rcx
vzeroupper
ret
Differential Revision: https://reviews.llvm.org/D67841
llvm-svn: 375025
Add specific scalar costs for CTLZ instructions, we can't discriminate between CTLZ and CTLZ_ZERO_UNDEF so we have to assume the worst. Given how BSR is often a microcoded nightmare on some older targets we might still be underestimating it.
For targets supporting LZCNT (Intel Haswell+ or AMD Fam10+), we provide overrides that assume 1cy costs.
llvm-svn: 374786
Add specific scalar costs for ctpop instructions, these are based on the llvm-mca's SLM throughput numbers (the oldest model we have).
For targets supporting POPCNT, we provide overrides that assume 1cy costs.
llvm-svn: 374775
I can't see any notable differences in costs between SSE2 and SSE42 arches for FADD/ADD reduction, so I've lowered the target to just SSE2.
I've also added vXi8 sum reduction costs in line with the PSADBW codegen and discussions on PR42674.
llvm-svn: 374655
We failed to account for the target register width (max vector factor)
when vectorizing starting from GEPs. This causes vectorization to
proceed to obviously illegal widths as in:
https://bugs.llvm.org/show_bug.cgi?id=43578
For x86, this also means that SLP can produce rogue AVX or AVX512
code even when the user specifies a narrower vector width.
The AArch64 test in ext-trunc.ll appears to be better using the
narrower width. I'm not exactly sure what getelementptr.ll is trying
to do, but it's testing with "-slp-threshold=-18", so I'm not worried
about those diffs. The x86 test is an over-reduction from SPEC h264;
this patch appears to restore the perf loss caused by SLP when using
-march=haswell.
Differential Revision: https://reviews.llvm.org/D68667
llvm-svn: 374183
This reverts SVN r373833, as it caused a failed assert "Non-zero loop
cost expected" on building numerous projects, see PR43582 for details
and reproduction samples.
llvm-svn: 373882
I don't see an ideal solution to these 2 related, potentially large, perf regressions:
https://bugs.llvm.org/show_bug.cgi?id=42708https://bugs.llvm.org/show_bug.cgi?id=43146
We decided that load combining was unsuitable for IR because it could obscure other
optimizations in IR. So we removed the LoadCombiner pass and deferred to the backend.
Therefore, preventing SLP from destroying load combine opportunities requires that it
recognizes patterns that could be combined later, but not do the optimization itself (
it's not a vector combine anyway, so it's probably out-of-scope for SLP).
Here, we add a scalar cost model adjustment with a conservative pattern match and cost
summation for a multi-instruction sequence that can probably be reduced later.
This should prevent SLP from creating a vector reduction unless that sequence is
extremely cheap.
In the x86 tests shown (and discussed in more detail in the bug reports), SDAG combining
will produce a single instruction on these tests like:
movbe rax, qword ptr [rdi]
or:
mov rax, qword ptr [rdi]
Not some (half) vector monstrosity as we currently do using SLP:
vpmovzxbq ymm0, dword ptr [rdi + 1] # ymm0 = mem[0],zero,zero,..
vpsllvq ymm0, ymm0, ymmword ptr [rip + .LCPI0_0]
movzx eax, byte ptr [rdi]
movzx ecx, byte ptr [rdi + 5]
shl rcx, 40
movzx edx, byte ptr [rdi + 6]
shl rdx, 48
or rdx, rcx
movzx ecx, byte ptr [rdi + 7]
shl rcx, 56
or rcx, rdx
or rcx, rax
vextracti128 xmm1, ymm0, 1
vpor xmm0, xmm0, xmm1
vpshufd xmm1, xmm0, 78 # xmm1 = xmm0[2,3,0,1]
vpor xmm0, xmm0, xmm1
vmovq rax, xmm0
or rax, rcx
vzeroupper
ret
Differential Revision: https://reviews.llvm.org/D67841
llvm-svn: 373833
Initially SLP vectorizer replaced all going-to-be-vectorized
instructions with Undef values. It may break ScalarEvaluation and may
cause a crash.
Reworked SLP vectorizer so that it does not replace vectorized
instructions by UndefValue anymore. Instead vectorized instructions are
marked for deletion inside if BoUpSLP class and deleted upon class
destruction.
