natively supported on Neon and SVE
PR #158641 refined and refactored the cost model for partial reductions.
While doing so, it missed out on certain constraints. Specifically,
cases like i32 -> i64 partial reduce are not natively supported. This
patch adds back the condition/constraint that was present before PR
#158641
Recipes in replicate regions implicitly depend on the region's
predicate. Limit CSE to recipes in the same block, when either recipe is
in a replicate region.
This allows handling VPPredInstPHIRecipe during CSE. If we perform CSE
on recipes inside a replicate region, we may end up with 2
VPPredInstPHIRecipes sharing the same operand. This is incompatible with
current VPPredInstPHIRecipe codegen, which re-sets the current value of
its operand in VPTransformState. This can cause crashes in the added
test cases.
Note that this patch only modifies ::isEqual to check for replicating
regions and not getHash, as CSE across replicating regions should be
uncommon.
Fixes https://github.com/llvm/llvm-project/issues/157314.
Fixes https://github.com/llvm/llvm-project/issues/161974.
PR: https://github.com/llvm/llvm-project/pull/162110
createWidenCast doesn't set the flag type, so when we simplify trunc
(zext nneg x) -> zext x we would hit an assertion in CSE that the flag
types don't match with other VPWidenCastRecipes that weren't simplified.
This fixes it the same way trunc flags are handled too.
As an aside I think it should be correct to preserve the nneg flag in
this case since the input operand is still non-negative after the
transform. But that's left to another PR.
Fixes https://github.com/llvm/llvm-project/issues/164171
When the legacy cost model scalarizes loads that are used as addresses
for other loads and stores, it looks to phi nodes, if they are direct
address operands of loads/stores. Match this behavior in
isUsedByLoadStoreAddress, to fix a divergence between legacy and
VPlan-based cost model.
The `masked.load`, `masked.store`, `masked.gather` and `masked.scatter`
intrinsics currently accept a separate alignment immarg. Replace this
with an `align` attribute on the pointer / vector of pointers argument.
This is the standard representation for alignment information on
intrinsics, and is already used by all other memory intrinsics. This
means the signatures now match llvm.expandload, llvm.vp.load, etc.
(Things like llvm.memcpy used to have a separate alignment argument as
well, but were already migrated a long time ago.)
It's worth noting that the masked.gather and masked.scatter intrinsics
previously accepted a zero alignment to indicate the ABI type alignment
of the element type. This special case is gone now: If the align
attribute is omitted, the implied alignment is 1, as usual. If ABI
alignment is desired, it needs to be explicitly emitted (which the
IRBuilder API already requires anyway).
Split off from PR #163525, this standalone patch replaces `ret * undef`
returns with `ret void` in order to reduce the likelihood of
contributors hitting the `undef deprecator` warning in github.
When narrowing stores of a single-scalar, we currently use
ExtractLastElement, which extracts the last element across all parts.
This is not correct if the store's address is not uniform across all
parts. If it is only uniform-per-part, the last lane per part must be
extracted. Add a new ExtractLastLanePerPart opcode to handle this
correctly. Most transforms apply to both ExtractLastElement and
ExtractLastLanePerPart, with the only difference being their treatment
during unrolling.
Fixes https://github.com/llvm/llvm-project/issues/162498.
PR: https://github.com/llvm/llvm-project/pull/163056
We have seen performance regression for several instances of the Numba
benchmark, with some ranging around 70%, on Neoverse-v2 post #158641.
The mentioned case is short reproducer of the same. See
https://godbolt.org/z/j9Mj5WM7c for the IR differences.. A future patch
will address this.
Add test coverage for min/max reductions with various combinations of
users (in and outside loops, used by stores) and predicated variants.
This adds missing test coverage for min/max reductions.
Currently we cannot vectorize loops with latch blocks terminated by a
switch. In the future this could be handled by materializing appropriate
compares.
Fixes https://github.com/llvm/llvm-project/issues/156894.
VPWidenCastRecipes with Trunc opcodes where missing the correct OpType
for IR flags. Update createWidenCast to set the correct flags for
truncs, and use it consistenly.
Fixes https://github.com/llvm/llvm-project/issues/162374.
Add extra test coverage for narrowing stores to single scalars, with the
store address being uniform-per-part, not uniform-across-all-parts.
Test for https://github.com/llvm/llvm-project/issues/162498.
Replication is currently not supported for scalable VFs. Make sure
VPReplicateRecipe::computeCost returns an invalid cost early, for
scalable VFs if the recipe is not a single-scalar.
Note that this moves the existing invalid-costs.ll out of the AArch64
subdirectory, as it does not use a target triple.
Fixes https://github.com/llvm/llvm-project/issues/160792.
