This patch adds metadata to disable runtime unrolling to the vectorized
loop. If runtime unrolling/interleaving is considered profitable, LV
will interleave the loop directly. There should be no need to perform
runtime unrolling at a later stage.
Note that we already add metadata to disable runtime unrolling to the
scalar loop after vectorization.
The additional unrolling unnecessarily increases code size and compile
time. In addition to that we have several bug reports of unncessary
runtime unrolling for vectorized loops, e.g. PR40961
Compile-time improvements:
NewPM-O3: -1.04%
NewPM-ReleaseThinLTO: -0.59%
NewPM-ReleaseLTO-g: -0.97%
https://llvm-compile-time-tracker.com/compare.php?from=ce1be13a868d0f8afa367975558c1a6175cce33a&to=78bc2e67f22e9e10e61cdb6cdac4bb857d95eb1b&stat=instructions:uFixes#40306.
Reviewed By: lebedev.ri, nikic
Differential Revision: https://reviews.llvm.org/D115261
We call tail-call-elim near the beginning of the pipeline,
but that is too early to annotate calls that get added later.
In the motivating case from issue #47852, the missing 'tail'
on memset leads to sub-optimal codegen.
I experimented with removing the early instance of
tail-call-elim instead of just adding another pass, but that
appears to be slightly worse for compile-time:
+0.15% vs. +0.08% time.
"tailcall" shows adding the pass; "tailcall2" shows moving
the pass to later, then adding the original early pass back
(so 1596886802 is functionally equivalent to 180b0439dc ):
https://llvm-compile-time-tracker.com/index.php?config=NewPM-O3&stat=instructions&remote=rotateright
Note that there was an effort to split the tail call functionality
into 2 passes - that could help reduce compile-time if we find
that this change costs more in compile-time than expected based
on the preliminary testing:
D60031
Differential Revision: https://reviews.llvm.org/D130374
Now that SimpleLoopUnswitch and other transforms no longer introduce
branch on poison, enable the -branch-on-poison-as-ub option by
default. The practical impact of this is mostly better flag
preservation in SCEV, and some freeze instructions no longer being
necessary.
Differential Revision: https://reviews.llvm.org/D125299
When computing the BECount for multi-exit loops, we need to combine
individual exit counts using umin_seq rather than umin. This is
because an earlier exit may exit on the first iteration, in which
case later exit expressions will not be evaluated and could be
poisonous. We cannot propagate potential poison values from later
exits.
In particular, this avoids the introduction of "branch on poison"
UB when optimizing multi-exit loops.
Differential Revision: https://reviews.llvm.org/D124910
This patch adds initial support for a pointer diff based runtime check
scheme for vectorization. This scheme requires fewer computations and
checks than the existing full overlap checking, if it is applicable.
The main idea is to only check if source and sink of a dependency are
far enough apart so the accesses won't overlap in the vector loop. To do
so, it is sufficient to compute the difference and compare it to the
`VF * UF * AccessSize`. It is sufficient to check
`(Sink - Src) <u VF * UF * AccessSize` to rule out a backwards
dependence in the vector loop with the given VF and UF. If Src >=u Sink,
there is not dependence preventing vectorization, hence the overflow
should not matter and using the ULT should be sufficient.
Note that the initial version is restricted in multiple ways:
1. Pointers must only either be read or written, by a single
instruction (this allows re-constructing source/sink for
dependences with the available information)
2. Source and sink pointers must be add-recs, with matching steps
3. The step must be a constant.
3. abs(step) == AccessSize.
Most of those restrictions can be relaxed in the future.
See https://github.com/llvm/llvm-project/issues/53590.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D119078
Given a shuffle with 4 elements size 16 or 32, we can use the costs
directly from the PerfectShuffle tables to get a slightly more accurate
cost for the resulting shuffle.
Differential Revision: https://reviews.llvm.org/D123409
LICM will speculatively hoist code outside of loops. This requires removing information, like alias analysis (https://github.com/llvm/llvm-project/issues/53794), range information (https://bugs.llvm.org/show_bug.cgi?id=50550), among others. Prior to https://reviews.llvm.org/D99249 , LICM would only be run after LoopRotate. Running Loop Rotate prior to LICM prevents a instruction hoist from being speculative, if it was conditionally executed by the iteration (as is commonly emitted by clang and other frontends). Adding the additional LICM pass first, however, forces all of these instructions to be considered speculative, even if they are not speculative after LoopRotate. This destroys information, resulting in performance losses for discarding this additional information.