Reviewers: mzolotukhin, mkuper, hfinkel, RKSimon, davide, spatel
Subscribers: RKSimon, Gerolf, anemet, hans, majnemer, llvm-commits, sanjoy
Differential Revision: https://reviews.llvm.org/D29641
llvm-svn: 373166
SLM is 2 x slower for <2 x i64> comparison ops than other vector types, we should account for this like we do for SLM <2 x i64> add/sub/mul costs.
This should remove some of the SLM codegen diffs in D43582
llvm-svn: 372954
Summary:
Initially SLP vectorizer replaced all going-to-be-vectorized
instructions with Undef values. It may break ScalarEvaluation and may
cause a crash.
Reworked SLP vectorizer so that it does not replace vectorized
instructions by UndefValue anymore. Instead vectorized instructions are
marked for deletion inside if BoUpSLP class and deleted upon class
destruction.
Reviewers: mzolotukhin, mkuper, hfinkel, RKSimon, davide, spatel
Subscribers: RKSimon, Gerolf, anemet, hans, majnemer, llvm-commits, sanjoy
Differential Revision: https://reviews.llvm.org/D29641
llvm-svn: 372626
We are missing costs for a lot of truncation cases, I'm hoping to address all the 'zero cost' cases in trunc.ll
I thought this was a vector widening side effect, but even before this we had some interesting LV decisions (notably over indvars) being made due to these zero costs.
llvm-svn: 372498
This is a fix for:
https://bugs.llvm.org/show_bug.cgi?id=33958
It seems universally true that we would not want to transform this kind of
sequence on any target, but if that's not correct, then we could view this
as a target-specific cost model problem. We could also white-list ConstantInt,
ConstantFP, etc. rather than blacklist Global and ConstantExpr.
Differential Revision: https://reviews.llvm.org/D67362
llvm-svn: 371931
Summary:
Similar to the previous prefer-256-bit flag. We might want to
enable this by default some CPUs. This just starts the initial
work to implement and prove that it effects TTI's vector width.
Reviewers: RKSimon, echristo, spatel, atdt
Reviewed By: RKSimon
Subscribers: lebedev.ri, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67311
llvm-svn: 371319
Now that we legalize by widening, the element types here won't change. Previously these were modeled as the elements being widened and then the instruction might become an AND or SHL/ASHR pair. But now they'll become something like a ZERO_EXTEND_VECTOR_INREG/SIGN_EXTEND_VECTOR_INREG.
For AVX2, when the destination type is legal its clear the cost should be 1 since we have extend instructions that can produce 256 bit vectors from less than 128 bit vectors. I'm a little less sure about AVX1 costs, but I think the ones I changed were definitely too high, but they might still be too high.
Differential Revision: https://reviews.llvm.org/D66169
llvm-svn: 368858
The assert that caused this to be reverted should be fixed now.
Original commit message:
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 368183
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 367901
Summary:
In D62801, new function attribute `willreturn` was introduced. In short, a function with `willreturn` is guaranteed to come back to the call site(more precise definition is in LangRef).
In this patch, willreturn is annotated for LLVM intrinsics.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: jvesely, nhaehnle, sstefan1, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64904
llvm-svn: 367184
Summary:
- As the pointer stripping now tracks through `addrspacecast`, prepare
to handle the bit-width difference from the result pointer.
Reviewers: jdoerfert
Subscribers: jvesely, nhaehnle, hiraditya, arphaman, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64928
llvm-svn: 366470
As there are some reported miscompiles with AVX512 and performance regressions
in Eigen. Verified with the original committer and testcases will be forthcoming.
This reverts commit r364964.
llvm-svn: 366154
Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for an example).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: hiraditya, phosek, rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
llvm-svn: 364964
Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for an example).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
llvm-svn: 364478
This is a pre-commit of the tests introduced by the SuperNode SLP patch D63661.
Committed on behalf of @vporpo (Vasileios Porpodas)
Differential Revision: https://reviews.llvm.org/D63664
llvm-svn: 364320
This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for an example).
Committed on behalf of @vporpo (Vasileios Porpodas)
Differential Revision: https://reviews.llvm.org/D60897
llvm-svn: 364084
This patch uses the mechanism from D62995 to strengthen the
definitions of the reduction intrinsics by letting the scalar
result/accumulator type be overloaded from the vector element type.
