For processors with low overhead branching (LOB), runtime unrolling the
innermost loop is often detrimental to performance. In these cases the
loop remainder gets unrolled into a series of compare-and-jump blocks,
which in deeply nested loops get executed multiple times, negating the
benefits of LOB.
This is particularly noticable when the loop trip count of the innermost
loop varies within the outer loop, such as in the case of triangular
matrix decompositions.
In these cases we will prefer to not unroll the innermost loop, with the
intention for it to be executed as a low overhead loop.
Add initial heuristics to selectively enable runtime unrolling for loops
where doing so is expected to be highly beneficial on Apple Silicon
CPUs.
To start with, we try to runtime-unroll small, single block loops, if
they have load/store dependencies, to expose more parallel memory
access streams [1] and to improve instruction delivery [2].
We also explicitly avoid runtime-unrolling for loop structures that may
limit the expected gains from runtime unrolling. Such loops include
loops with complex control flow (aren't innermost loops, have multiple
exits, have a large number of blocks), trip count expansion is
expensive and are expected to execute a small number of iterations.
Note that the heuristics here may be overly conservative and we err on
the side of avoiding runtime unrolling rather than unroll excessively.
They are all subject to further refinement.
Across a large set of workloads, this increase the total number of
unrolled loops by 2.9%.
[1] 4.6.10 in Apple Silicon CPU Optimization Guide
[2] 4.4.4 in Apple Silicon CPU Optimization Guide
Depends on https://github.com/llvm/llvm-project/pull/118316 for TTI
changes.
PR: https://github.com/llvm/llvm-project/pull/118317
Add test for existing loop unroll behaviour.
Current behaviour is the single loop with fmul gets runtime unrolled by
count of 4, with the loop remainder unrolled as the 3 for.body9.us.prol
sections. This is quite a lot of compare and branch, negating the
benefits of the low overhead loop mechanism.
This patch fixes a simple error where as part of loop unrolling we
optimize conditional loop-exiting branches into unconditional branches
when we know that they will or won't exit the loop, but does not
propagate the source location of the original branch to the new one.
Found using https://github.com/llvm/llvm-project/pull/107279.
In the test case, the BECount of the second loop uses %load,
but we only have an LCSSA phi node for %add, so that is what
gets invalidated. Use the forgetLcssaPhiWithNewPredecessor()
API instead, which will invalidate the roots of the expression
instead.
Fixes https://github.com/llvm/llvm-project/issues/109333.
The current isHighCostExpansion cost model for addrecs computes the cost
for some kind of polynomial expansion that does not appear to have any
relation to addrec expansion whatsoever.
A literal expansion of an affine addrec is a phi and add (plus the
expansion of start and step). For a non-affine addrec, we get another
phi+add for each additional addrec nested in the step recurrence.
This partially `fixes` https://github.com/llvm/llvm-project/issues/53205
(the runtime unroll test case in this PR).
Use ConstantFoldLoadFromConst() instead of a partial re-implementation.
This makes the code slightly more generic by not depending on the
exact structure of the constant.
It is now translated to `<1 x i64>`, which allows the removal of a bunch
of special casing.
This _incompatibly_ changes the ABI of any LLVM IR function with
`x86_mmx` arguments or returns: instead of passing in mmx registers,
they will now be passed via integer registers. However, the real-world
incompatibility caused by this is expected to be minimal, because Clang
never uses the x86_mmx type -- it lowers `__m64` to either `<1 x i64>`
or `double`, depending on ABI.
This change does _not_ eliminate the SelectionDAG `MVT::x86mmx` type.
That type simply no longer corresponds to an IR type, and is used only
by MMX intrinsics and inline-asm operands.
Because SelectionDAGBuilder only knows how to generate the
operands/results of intrinsics based on the IR type, it thus now
generates the intrinsics with the type MVT::v1i64, instead of
MVT::x86mmx. We need to fix this before the DAG LegalizeTypes, and thus
have the X86 backend fix them up in DAGCombine. (This may be a
short-lived hack, if all the MMX intrinsics can be removed in upcoming
changes.)
Works towards issue #98272.
The use-def walk in forgetValue() was skipping instructions with
non-SCEVable types. However, SCEV may look past with.overflow
intrinsics returning aggregates.
Fixes#97586.