Consider the following transform:
```
C = binop float A, nnan OOp
D = select ninf, i1 cond, float C, float A
->
E = select ninf, i1 cond, float OOp, float Identity
F = binop float A, E
```
We cannot propagate ninf from the original select, because OOp may be
inf, and the flag only guarantees that FalseVal (op OOp) is never
infinity.
Examples: -inf + +inf = NaN, -inf - -inf = NaN, 0 * inf = NaN
Specifically, if the original select has both ninf and nnan, we can
safely propagate the flag.
Alive2:
+ fadd: https://alive2.llvm.org/ce/z/TWfktv
+ fsub: https://alive2.llvm.org/ce/z/RAsjJb
+ fmul: https://alive2.llvm.org/ce/z/8eg4ND
Closes https://github.com/llvm/llvm-project/issues/161634.
Consistently scalarize loads used as part of address computations across
all uses in the loop. This aligns the VPlan and legacy cost model and
fixes a divergence crash. It doesn't matter if the load and address
users are in different blocks, as long as they are in the same loop, the
scalar value can be used. This removes a number of insert/extracts.
The vector intrinsics in question have no undefined behavior, and have
no other effect besides returning the result: they should hence be
marked speculatable.
VPBlendRecipes are introduced as part of if-conversion, potentially adding
a def-use chain from a load used in a compare to another load/store. In
the scalar IR, there is no connection via def-use chains, so the legacy
cost model won't consider the load used by memory operation.
Skipping blends brings the VPlan-based cost-computation in line with the
legacy cost model after https://github.com/llvm/llvm-project/pull/162157.
VPExpandSCEVRecipes must be at the beginning of the entry block.
addMinimumEpilogueIterationCheck currently creates VPInstructions to
compute the remaining iterations before potentially creating
VPExpandSCEVRecipes.
Fix this by first creating any SCEV expansions if needed.
Fixes https://github.com/llvm/llvm-project/issues/162128.
The original loop (OL) that serves as input to LoopUnroll has basic
blocks that are arranged as follows:
```
OLPreHeader
OLHeader <-.
... |
OLLatch ---'
OLExit
```
In this depiction, every block has an implicit edge to the next block
below, so any explicit edge indicates a conditional branch.
Given OL and unroll count N, LoopUnroll sometimes creates an unrolled
loop (UL) with a remainder loop (RL) epilogue arranged like this:
```
,-- ULGuard
| ULPreHeader
| ULHeader <-.
| ... |
| ULLatch ---'
| ULExit
`-> RLGuard -----.
RLPreHeader |
,-> RLHeader |
| ... |
`-- RLLatch |
RLExit |
OLExit <-----'
```
Each UL iteration executes N OL iterations, but each RL iteration
executes 1 OL iteration. ULGuard or RLGuard checks whether the first
iteration of UL or RL should execute, respectively. If so, ULLatch or
RLLatch checks whether to execute each subsequent iteration.
Once reached, OL always executes its first iteration but not necessarily
the next N-1 iterations. Thus, ULGuard is always required before the
first UL iteration. However, when control flows from ULGuard directly to
RLGuard, the first OL iteration has yet to execute, so RLGuard is then
redundant before the first RL iteration.
Thus, this patch makes the following changes:
- Adjust ULGuard to branch to RLPreHeader instead of RLGuard, thus
eliminating RLGuard's unnecessary branch instruction for that path.
- Eliminate the creation of RLGuard phi node poison values. Without this
patch, RLGuard has such a phi node for each value that is defined by any
OL iteration and used in OLExit. The poison value is required where
ULGuard is the predecessor. The poison value indicates that control flow
from ULGuard to RLGuard to Exit has no counterpart in OL because the
first OL iteration must execute either in UL or RL.
- Simplify the CFG by not splitting ULExit and RLGuard because, without
the ULGuard predecessor, the single block can now be a dedicated UL
exit.
- To RLPreHeader, add an `llvm.assume` call that asserts the RL trip
count is non-zero. Without this patch, RLPreHeader is reachable only
when RLGuard guarantees that assertion is true. With this patch, RLGuard
guarantees it only when RLGuard is the predecessor, and the OL structure
guarantees it when ULGuard is the predecessor. If RL itself is unrolled
later, this guarantee somehow prevents ScalarEvolution from giving up
when trying to compute a maximum trip count for RL. That maximum trip
count enables the branch instruction in the final unrolled instance of
RLLatch to be eliminated. Without the `llvm.assume` call, some existing
unroll tests start to fail because that instruction is not eliminated.
The original motivation for this patch is to facilitate later patches
that fix LoopUnroll's computation of branch weights so that they
maintain the block frequency of OL's body (see #135812). Specifically,
this patch ensures RLGuard's branch weights do not affect RL's
contribution to the block frequency of OL's body in the case that
ULGuard skips UL.