For example:
; The LLVM LangRef specifies that the scalar result must equal the
; vector element type, but this is not checked/enforced by LLVM.
declare i32 @llvm.experimental.vector.reduce.or.i32.v4i32(<4 x i32> %a)
This patch changes that into:
declare i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32> %a)
Which has the type-constraint more explicit and causes LLVM to check
the result type with the vector element type.
Reviewers: RKSimon, arsenm, rnk, greened, aemerson
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D62996
llvm-svn: 363240
This patch fixes a regression caused by the operand reordering refactoring patch https://reviews.llvm.org/D59973 .
The fix changes the strategy to Splat instead of Opcode, if broadcast opportunities are found.
Please see the lit test for some examples.
Committed on behalf of @vporpo (Vasileios Porpodas)
Differential Revision: https://reviews.llvm.org/D62427
llvm-svn: 362613
The original costs stopped at SSE42, I've added conservative estimates for everything down to SSE1/SSE2 and moved some of the SSE42 costs to SSE41 (really only the addition of PCMPGT makes any difference).
I've also added missing vXi8 costs (we use PHMINPOSUW for i8/i16 for scarily quick results) and 256-bit vector costs for AVX1.
llvm-svn: 360528
The code in this test is not vectorized by SLP because its operand reordering cannot look beyond the immediate predecessors.
This will get fixed in a follow-up patch that introduces the look-ahead operand reordering heuristic.
Committed on behalf of @vporpo (Vasileios Porpodas)
Differential Revision: https://reviews.llvm.org/D61283
llvm-svn: 359553
Summary: The code did not check if operand was undef before casting it to Instruction.
Reviewers: RKSimon, ABataev, dtemirbulatov
Reviewed By: ABataev
Subscribers: uabelho
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61024
llvm-svn: 359136
As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
This is a refactoring patch which should have all the functionality of the current code. Its goal is twofold:
i. Cleanup and simplify the reordering code, and
ii. Generalize reordering so that it will work for an arbitrary number of operands, not just 2.
This is the second patch in a series of patches that will enable operand reordering across chains of operations. An example of this was presented in EuroLLVM'18 https://www.youtube.com/watch?v=gIEn34LvyNo .
Committed on behalf of @vporpo (Vasileios Porpodas)
Differential Revision: https://reviews.llvm.org/D59973
llvm-svn: 358519
In PR41304:
https://bugs.llvm.org/show_bug.cgi?id=41304
...we have a case where we want to fold a binop of select-shuffle (blended) values.
Rather than try to match commuted variants of the pattern, we can canonicalize the
shuffles and check for mask equality with commuted operands.
We don't produce arbitrary shuffle masks in instcombine, but select-shuffles are a
special case that the backend is required to handle because we already canonicalize
vector select to this shuffle form.
So there should be no codegen difference from this change. It's possible that this
improves CSE in IR though.
Differential Revision: https://reviews.llvm.org/D60016
llvm-svn: 357366
For the cases where the icmp/fcmp predicate is commutative, use reorderInputsAccordingToOpcode to collect and commute the operands.
This requires a helper to recognise commutativity in both general Instruction and CmpInstr types - the CmpInst::isCommutative doesn't overload the Instruction::isCommutative method for reasons I'm not clear on (maybe because its based on predicate not opcode?!?).
Differential Revision: https://reviews.llvm.org/D59992
llvm-svn: 357266
We should be able to match elements with the swapped predicate as well - as long as we commute the source operands.
Differential Revision: https://reviews.llvm.org/D59956
llvm-svn: 357243
As discussed on D59738, this generalizes reorderInputsAccordingToOpcode to handle multiple + non-commutative instructions so we can get rid of reorderAltShuffleOperands and make use of the extra canonicalizations that reorderInputsAccordingToOpcode brings.
Differential Revision: https://reviews.llvm.org/D59784
llvm-svn: 356939
Remove attempts to commute non-Instructions to the LHS - the codegen changes appear to rely on chance more than anything else and also have a tendency to fight existing instcombine canonicalization which moves constants to the RHS of commutable binary ops.
This is prep work towards:
(a) reusing reorderInputsAccordingToOpcode for alt-shuffles and removing the similar reorderAltShuffleOperands
(b) improving reordering to optimized cases with commutable and non-commutable instructions to still find splat/consecutive ops.
Differential Revision: https://reviews.llvm.org/D59738
llvm-svn: 356913
x86-64 is an invalid architecture in triples. Changing it to the correct
triple (x86_64) changes some tests, because SLP is not deemed profitable
any more.