This patch makes the final major change of the RemoveDIs project, changing the
default IR output from debug intrinsics to debug records. This is expected to
break a large number of tests: every single one that tests for uses or
declarations of debug intrinsics and does not explicitly disable writing
records.
If this patch has broken your downstream tests (or upstream tests on a
configuration I wasn't able to run):
1. If you need to immediately unblock a build, pass
`--write-experimental-debuginfo=false` to LLVM's option processing for all
failing tests (remember to use `-mllvm` for clang/flang to forward arguments to
LLVM).
2. For most test failures, the changes are trivial and mechanical, enough that
they can be done by script; see the migration guide for a guide on how to do
this: https://llvm.org/docs/RemoveDIsDebugInfo.html#test-updates
3. If any tests fail for reasons other than FileCheck check lines that need
updating, such as assertion failures, that is most likely a real bug with this
patch and should be reported as such.
For more information, see the recent PSA:
https://discourse.llvm.org/t/psa-ir-output-changing-from-debug-intrinsics-to-debug-records/79578
- There is no restriction on a loop with controlled convergent
operations when
the relevant tokens are defined and used within the loop.
- When a token defined outside a loop is used inside (also called a loop
convergence heart), unrolling is allowed only in the absence of
remainder or
runtime checks.
- When a token defined inside a loop is used outside, such a loop is
said to be
"extended". This loop can only be unrolled by also duplicating the
extended part
lying outside the loop. Such unrolling is disabled for now.
- Clean up loop hearts: When unrolling a loop with a heart, duplicating
the
heart will introduce multiple static uses of a convergence control token
in a
cycle that does not contain its definition. This violates the static
rules for
tokens, and needs to be cleaned up into a single occurrence of the
intrinsic.
- Spell out the initializer for UnrollLoopOptions to improve
readability.
Original implementation [D85605] by Nicolai Haehnle
<nicolai.haehnle@amd.com>.
This patch adds processing of min/max intrinsics in LoopPeel in the
similar way as it was done for conditional statements: for
min/max(IterVal, BoundVal) we peel iterations where IterVal < BoundVal
for monotonically increasing IterVal; for monotonically decreasing
IterVal we peel iterations where IterVal > BoundVal (strict comparision
predicates are used to minimize number of peeled iterations).
The znver3/4 scheduler models have previously associated the LoopMicroOpBufferSize with the maximum size of their op caches, and when this led to quadratic complexity issues this were reduced to a value of 512 uops, based mainly on compilation time and not its effectiveness on runtime performance.
From a runtime performance POV, a large LoopMicroOpBufferSize leads to a higher number of loop unrolls, meaning the cpu has to rely on the frontend decode rate (4 ins/cy max) for much longer to fill the op cache before looping begins and we make use of the faster op cache rate (8/9 ops/cy).
This patch proposes we instead cap the size of the LoopMicroOpBufferSize based off the maximum rate from the op cache (znver3 = 8op/cy, znver4 = 9op/cy) and the branch misprediction penalty from the opcache (~12cy) as a estimate of the useful number of ops we can unroll a loop by before mispredictions are likely to cause stalls. This isn't a perfect metric, but does try to be closer to the spirit of how we use LoopMicroOpBufferSize in the compiler vs the size of a similar naming buffer in the cpu.
This reverts commit f0b3654701bde1cf7821d60698b42383edaff9f3.
This commit triggers UB by reading an uninitialized variable.
`UP.PartialThreshold` is used uninitialized in `getUnrollingPreferences()` when
it is called from `LoopVectorizationPlanner::executePlan()`. In this case the
`UP` variable is created on the stack and its fields are not initialized.
