Move the X86 VSELECT->UADDSAT fold to DAGCombiner - there's nothing target specific about these folds.
The SSE42 test diffs are relatively benign - its avoiding an extra constant load in exchange for an extra xor operation - there are extra register moves, which is annoying as all those operations should commute them away.
Differential Revision: https://reviews.llvm.org/D91876
For now, we will hardcode the result as 0.0 if the input is denormal or 0. That will
have the impact the precision. As the fsqrt added belong to the cold path of the
cmp+branch, it won't impact the performance for normal inputs for PowerPC, but improve
the precision if the input is denormal.
Reviewed By: Spatel
Differential Revision: https://reviews.llvm.org/D80974
PowerPC has instruction ftsqrt/xstsqrtdp etc to do the input test for software square root.
LLVM now tests it with smallest normalized value using abs + setcc. We should add hook to
target that has test instructions.
Reviewed By: Spatel, Chen Zheng, Qiu Chao Fang
Differential Revision: https://reviews.llvm.org/D80706
If the scatter store is able to perform the sign/zero extend of
its index, this is folded into the instruction with refineIndexType().
Additionally, refineUniformBase() will return the base pointer and index
from an add + splat_vector.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D90942
Fold
VT = (and (sign_extend NarrowVT to VT) #bitmask)
into
VT = (zero_extend NarrowVT)
With this combine, the test replaces a sign extended load + an
unsigned extention with a zero extended load to render one of the
operands of the last multiplication.
BEFORE | AFTER
f_i16_i32: | f_i16_i32:
.fnstart | .fnstart
ldrsh r0, [r0] | ldrh r1, [r1]
ldrsh r1, [r1] | ldrsh r0, [r0]
smulbb r0, r1, r0 | smulbb r0, r0, r1
uxth r1, r1 | mul r0, r0, r1
mul r0, r0, r1 | bx lr
bx lr |
Reviewed By: resistor
Differential Revision: https://reviews.llvm.org/D90605
Summary:
For vector element types which are not byte-sized, we would generate
incorrect scalar offsets and produce incorrect codegen.
This optimization could potentially be supported in the future, e.g. by
loading in bytes, then shifting and masking out the remaining bits of
the vector element. However, without an upstream target to test against
it's best to avoid the bad codegen in the simplest possible way.
Related to this bug:
https://bugs.llvm.org/show_bug.cgi?id=27600
Reviewed by: foad
Differential Revision: https://reviews.llvm.org/D78568
This patch uses the existing LowerFixedLengthReductionToSVE function to also lower
scalable vector reductions. A separate function has been added to lower VECREDUCE_AND
& VECREDUCE_OR operations with predicate types using ptest.
Lowering scalable floating-point reductions will be addressed in a follow up patch,
for now these will hit the assertion added to expandVecReduce() in TargetLowering.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D89382
The modified code in visitSTORE was missing a scalable vector check, and still
using the now deprecated implicit cast of TypeSize to uint64_t through the
overloaded operator. This patch fixes these issues.
This brings the logic in line with the comment on the context line immediately
above the added precondition.
Add a test in sve-redundant-store.ll that the warning is not triggered.
Differential Revision: https://reviews.llvm.org/D89701
This reverts commit 4604441386dc5fcd3165f4b39f5fa2e2c600f1bc.
Reverting because it was not the intended version of the patch, which
follows this patch.
The modified code in visitSTORE was missing a scalable vector check, and still
using the now deprecated implicit cast of TypeSize to uint64_t through the
overloaded operator. This patch fixes these issues.
This brings the logic in line with the comment on the context line immediately
above the added precondition.
Add a test in Redundantstores.ll that the warning is not triggered.
From LangRef, FMF contract should not enable reassociating to form
arbitrary contractions. So it should not help rearrange nodes like
(fma (fmul x, c1), c2, y) into (fma x, c1*c2, y).
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D89527
MULH is often expanded on targets.
This patch removes the isMulhCheaperThanMulShift hook and uses
isOperationLegalOrCustom instead.