When using information from dereferenceable assumptions, we need to make
sure that the memory is not freed between the assume and the specified
context instruction. Instead of just checking canBeFreed, check if there
any calls that may free between the assume and the context instruction.
Note that this also adjusts the context instruction to be the terminator
in the loop predecessor, if there is one and it is a branch (to avoid
things like invoke).
PR: https://github.com/llvm/llvm-project/pull/161255
Currently there's a crash when trying to construct VPExpressionRecipes
for a mul (ext, ext), if the multiply has outside users; the mul will be
cloned to serve its external users, but the extends won't get cloned and
will stay connected to users outside the loop (the cloned multiply).
To fix this, process recipes in reverse order. This ensures that we
visit bundled users before their operands, properly ensuring that the
extends for the external user are cloned as well.
This reverts commit f80c0baf058dbdc5 and 94eade61a02ae5.
Recommit a small fix for targets using prefersVectorizedAddressing.
Original message:
Update VPReplicateRecipe::computeCost to compute costs of more
replicating loads/stores.
There are 2 cases that require extra checks to match the legacy cost
model:
1. If the pointer is based on an induction, the legacy cost model passes
its SCEV to getAddressComputationCost. In those cases, still fall back
to the legacy cost. SCEV computations will be added as follow-up
2. If a load is used as part of an address of another load, the legacy
cost model skips the scalarization overhead. Those cases are currently
handled by a usedByLoadOrStore helper.
Note that getScalarizationOverhead also needs updating, because when the
legacy cost model computes the scalarization overhead, scalars have not
been collected yet, so we can't each for replicating recipes to skip
their cost, except other loads. This again can be further improved by
modeling inserts/extracts explicitly and consistently, and compute costs
for those operations directly where needed.
PR: https://github.com/llvm/llvm-project/pull/160053
PR #158641 introduced an issue where i128 accumulator types resulted
in a valid cost, because for a <2 x i128> type the code that
checks for unsupported type legalization would see a type action
of 'TypeSplitVector' which is supported, even though the legalised
type of <1 x i128> would require further scalarization.
This fixes https://github.com/llvm/llvm-project/issues/162009
If a load is scalarized because it is used by a load/store address, the
legacy cost model does not pass ScalarEvolution to getAddressComputationCost.
Match the behavior in VPReplicateRecipe::computeCost.
When using information from dereferenceable assumptions, we need to make
sure that the memory is not freed between the assume and the specified
context instruction. Instead of just checking canBeFreed, check if there
any calls that may free between the assume and the context instruction.
This patch introduces a willNotFreeBetween to check for calls that may
free between an assume and a context instructions, to also be used in
https://github.com/llvm/llvm-project/pull/161255.
PR: https://github.com/llvm/llvm-project/pull/161725
This cost-model takes into account any type-legalisation that would
happen on vectors such as splitting and promotion. This results in wider
VFs being chosen for loops that can use partial reductions.
The cost-model now also assumes that when SVE is available, the SVE dot
instructions for i16 -> i64 dot products can be used for fixed-length
vectors. In practice this means that loops with non-scalable VFs are
vectorized using partial reductions where they wouldn't before, e.g.
```
int64_t foo2(int8_t *src1, int8_t *src2, int N) {
int64_t sum = 0;
for (int i=0; i<N; ++i)
sum += (int64_t)src1[i] * (int64_t)src2[i];
return sum;
}
```
These changes also fix an issue where previously a partial reduction
would be used for mixed sign/zero-extends (USDOT), even when +i8mm was
not available.
Add additional test coverage for tail-folding loops with first-order
recurrences and users outside the loop. Test a combination of
vectorization factors and interleave counts.
Also update check lines in reduction-order.ll and adjust naming for
clarity.
This adds extra test coverage for
https://github.com/llvm/llvm-project/pull/149042.
We can create partial reductions for multiplies with constants, if the
constant is small enough to be extended from source to destination type
w/o changing the value.
This only handles constant on the right side of a multiply, relying on
other passes to canonicalize the input.
Alive2 Proofs: https://alive2.llvm.org/ce/z/iWRMr6
PR: https://github.com/llvm/llvm-project/pull/161092
If there are direct memory op users of the newly scalarized load,
their cost may have changed because there's no scalarization
overhead for the operand. Update it.
This ensures assigning consistent costs to scalarized memory
instructions that themselves have scalarized memory instructions as
operands.
The VPExpressionRecipe class uses a set to store its bundled recipes. If
repeated recipes are bundled then the duplicates will be lost, causing
the following recipes to not be at the expected place in the set.
When printing a reduce.add(mul(ext, ext)) bundle, for example, if the
extends are the same then the 3rd element of the set will be the
reduction, rather than the expected mul, causing a cast error. With this
change, the recipes are at the expected index in the set.
Fixes#156464