```
==8802==WARNING: MemorySanitizer: use-of-uninitialized-value
#0 0x557c0b081b99 in llvm::BasicTTIImplBase<llvm::X86TTIImpl>::getUnrollingPreferences(llvm::Loop*, llvm::ScalarEvolution&, llvm::TargetTransformInfo::UnrollingPreferences&, llvm::OptimizationRemarkEmitter*) llvm-project/llvm/include/llvm/CodeGen/BasicTTIImpl.h
#1 0x557c0b07a40c in llvm::TargetTransformInfo::Model<llvm::X86TTIImpl>::getUnrollingPreferences(llvm::Loop*, llvm::ScalarEvolution&, llvm::TargetTransformInfo::UnrollingPreferences&, llvm::OptimizationRemarkEmitter*) llvm-project/llvm/include/llvm/Analysis/TargetTransformInfo.h:2277:17
#2 0x557c0f5d69ee in llvm::TargetTransformInfo::getUnrollingPreferences(llvm::Loop*, llvm::ScalarEvolution&, llvm::TargetTransformInfo::UnrollingPreferences&, llvm::OptimizationRemarkEmitter*) const llvm-project/llvm/lib/Analysis/TargetTransformInfo.cpp:387:19
#3 0x557c0e6b96a0 in llvm::LoopVectorizationPlanner::executePlan(llvm::ElementCount, unsigned int, llvm::VPlan&, llvm::InnerLoopVectorizer&, llvm::DominatorTree*, bool, llvm::DenseMap<llvm::SCEV const*, llvm::Value*, llvm::DenseMapInfo<llvm::SCEV const*, void>, llvm::detail::DenseMapPair<llvm::SCEV const*, llvm::Value*>> const*) llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:7624:7
#4 0x557c0e6e4b63 in llvm::LoopVectorizePass::processLoop(llvm::Loop*) llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:10253:13
#5 0x557c0e6f2429 in llvm::LoopVectorizePass::runImpl(llvm::Function&, llvm::ScalarEvolution&, llvm::LoopInfo&, llvm::TargetTransformInfo&, llvm::DominatorTree&, llvm::BlockFrequencyInfo*, llvm::TargetLibraryInfo*, llvm::DemandedBits&, llvm::AssumptionCache&, llvm::LoopAccessInfoManager&, llvm::OptimizationRemarkEmitter&, llvm::ProfileSummaryInfo*) llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:10344:30
#6 0x557c0e6f2f97 in llvm::LoopVectorizePass::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:10383:9
[...]
Uninitialized value was created by an allocation of 'UP' in the stack frame
#0 0x557c0e6b961e in llvm::LoopVectorizationPlanner::executePlan(llvm::ElementCount, unsigned int, llvm::VPlan&, llvm::InnerLoopVectorizer&, llvm::DominatorTree*, bool, llvm::DenseMap<llvm::SCEV const*, llvm::Value*, llvm::DenseMapInfo<llvm::SCEV const*, void>, llvm::detail::DenseMapPair<llvm::SCEV const*, llvm::Value*>> const*) llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:7623:3
```
The znver3/znver4 scheduler models are outliers, specifying very large
LoopMicroOpBufferSizes at 512, while typical values for other subtargets
are on the order of ~50. Even if this information is
micro-architecturally correct (*), this does not mean that we want to
runtime unroll all loops to a size that completely fills the loop
buffer. Unless this is the single hot loop in the entire application,
the massive code size increase will bust the micro-op and instruction
caches.
Protect against this by clamping to the default PartialThreshold of 150,
which is the same as the default full-unroll threshold and half the
aggressive full-unroll threshold. Allowing more partial unrolling than
full unrolling certainly does not make sense.
(*) I strongly doubt that this is actually correct -- I believe this may
derive from an incorrect reading of Agner Fog's micro-architecture
guide. The number 4096 that was originally used here is the size of the
general micro-op cache, not that of a loop buffer. A separate loop
buffer is not listed for the Zen microarchitecture. Comparing this to
the listing for Skylake, it has a 1536 micro-op buffer, but only a 64
micro-op loopback buffer, with a note that it's rarely fully utilized.
Our scheduling model specifies LoopMicroOpBufferSize of 50 in that case.
Remove redundant debug instructions after blocks have been merged into
the predecessor, It can reduce some compile time in some cases.
This change only fixes the situation of loop unrolling, and other
situations are not considered. "RemoveRedundantDbgInstrs" seems to be
very time-consuming. Thus, we just add here after the "Dest" has been
merged into the "Fold", this may be a more targeted solution!!!
fixes: https://github.com/llvm/llvm-project/issues/89073
This reverts commit 2e3e0868748635b779ba89a772eae3664bd822e4. It caused
quadratic slowdown at compilation time in some cases. See the comments
in the original PR: https://github.com/llvm/llvm-project/pull/89069
This patch adds loadCSE support to simplifyLoopAfterUnroll. It is based
on EarlyCSE's implementation using ScopeHashTable and is using SCEV for
accessed pointers to check to find redundant loads after unrolling.