This PR modifies LICM to accept a ``speculative'' parameter which allows LICM to be set to perform information-loss speculative hoists or not. Phase ordering is then modified to not perform the information-losing speculative hoists until after loop rotate is performed, preserving this additional information.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D119965
In D110057 we moved LoopFlatten to a LoopPassManager. This caused a performance
regression for our 64-bit targets (the 32-bit were unaffected), the pass is no
longer triggering for a motivating example. The reason is that the IR is just
very different than expected; we try to match loop statements and particular
uses of induction variables. The easiest is to just move LoopFlatten to a place
in the pipeline where the IR is as expected, which is just before
IndVarSimplify. This means we move it from LPM2 to LPM1, so that it actually
runs just a bit earlier from where it was running before. IndVarSimplify is
responsible for significant rewrites that are difficult to "look through" in
LoopFlatten.
Differential Revision: https://reviews.llvm.org/D116612
The basic idea to this is that a) having a single canonical type makes CSE easier, and b) many of our transforms are inconsistent about which types we end up with based on visit order.
I'm restricting this to constants as for non-constants, we'd have to decide whether the simplicity was worth extra instructions. For constants, there are no extra instructions.
We chose the canonical type as i64 arbitrarily. We might consider changing this to something else in the future if we have cause.
Differential Revision: https://reviews.llvm.org/D115387
This patch fixes PR52111. The problem is that LV propagates poison-generating flags (`nuw`/`nsw`, `exact`
and `inbounds`) in instructions that contribute to the address computation of widen loads/stores that are
guarded by a condition. It may happen that when the code is vectorized and the control flow within the loop
is linearized, these flags may lead to generating a poison value that is effectively used as the base address
of the widen load/store. The fix drops all the integer poison-generating flags from instructions that
contribute to the address computation of a widen load/store whose original instruction was in a basic block
that needed predication and is not predicated after vectorization.
Reviewed By: fhahn, spatel, nlopes
Differential Revision: https://reviews.llvm.org/D111846
The unrolling code was previously inserting new cloned blocks at the end of the function. The result of this with typical loop structures is that the new iterations are placed far from the initial iteration.
With unrolling, the general assumption is that the a) the loop is reasonable hot, and b) the first Count-1 copies of the loop are rarely (if ever) loop exiting. As such, placing Count-1 copies out of line is a fairly poor code placement choice. We'd much rather fall through into the hot (non-exiting) path. For code with branch profiles, later layout would fix this, but this may have a positive impact on non-PGO compiled code.
However, the real motivation for this change isn't performance. Its readability and human understanding. Having to jump around long distances in an IR file to trace an unrolled loop structure is error prone and tedious.
This patch adds a pass option to only run transforms that scalarize
vector operations and do not create new vector instructions.
When running VectorCombine early in the pipeline introducing new vector
operations can have negative effects, like blocking loop or SLP
vectorization. To avoid regressions, restrict the early VectorCombine
run (when using -enable-matrix) to only perform scalarization and not
introduce new vector operations.
This is done as option to the pass directly, which is then set when
adding the pass to the pipeline. This is done for the new pass manager
only.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D111800
Running -vector-combine early can introduce new vector operations,
blocking loop/SLP vectorization. The added test case could be better
optimized by the SLPVectorizer if no new vector operations are added
early.
This patch adds a new cost heuristic that allows peeling a single
iteration off read-only loops, if the loop contains a load that
1. is feeding an exit condition,
2. dominates the latch,
3. is not already known to be dereferenceable,
4. and has a loop invariant address.
If all non-latch exits are terminated with unreachable, such loads
in the loop are guaranteed to be dereferenceable after peeling,
enabling hoisting/CSE'ing them.
This enables vectorization of loops with certain runtime-checks, like
multiple calls to `std::vector::at` if the vector is passed as pointer.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D108114
IR with matrix intrinsics is likely to also contain large vector
operations, which can benefit from early simplifications.
This is the last step in a series of changes to improve code-gen for
code using matrix subscript operators with the C/C++ matrix extension in
CLang, like
using matrix_t = double __attribute__((matrix_type(15, 15)));
void foo(unsigned i, matrix_t &A, matrix_t &B) {
for (unsigned j = 0; j < 4; ++j)
for (unsigned k = 0; k < i; k++)
B[k][j] -= A[k][j] * B[i][j];
}
https://clang.godbolt.org/z/6dKxK1Ed7
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D102496
This patch updates VectorCombine to use a worklist to allow iterative
simplifications where a combine enables other combines.
Suggested in D100302.
The main use case at the moment is foldSingleElementStore and
scalarizeLoadExtract working together to improve scalarization.
Note that we now also do not run SimplifyInstructionsInBlock on the
whole function if there have been changes. This means we fail to
remove/simplify instructions not related to any of the vector combines.
IMO this is fine, as simplifying the whole function seems more like a
workaround for not tracking the changed instructions.