Differential Revision: https://reviews.llvm.org/D80485
In certain places in llvm/lib/CodeGen we were relying upon the TypeSize
comparison operators when in fact the code was only ever expecting
either scalar values or fixed width vectors. This patch changes a few
functions that were always expecting to work on scalar or fixed width
types:
1. DAGCombiner::mergeTruncStores - deals with scalar integers only.
2. DAGCombiner::ReduceLoadWidth - not valid for vectors.
3. DAGCombiner::createBuildVecShuffle - should only be used for
fixed width vectors.
4. SelectionDAGLegalize::ExpandFCOPYSIGN and
SelectionDAGLegalize::getSignAsIntValue - only work on scalars.
Differential Revision: https://reviews.llvm.org/D88562
In certain places in llvm/lib/CodeGen we were relying upon the TypeSize
comparison operators when in fact the code was only ever expecting
either scalar values or fixed width vectors. I've changed some of these
places to use the equivalent scalar operator.
Differential Revision: https://reviews.llvm.org/D88482
In certain places in the code we can never end up in a situation where
we're mixing fixed width and scalable vector types. For example,
we can't have truncations and extends that change the lane count. Also,
in other places such as GenWidenVectorStores and GenWidenVectorLoads we
know from the behaviour of FindMemType that we can never choose a vector
type with a different scalable property.
In various places I have used EVT::bitsXY functions instead of
TypeSize::isKnownXY, where it probably makes sense to keep an assert
that scalable properties match.
Differential Revision: https://reviews.llvm.org/D88654
This passes existing X86 test but I'm not sure if it handles all type
legalization cases it needs to.
Alternative to D89200
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D89222
I have introduced a new template PolySize class, where the template
parameter determines the type of quantity, i.e. for an element
count this is just an unsigned value. The ElementCount class is
now just a simple derivation of PolySize<unsigned>, whereas TypeSize
is more complicated because it still needs to contain the uint64_t
cast operator, since there are still many places in the code that
rely upon this implicit cast. As such the class also still needs
some of it's own operators.
I've tried to minimise the amount of code in the base PolySize
class, which led to a couple of changes:
1. In some places we were relying on '==' operator comparisons
between ElementCounts and the scalar value 1. I didn't put this
operator in the new PolySize class, and thought it was actually
clearer to use the isScalar() function instead.
2. I removed the isByteSized function and replaced it with calls
to isKnownMultipleOf(8).
I've also renamed NextPowerOf2 to be coefficientNextPowerOf2 so
that it's more consistent with coefficientDivideBy.
Differential Revision: https://reviews.llvm.org/D88409
Summary: This patch is derived from D87384.
In this patch we expand the existing decomposition of mul-by-constant to be more general by implementing 2 patterns:
```
mul x, (2^N + 2^M) --> (add (shl x, N), (shl x, M))
mul x, (2^N - 2^M) --> (sub (shl x, N), (shl x, M))
```
The conversion will be trigged if the multiplier is a big constant that the target can't use a single multiplication instruction to handle. This is controlled by the hook `decomposeMulByConstant`.
More over, the conversion benefits from an ILP improvement since the instructions are independent. A case with the sequence like following also gets benefit since a shift instruction is saved.
```
*res1 = a * 0x8800;
*res2 = a * 0x8080;
```
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D88201
In DAGCombiner::ForwardStoreValueToDirectLoad I have fixed up some
implicit casts from TypeSize -> uint64_t and replaced calls to
getVectorNumElements() with getVectorElementCount(). There are some
simple cases of forwarding that we can definitely support for
scalable vectors, i.e. when the store and load are both scalable
vectors and have the same size. I have added tests for the new
code paths here:
CodeGen/AArch64/sve-forward-st-to-ld.ll
Differential Revision: https://reviews.llvm.org/D87098
getNode handling for ISD:SETCC calls FoldSETCC which can canonicalize
FP constants to the RHS. When this happens we should create the node
with the FMF that was requested. By using FlagInserter when can ensure
any calls to getNode/getSetcc during canonicalization will also get the flags.