Reviewers: ABataev, RKSimon, spatel
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D58931
llvm-svn: 355420
This requires a couple of tweaks to existing vectorization functions as they were assuming that only the second call argument (ctlz/cttz/powi) could ever be the 'always scalar' argument, but for smul.fix + umul.fix its the third argument.
Differential Revision: https://reviews.llvm.org/D58616
llvm-svn: 354790
As this has broken the lto bootstrap build for 3 days and is
showing a significant regression on the Dither_benchmark results (from
the LLVM benchmark suite) -- specifically, on the
BENCHMARK_FLOYD_DITHER_128, BENCHMARK_FLOYD_DITHER_256, and
BENCHMARK_FLOYD_DITHER_512; the others are unchanged. These have
regressed by about 28% on Skylake, 34% on Haswell, and over 40% on
Sandybridge.
This reverts commit r353923.
llvm-svn: 354434
Try to use 64-bit SLP vectorization. In addition to horizontal instrs
this change triggers optimizations for partial vector operations (for instance,
using low halfs of 128-bit registers xmm0 and xmm1 to multiply <2 x float> by
<2 x float>).
Fixes llvm.org/PR32433
llvm-svn: 353923
Add generic costs calculation for SADDSAT/SSUBSAT intrinsics, this uses generic costs for sadd_with_overflow/ssub_with_overflow, an extra sign comparison + a selects based on the sign/overflow.
This completes PR40316
Differential Revision: https://reviews.llvm.org/D57239
llvm-svn: 352315
For the power9 CPU, vector operations consume a pair of execution units rather
than one execution unit like a scalar operation. Update the target transform
cost functions to reflect the higher cost of vector operations when targeting
Power9.
Patch by RolandF.
Differential revision: https://reviews.llvm.org/D55461
llvm-svn: 352261
Add generic costs calculation for UADDSAT/USUBSAT intrinsics, this fallbacks to using generic costs for uadd_with_overflow/usub_with_overflow + a select.
Differential Revision: https://reviews.llvm.org/D56907
llvm-svn: 352044
Prior to SSE41 (and sometimes on AVX1), vector select has to be performed as a ((X & C)|(Y & ~C)) bit select.
Exposes a couple of issues with the min/max reduction costs (which only go down to SSE42 for some reason).
The increase pre-SSE41 selection costs also prevent a couple of tests from firing any longer, so I've either tweaked the target or added AVX tests as well to the existing SSE2 tests.
llvm-svn: 351685
Summary:
Sometimes the SLP vectorizer tries to vectorize the horizontal reduction
nodes during regular vectorization. This may happen inside of the loops,
when there are some vectorizable PHIs. Patch fixes this by checking if
the node is the reduction node and thus it must not be vectorized, it must
be gathered.
Reviewers: RKSimon, spatel, hfinkel, fedor.sergeev
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D56783
llvm-svn: 351349
official Git repository.
Remove the directions for using git-svn, and demote the prominence of
the svn instructions.
Also, fix a few other issues while I'm in there:
* Mention LLVM_ENABLE_PROJECTS more.
* Getting started doesn't need to mention test-suite, but should
mention clang and the other projects.
* Remove mentions of "configure", since that's long gone.
I've also adjusted a few other mentions of svn to point to github, but
have not done so comprehensively.
Differential Revision: https://reviews.llvm.org/D56654
llvm-svn: 351130
Summary: The comment says we need 3 extracts and a select at the end. But didn't we just account for the select in the vector cost above. Aren't we just extracting the single element after taking the min/max in the vector register?
Reviewers: RKSimon, spatel, ABataev
Reviewed By: RKSimon
Subscribers: javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D55480
llvm-svn: 348739
We were overcounting the number of arithmetic operations needed at each level before we reach a legal type. We were using the full vector type for that level, but we are going to split the input vector at that level in half. So the effective arithmetic operation cost at that level is half the width.
So for example on 8i32 on an sse target. Were were calculating the cost of an 8i32 op which is likely 2 for basic integer. Then after the loop we count 2 more v4i32 ops. For a total arith cost of 4. But if you look at the assembly there would only be 3 arithmetic ops.
There are still more bugs in this code that I'm going to work on next. The non pairwise code shouldn't count extract subvectors in the loop. There are no extracts, the types are split in registers. For pairwise we need to use 2 two src permute shuffles.