This applies to the late unroll pass only, for full unrolling those
redundant loads will be cleaned up by the regular pipeline.
The current approach constructs MSSA on-demand per-loop, but there is
still small but notable compile-time impact:
stage1-O3 +0.04%
stage1-ReleaseThinLTO +0.06%
stage1-ReleaseLTO-g +0.05%
stage1-O0-g +0.02%
stage2-O3 +0.09%
stage2-O0-g +0.04%
stage2-clang +0.02%
https://llvm-compile-time-tracker.com/compare.php?from=c089fa5a729e217d0c0d4647656386dac1a1b135&to=ec7c0f27cb5c12b600d9adfc8543d131765ec7be&stat=instructions:u
This benefits some workloads with runtime-unrolling disabled,
where users use pragmas to force unrolling, as well as with
runtime unrolling enabled.
On SPEC/MultiSource, this removes a number of loads after unrolling
on AArch64 with runtime unrolling enabled.
```
External/S...te/526.blender_r/526.blender_r 96
MultiSourc...rks/mediabench/gsm/toast/toast 39
SingleSource/Benchmarks/Misc/ffbench 4
External/SPEC/CINT2006/403.gcc/403.gcc 18
MultiSourc.../Applications/JM/ldecod/ldecod 4
MultiSourc.../mediabench/jpeg/jpeg-6a/cjpeg 6
MultiSourc...OE-ProxyApps-C/miniGMG/miniGMG 9
MultiSourc...e/Applications/ClamAV/clamscan 4
MultiSourc.../MallocBench/espresso/espresso 3
MultiSourc...dence-flt/LinearDependence-flt 2
MultiSourc...ch/office-ispell/office-ispell 4
MultiSourc...ch/consumer-jpeg/consumer-jpeg 6
MultiSourc...ench/security-sha/security-sha 11
MultiSourc...chmarks/McCat/04-bisect/bisect 3
SingleSour...tTests/2020-01-06-coverage-009 12
MultiSourc...ench/telecomm-gsm/telecomm-gsm 39
MultiSourc...lds-flt/CrossingThresholds-flt 24
MultiSourc...dence-dbl/LinearDependence-dbl 2
External/S...C/CINT2006/445.gobmk/445.gobmk 6
MultiSourc...enchmarks/mafft/pairlocalalign 53
External/S...31.deepsjeng_r/531.deepsjeng_r 3
External/S...rate/510.parest_r/510.parest_r 58
External/S...NT2006/464.h264ref/464.h264ref 29
External/S...NT2017rate/502.gcc_r/502.gcc_r 45
External/S...C/CINT2006/456.hmmer/456.hmmer 6
External/S...te/538.imagick_r/538.imagick_r 18
External/S.../CFP2006/447.dealII/447.dealII 4
MultiSourc...OE-ProxyApps-C++/miniFE/miniFE 12
External/S...2017rate/525.x264_r/525.x264_r 36
MultiSourc...Benchmarks/7zip/7zip-benchmark 33
MultiSourc...hmarks/ASC_Sequoia/AMGmk/AMGmk 2
MultiSourc...chmarks/VersaBench/8b10b/8b10b 1
MultiSourc.../Applications/JM/lencod/lencod 116
MultiSourc...lds-dbl/CrossingThresholds-dbl 24
MultiSource/Benchmarks/McCat/05-eks/eks 15
```
PR: https://github.com/llvm/llvm-project/pull/83860
Currently we model subvector inserts and extracts as shuffles,
potentially going as far as scalarizing. If the types are legal then
they can just be simple zip/unzip operations, or possible even no-ops.
Change the cost to a relatively small one to ensure that simple loops
featuring such operations between fixed and scalable vector types that
are effectively the same at a given sve width can be unrolled and
further optimized.
In LoopUnroll, peelLoop is called on the loop. After the loop is peeled
it calls simplifyLoopAfterUnroll on the loop. This call to
simplifyLoopAfterUnroll doesn't preserve the LCSSA form of the parent
loop and thus during the next call to peelLoop the LCSSA form is already
broken.
LoopPeel util takes in the PreserveLCSSA argument and it passes
on the same argument to simplifyLoop which checks if the loop is in a
valid LCSSA form, when (PreserveLCSSA = true).