Differential Revision: https://reviews.llvm.org/D88063
In the motivating case from https://llvm.org/PR47517
we create a node that does not get constant folded
before getNegatedExpression is attempted from some
other node, and we crash.
By moving the fold into SelectionDAG::simplifyFPBinop(),
we get the constant fold sooner and avoid the problem.
After some recent upstream discussion we decided that it was best
to avoid having the / operator for both ElementCount and TypeSize,
since this could give the impression that these classes can be used
in the same way as basic integer integer types. However, division
for scalable types is a bit odd because we are only dividing the
minimum quantity by a value, as opposed to something like:
(MinSize * Vscale) / SomeValue
This is why when performing division it's important the caller
first establishes whether the operation makes sense, perhaps by
calling isKnownMultipleOf() prior to division. The caller must now
explictly call divideCoefficientBy() on the class to perform the
operation.
Differential Revision: https://reviews.llvm.org/D87700
If we're multiplying all elements of a vector by '0' or '1' then we can more efficiently perform this as a clearing mask (that is likely to further simplify to a shuffle blend).
This was noticed when reviewing D87502 but seems to help idiv/irem by constant cases even more as '0'/'1' values are often used for 'passthrough' cases.
Differential Revision: https://reviews.llvm.org/D88225
This is like FastMathFlagGuard in IR. Since we use SDAG instance to get
values, it's with SelectionDAG. By creating a FlagInserter in current
scope, all values created by getNode will get the flags if no Flags
argument provided.
In this patch, I applied it to floating point operations folding part in
DAG combiner, and removed Flags passing to getNode to show its effect.
Other places in DAG combiner and other helper methods similar to getNode
also need this. They can be done in follow-up patches.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D87361
An existing function Type::getScalarSizeInBits returns a uint64_t
instead of a TypeSize class because the caller is requesting a
scalar size, which cannot be scalable. This patch makes other
similar functions requesting a scalar size consistent with that,
thereby eliminating more than 1000 implicit TypeSize -> uint64_t
casts.
Differential revision: https://reviews.llvm.org/D87889
If we have an all ones mask, we can just a regular masked load. InstCombine already gets this in IR. But the all ones mask can appear after type legalization.
Only avx512 test cases are affected because X86 backend already looks for element 0 and the last element being 1. It replaces this with an unmasked load and blend. The all ones mask is a special case of that where the blend will be removed. That transform is only enabled on avx2 targets. I believe that's because a non-zero passthru on avx2 already requires a separate blend so its more profitable to handle mixed constant masks.
This patch adds a dedicated all ones handling to the target independent DAG combiner. I've skipped extending, expanding, and index loads for now. X86 doesn't use index so I don't know much about it. Extending made me nervous because I wasn't sure I could trust the memory VT had the right element count due to some weirdness in vector splitting. For expanding I wasn't sure if we needed different undef handling.
Differential Revision: https://reviews.llvm.org/D87788
The versions that take 'unsigned' will be removed in the future.
I tried to use getOriginalAlign instead of getAlign in some
places. getAlign factors in the minimum alignment implied by
the offset in the pointer info. Since we're also passing the
pointer info we can use the original alignment.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D87592
Similar to D87415, this folds the various float min/max opcodes
with a constant INF or -INF operand, or FLT_MAX / -FLT_MAX operand
if the ninf flag is set. Some of the folds are only possible under
nnan.
The fminnum(X, INF) with nnan and fmaxnum(X, -INF) with nnan cases
are needed to improve the VECREDUCE_FMIN/FMAX lowerings on X86,
the rest is here for the sake of completeness.
Differential Revision: https://reviews.llvm.org/D87571
Previously, we formed ISD::PARITY by looking for (and (ctpop X), 1)
but the AND might be separated from the ctpop. For example if the
parity result is multiplied by 2, we'll pull the AND through the
shift.
So to handle more cases, move to SimplifyDemandedBits where we
can handle more cases that result in only the LSB of the CTPOP
being used.