Differential Revision: https://reviews.llvm.org/D55397
llvm-svn: 348621
We were adding the entire scalarization extraction cost for reductions, which returns the total cost of extracting every element of a vector type.
For reductions we don't need to do this - we just need to extract the 0'th element after the reduction pattern has completed.
Fixes PR37731
Rebased and reapplied after being reverted in rL347541 due to PR39774 - which was fixed by D54955/rL347759 and D55017/rL347997
Differential Revision: https://reviews.llvm.org/D54585
llvm-svn: 348076
Summary:
An additional fix for PR39774. Need to update the references for the
RedcutionRoot instruction when it is replaced during the vectorization
phase to avoid compiler crash on reduction vectorization.
Reviewers: RKSimon, spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D55017
llvm-svn: 347997
Summary:
If the original reduction root instruction was vectorized, it might be
removed from the tree. It means that the insertion point may become
invalidated and the whole vectorization of the reduction leads to the
incorrect output result.
The ReductionRoot instruction must be marked as externally used so it
could not be removed. Otherwise it might cause inconsistency with the
cost model and we may end up with too optimistic optimization.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54955
llvm-svn: 347759
This reverts commit r346970.
It was causing PR39774, a crash in slp-vectorizer on a rather simple loop
with just a bunch of 'and's in the body.
llvm-svn: 347541
We were adding the entire scalarization extraction cost for reductions, which returns the total cost of extracting every element of a vector type.
For reductions we don't need to do this - we just need to extract the 0'th element after the reduction pattern has completed.
Fixes PR37731
Differential Revision: https://reviews.llvm.org/D54585
llvm-svn: 346970
Expand arithmetic reduction to include mul/and/or/xor instructions.
This patch just fixes the SLPVectorizer - the effective reduction costs for AVX1+ are still poor (see rL344846) and will need to be improved before SLP sees this as a valid transform - but we can already see the effect on SSE2 tests.
This partially helps PR37731, but doesn't fix it all as it still falls over on the extraction/reduction order for some reason.
Differential Revision: https://reviews.llvm.org/D53473
llvm-svn: 345037
We miss arithmetic reduction for everything but Add/FAdd (I assume because that's the only cases which x86 has horizontal ops for.....)
llvm-svn: 344849
In the case of soft-fp (e.g. fp128 under wasm) the result of
getTypeLegalizationCost() can be an integer type even if the input is
floating point (See LegalizeTypeAction::TypeSoftenFloat).
Before calling isFabsFree() (which asserts if given a non-fp
type) we need to check that that result is fp. This is safe since in
fabs is certainly not free in the soft-fp case.
Fixes PR39168
Differential Revision: https://reviews.llvm.org/D52899
llvm-svn: 344069
Summary:
Reworked the previously committed patch to insert shuffles for reused
extract element instructions in the correct position. Previous logic was
incorrect, and might lead to the crash with PHIs and EH instructions.
Reviewers: efriedma, javed.absar
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D50143
llvm-svn: 339166
Summary:
If the ExtractElement instructions can be optimized out during the
vectorization and we need to reshuffle the parent vector, this
ShuffleInstruction may be inserted in the wrong place causing compiler
to produce incorrect code.
Reviewers: spatel, RKSimon, mkuper, hfinkel, javed.absar
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49928
llvm-svn: 338380
as well as sext(C + x + ...) -> (D + sext(C-D + x + ...))<nuw><nsw>
similar to the equivalent transformation for zext's
if the top level addition in (D + (C-D + x * n)) could be proven to
not wrap, where the choice of D also maximizes the number of trailing
zeroes of (C-D + x * n), ensuring homogeneous behaviour of the
transformation and better canonicalization of such AddRec's
(indeed, there are 2^(2w) different expressions in `B1 + ext(B2 + Y)` form for
the same Y, but only 2^(2w - k) different expressions in the resulting `B3 +
ext((B4 * 2^k) + Y)` form, where w is the bit width of the integral type)
This patch generalizes sext(C1 + C2*X) --> sext(C1) + sext(C2*X) and
sext{C1,+,C2} --> sext(C1) + sext{0,+,C2} transformations added in
r209568 relaxing the requirements the following way:
1. C2 doesn't have to be a power of 2, it's enough if it's divisible by 2
a sufficient number of times;
2. C1 doesn't have to be less than C2, instead of extracting the entire
C1 we can split it into 2 terms: (00...0XXX + YY...Y000), keep the
second one that may cause wrapping within the extension operator, and
move the first one that doesn't affect wrapping out of the extension
operator, enabling further simplifications;
3. C1 and C2 don't have to be positive, splitting C1 like shown above
produces a sum that is guaranteed to not wrap, signed or unsigned;
4. in AddExpr case there could be more than 2 terms, and in case of
AddExpr the 2nd and following terms and in case of AddRecExpr the
Step component don't have to be in the C2*X form or constant
(respectively), they just need to have enough trailing zeros,
which in turn could be guaranteed by means other than arithmetics,
e.g. by a pointer alignment;
5. the extension operator doesn't have to be a sext, the same
transformation works and profitable for zext's as well.