Compile-time impact looks neutral:
NewPM-O3: +0.02%
NewPM-ReleaseThinLTO: -0.00%
NewPM-ReleaseLTO-g: -0.02%
http://llvm-compile-time-tracker.com/compare.php?from=52832cd917af00e2b9c6a9d1476ba79754dcabff&to=e66520a4637290550a945d528e3e59573485dd40&stat=instructions
Reviewed By: spatel, lebedev.ri
Differential Revision: https://reviews.llvm.org/D110171
This test illustrates missed vectorization of loops with multiple
std::vector::at calls, like
int sum(std::vector<int> *A, std::vector<int> *B, int N) {
int cost = 0;
for (int i = 0; i < N; ++i)
cost += A->at(i) + B->at(i);
return cost;
}
https://clang.godbolt.org/z/KbYoaPhvq
Addition of this pass has been botched.
There is no particular reason why it had to be sold as an inseparable part
of new-pm transition. It was added when old-pm was still the default,
and very *very* few users were actually tracking new-pm,
so it's effects weren't measured.
Which means, some of the turnoil of the new-pm transition
are actually likely regressions due to this pass.
Likewise, there has been a number of post-commit feedback
(post new-pm switch), namely
* https://reviews.llvm.org/D37467#2787157 (regresses HW-loops)
* https://reviews.llvm.org/D37467#2787259 (should not be in middle-end, should run after LSR, not before)
* https://reviews.llvm.org/D95789 (an attempt to fix bad loop backedge metadata)
and in the half year past, the pass authors (google) still haven't found time to respond to any of that.
Hereby it is proposed to backout the pass from the pipeline,
until someone who cares about it can address the issues reported,
and properly start the process of adding a new pass into the pipeline,
with proper performance evaluation.
Furthermore, neither google nor facebook reports any perf changes
from this change, so i'm dropping the pass completely.
It can always be re-reverted should/if anyone want to pick it up again.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D104099
We can only scalarize memory accesses if we know the index is valid.
This patch adjusts canScalarizeAcceess to fall back to
computeConstantRange to check if the index is known to be valid.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D102476
This patch adjusts the LTO pipeline in the new PM to run GlobalsAA
before LICM to match the legacy PM.
This fixes a regression where the new PM failed to vectorize loops that
require hoisting/sinking by LICM depending on GlobalsAA info.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D102345
Hoisting and sinking instructions out of conditional blocks enables
additional vectorization by:
1. Executing memory accesses unconditionally.
2. Reducing the number of instructions that need predication.
After disabling early hoisting / sinking, we miss out on a few
vectorization opportunities. One of those is causing a ~10% performance
regression in one of the Geekbench benchmarks on AArch64.
This patch tires to recover the regression by running hoisting/sinking
as part of a SimplifyCFG run after LoopRotate and before LoopVectorize.
Note that in the legacy pass-manager, we run LoopRotate just before
vectorization again and there's no SimplifyCFG run in between, so the
sinking/hoisting may impact the later run on LoopRotate. But the impact
should be limited and the benefit of hosting/sinking at this stage
should outweigh the risk of not rotating.
Compile-time impact looks slightly positive for most cases.
http://llvm-compile-time-tracker.com/compare.php?from=2ea7fb7b1c045a7d60fcccf3df3ebb26aa3699e5&to=e58b4a763c691da651f25996aad619cb3d946faf&stat=instructions
NewPM-O3: geomean -0.19%
NewPM-ReleaseThinLTO: geoman -0.54%
NewPM-ReleaseLTO-g: geomean -0.03%
With a few benchmarks seeing a notable increase, but also some
improvements.
Alternative to D101290.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D101468
As a side-effect of the change to default HoistCommonInsts to false
early in the pipeline, we fail to convert conditional branch & phis to
selects early on, which prevents vectorization for loops that contain
conditional branches that effectively are selects (or if the loop gets
vectorized, it will get vectorized very inefficiently).
This patch updates SimplifyCFG to perform hoisting if the only
instruction in both BBs is an equal branch. In this case, the only
additional instructions are selects for phis, which should be cheap.
Even though we perform hoisting, the benefits of this kind of hoisting
should by far outweigh the negatives.
For example, the loop in the code below will not get vectorized on
AArch64 with the current default, but will with the patch. This is a
fundamental pattern we should definitely vectorize. Besides that, I
think the select variants should be easier to use for reasoning across
other passes as well.
https://clang.godbolt.org/z/sbjd8Wshx
```
double clamp(double v) {
if (v < 0.0)
return 0.0;
if (v > 6.0)
return 6.0;
return v;
}
void loop(double* X, double *Y) {
for (unsigned i = 0; i < 20000; i++) {
X[i] = clamp(Y[i]);
}
}
```
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D100329