This causes an assert in simplifyLoop when (PreserveLCSSA = true), as
during the last call LCSSA for the loop wasn't preserved, and thus
crashes at the following assert.
assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&
"Requested to preserve LCSSA, but it's already broken.");
Upon debugging, it is evident that simplifyLoopIVs call inside
simplifyLoopAfterUnroll breaks the LCSSA form. This patch fixes
llvm#77118, it checks if the replacement of IV Users with Loop Invariant
preserves the LCSSA form. If it does not, it emits the required LCSSA
Phi instructions.
Get more precise cost of instruction after LoopUnroll considering that
some operands of it can be simplified, e.g. induction variable will be
replaced by constant after full unrolling.
Fixes [PR77842](https://github.com/llvm/llvm-project/issues/77842) where
UBSAN causes pragma full unroll to try and unroll INT_MAX times. This
sets a cap to make sure we don't attempt this and crash the compiler.
Testing:
ninja check-all with new test
---------
Co-authored-by: Nikita Popov <github@npopov.com>
This patch canonicalizes getelementptr instructions with constant
indices to use the `i8` source element type. This makes it easier for
optimizations to recognize that two GEPs are identical, because they
don't need to see past many different ways to express the same offset.
This is a first step towards
https://discourse.llvm.org/t/rfc-replacing-getelementptr-with-ptradd/68699.
This is limited to constant GEPs only for now, as they have a clear
canonical form, while we're not yet sure how exactly to deal with
variable indices.
The test llvm/test/Transforms/PhaseOrdering/switch_with_geps.ll gives
two representative examples of the kind of optimization improvement we
expect from this change. In the first test SimplifyCFG can now realize
that all switch branches are actually the same. In the second test it
can convert it into simple arithmetic. These are representative of
common optimization failures we see in Rust.
Fixes https://github.com/llvm/llvm-project/issues/69841.
Fix loop unroll fail caused by branches folding.
For example:
SimplifyCFG foldloop branches then cause loop unroll failed for "#program unroll" loop.
```
#program unroll
for (int I = 0; I < ConstNum; ++I) { // folding "I < ConstNum" and "Cond2"
if (Cond2) {
break;
}
xxx loop body;
}
```
The pragma unroll metadata only takes effect if there is an exact trip
count, but not if there is an upper bound trip count. This patch make it
work with an upper bound trip count as well in shouldPragmaUnroll().
Loop unroll is important in stack nervous devices (e.g. GPU, and that is
why a lot of GPU code mark loop with "#program unroll").
It usually much simplify the address (offset) calculations in old
iterations, then we can do a lot of others optimizations, e.g, SROA, for
these simplifed address (escape alloca the whole aggregates).
LSR uses SCEVExpander to generate induction formulas. The expander
internally tries to reuse existing IR expressions. To do that, it needs
to strip any poison generating flags (nsw, nuw, exact, nneg, etc..)
which may not be valid for the newly added users.
This is conservatively correct, but has the effect that LSR will strip
nneg flags on zext instructions involved in trip counts in loop
preheaders. To avoid this, this patch adjusts the expanded to reinfer
the flags on the CSE candidate if legal for all possible users.
This should fix the regression reported in
https://github.com/llvm/llvm-project/issues/71200.
This should arguably be done inside canReuseInstruction instead, but
doing it outside is more conservative compile time wise. Both
canReuseInstruction and isGuaranteedNotToBePoison walk operand lists, so
right now we are performing work which is roughly O(N^2) in the size of
the operand graph. We should fix that before making the per operand step
more expensive. My tenative plan is to land this, and then rework the
code to sink the logic into more core interfaces.
These tests rely on SCEV looking recognizing an "or" with no common
bits as an "add". Add the disjoint flag to relevant or instructions
in preparation for switching SCEV to use the flag instead of the
ValueTracking query. The IR with disjoint flag matches what
InstCombine would produce.
For example, this allows us to peel this loop with a `and`:
```
for (int i = 0; i < N; ++i) {
if (i % 2 == 0 && i < 3) // can peel based on || as well
f1();
f2();
```
into:
```
for (int i = 0; i < 3; ++i) { // peel three iterations
if (i % 2 == 0)
f1();
f2();
}
for (int i = 3; i < N; ++i)
f2();
```