Apparently, optimizations like SLPVectorizer currently fail to
vectorize even rather trivial cases like the following:
double bar(double *a, unsigned n) {
double x = 0.0;
double y = 0.0;
for (unsigned i = 0; i < n; i += 2) {
x += a[i];
y += a[i + 1];
}
return x * y;
}
If compiled with `clang -std=c11 -Wpedantic -Wall -O3 main.c -S -o - -emit-llvm`
(!{!"clang version 7.0.0 (trunk 337339) (llvm/trunk 337344)"})
it produces scalar code with the loop not unrolled with the unsigned `n` and
`i` (like shown above), but vectorized and unrolled loop with signed `n` and
`i`. With the changes made in this commit the unsigned version will be
vectorized (though not unrolled for unclear reasons).
How it all works:
Let say we have an AddExpr that looks like (C + x + y + ...), where C
is a constant and x, y, ... are arbitrary SCEVs. Let's compute the
minimum number of trailing zeroes guaranteed of that sum w/o the
constant term: (x + y + ...). If, for example, those terms look like
follows:
i
XXXX...X000
YYYY...YY00
...
ZZZZ...0000
then the rightmost non-guaranteed-zero bit (a potential one at i-th
position above) can change the bits of the sum to the left (and at
i-th position itself), but it can not possibly change the bits to the
right. So we can compute the number of trailing zeroes by taking a
minimum between the numbers of trailing zeroes of the terms.
Now let's say that our original sum with the constant is effectively
just C + X, where X = x + y + .... Let's also say that we've got 2
guaranteed trailing zeros for X:
j
CCCC...CCCC
XXXX...XX00 // this is X = (x + y + ...)
Any bit of C to the left of j may in the end cause the C + X sum to
wrap, but the rightmost 2 bits of C (at positions j and j - 1) do not
affect wrapping in any way. If the upper bits cause a wrap, it will be
a wrap regardless of the values of the 2 least significant bits of C.
If the upper bits do not cause a wrap, it won't be a wrap regardless
of the values of the 2 bits on the right (again).
So let's split C to 2 constants like follows:
0000...00CC = D
CCCC...CC00 = (C - D)
and represent the whole sum as D + (C - D + X). The second term of
this new sum looks like this:
CCCC...CC00
XXXX...XX00
----------- // let's add them up
YYYY...YY00
The sum above (let's call it Y)) may or may not wrap, we don't know,
so we need to keep it under a sext/zext. Adding D to that sum though
will never wrap, signed or unsigned, if performed on the original bit
width or the extended one, because all that that final add does is
setting the 2 least significant bits of Y to the bits of D:
YYYY...YY00 = Y
0000...00CC = D
----------- <nuw><nsw>
YYYY...YYCC
Which means we can safely move that D out of the sext or zext and
claim that the top-level sum neither sign wraps nor unsigned wraps.
Let's run an example, let's say we're working in i8's and the original
expression (zext's or sext's operand) is 21 + 12x + 8y. So it goes
like this:
0001 0101 // 21
XXXX XX00 // 12x
YYYY Y000 // 8y
0001 0101 // 21
ZZZZ ZZ00 // 12x + 8y
0000 0001 // D
0001 0100 // 21 - D = 20
ZZZZ ZZ00 // 12x + 8y
0000 0001 // D
WWWW WW00 // 21 - D + 12x + 8y = 20 + 12x + 8y
therefore zext(21 + 12x + 8y) = (1 + zext(20 + 12x + 8y))<nuw><nsw>
This approach could be improved if we move away from using trailing
zeroes and use KnownBits instead. For instance, with KnownBits we could
have the following picture:
i
10 1110...0011 // this is C
XX X1XX...XX00 // this is X = (x + y + ...)
Notice that some of the bits of X are known ones, also notice that
known bits of X are interspersed with unknown bits and not grouped on
the rigth or left.
We can see at the position i that C(i) and X(i) are both known ones,
therefore the (i + 1)th carry bit is guaranteed to be 1 regardless of
the bits of C to the right of i. For instance, the C(i - 1) bit only
affects the bits of the sum at positions i - 1 and i, and does not
influence if the sum is going to wrap or not. Therefore we could split
the constant C the following way:
i
00 0010...0011 = D
10 1100...0000 = (C - D)
Let's compute the KnownBits of (C - D) + X:
XX1 1 = carry bit, blanks stand for known zeroes
10 1100...0000 = (C - D)
XX X1XX...XX00 = X
--- -----------
XX X0XX...XX00
Will this add wrap or not essentially depends on bits of X. Adding D
to this sum, however, is guaranteed to not to wrap:
0 X
00 0010...0011 = D
sX X0XX...XX00 = (C - D) + X
--- -----------
sX XXXX XX11
As could be seen above, adding D preserves the sign bit of (C - D) +
X, if any, and has a guaranteed 0 carry out, as expected.
The more bits of (C - D) we constrain, the better the transformations
introduced here canonicalize expressions as it leaves less freedom to
what values the constant part of ((C - D) + x + y + ...) can take.
Reviewed By: mzolotukhin, efriedma
Differential Revision: https://reviews.llvm.org/D48853
llvm-svn: 337943
TTI::getMinMaxReductionCost typically can't handle pointer types - until this is changed its better to limit horizontal reduction to integer/float vector types only.
llvm-svn: 337280
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
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
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
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
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
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
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
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
These were being over cautious for costs for one/two op general shuffles - VSHUFPD doesn't have to replicate the same shuffle in both lanes like VSHUFPS does.
llvm-svn: 335216
D47985 saw the old SK_Alternate 'alternating' shuffle mask replaced with the SK_Select mask which accepts either input operand for each lane, equivalent to a vector select with a constant condition operand.
This patch updates SLPVectorizer to make full use of this SK_Select shuffle pattern by removing the 'isOdd()' limitation.
The AArch64 regression will be fixed by D48172.
Differential Revision: https://reviews.llvm.org/D48174
llvm-svn: 335130
This usually results in better code. Fixes using
inline asm with short2, and also fixes having a different
ABI for function parameters between VI and gfx9.
Partially cleans up the mess used for lowering of the d16
operations. Making v4f16 legal will help clean this up more,
but this requires additional work.
llvm-svn: 332953
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
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
We use getExtractWithExtendCost to calculate the cost of extractelement and
s|zext together when computing the extract cost after vectorization, but we
calculate the cost of extractelement and s|zext separately when computing the
scalar cost which is larger than it should be.
Differential Revision: https://reviews.llvm.org/D45469
llvm-svn: 330143
Summary:
If the load/extractelement/extractvalue instructions are not originally
consecutive, the SLP vectorizer is unable to vectorize them. Patch
allows reordering of such instructions.
Patch does not support reordering of the repeated instruction, this must
be handled in the separate patch.
Reviewers: RKSimon, spatel, hfinkel, mkuper, Ayal, ashahid
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43776
llvm-svn: 329085
We use two approaches for determining the minimum bitwidth.
* Demanded bits
* Value tracking
If demanded bits doesn't result in a narrower type, we then try value tracking.
We need this if we want to root SLP trees with the indices of getelementptr
instructions since all the bits of the indices are demanded.
But there is a missing piece though. We need to be able to distinguish "demanded
and shrinkable" from "demanded and not shrinkable". For example, the bits of %i
in
%i = sext i32 %e1 to i64
%gep = getelementptr inbounds i64, i64* %p, i64 %i
are demanded, but we can shrink %i's type to i32 because it won't change the
result of the getelementptr. On the other hand, in
%tmp15 = sext i32 %tmp14 to i64
%tmp16 = insertvalue { i64, i64 } undef, i64 %tmp15, 0
it doesn't make sense to shrink %tmp15 and we can skip the value tracking.
Ideas are from Matthew Simpson!
Differential Revision: https://reviews.llvm.org/D44868
llvm-svn: 329035
Benjamin Kramer