DAG combiner folds (fma a 1.0 b) into (fadd a b) but the flag isn't
propagated into new fadd. This patch fixes that.
Some code in visitFMA is redundant and such support for vector constants
is missing. Need follow-up patch to clean.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D87037
Clang emits (and (ctpop X), 1) for __builtin_parity. If ctpop
isn't natively supported by the target, this leads to poor codegen
due to the expansion of ctpop being more complex than what is needed
for parity.
This adds a DAG combine to convert the pattern to ISD::PARITY
before operation legalization. Type legalization is updated
to handled Expanding and Promoting this operation. If after type
legalization, CTPOP is supported for this type, LegalizeDAG will
turn it back into CTPOP+AND. Otherwise LegalizeDAG will emit a
series of shifts and xors followed by an AND with 1.
I've avoided vectors in this patch to avoid more legalization
complexity for this patch.
X86 previously had a custom DAG combiner for this. This is now
moved to Custom lowering for the new opcode. There is a minor
regression in vector-reduce-xor-bool.ll, but a follow up patch
can easily fix that.
Fixes PR47433
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D87209
fminnum(X, NaN) is X, fminimum(X, NaN) is NaN. This mirrors the
behavior of existing InstSimplify folds.
This is expected to improve the reduction lowerings in D87391,
which use NaN as a neutral element.
Differential Revision: https://reviews.llvm.org/D87415
During the main DAGCombine loop, whenever a node gets replaced, the new
node and all its users are pushed onto the worklist. Omit this if the
new node is the EntryToken (e.g. if a store managed to get optimized
out), because re-visiting the EntryToken and its users will not uncover
any additional opportunities, but there may be a large number of such
users, potentially causing compile time explosion.
This compile time explosion showed up in particular when building the
SingleSource/UnitTests/matrix-types-spec.cpp test-suite case on any
platform without SIMD vector support.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D86963
This removes the after the fact FMF handling from D46854 in favor of passing fast math flags to getNode. This should be a superset of D87130.
This required adding a SDNodeFlags to SelectionDAG::getSetCC.
Now we manage to contant fold some stuff undefs during the
initial getNode that we don't do in later DAG combines.
Differential Revision: https://reviews.llvm.org/D87200
This is a follow-up suggested in D86420 - if we have a pair of stores
in inverted order for the target endian, we can rotate the source
bits into place.
The "be_i64_to_i16_order" test shows a limitation of the current
function (which might be avoided if we integrate this function with
the other cases in mergeConsecutiveStores). In the earlier
"be_i64_to_i16" test, we skip the first 2 stores because we do not
match the full set as consecutive or rotate-able, but then we reach
the last 2 stores and see that they are an inverted pair of 16-bit
stores. The "be_i64_to_i16_order" test alters the program order of
the stores, so we miss matching the sub-pattern.
Differential Revision: https://reviews.llvm.org/D87112
When lowering fixed length vector operations for SVE the subvector
operations are used extensively to marshall data between scalable
and fixed-length vectors. This means that sequences like:
extract_subvec(binop(insert_subvec(a), insert_subvec(b)))
are very common. DAGCombine only checks if the resulting binop is
legal or can be custom lowered when undoing such sequences. When
it's custom lowering that is introducing them the result is an
infinite legalise->combine->legalise loop.
This patch extends the isOperationLegalOr... functions to include
a "LegalOnly" parameter to restrict the check to legal operations
only. Although isOperationLegal could be used it's common for
the affected code paths to be visited pre and post legalisation,
so the extra parameter keeps the code tidy.
Differential Revision: https://reviews.llvm.org/D86450
fabs and fneg share a common transformation:
(fneg (bitconvert x)) -> (bitconvert (xor x sign))
(fabs (bitconvert x)) -> (bitconvert (and x ~sign))
This patch separate the code into a single method.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D86862
I tried to fix this in:
rG716e35a0cf53
...but that patch depends on the order that we encounter the
magic "x/sqrt(x)" expression in the combiner's worklist.
This patch should improve that by waiting until we walk the
user list to decide if there's a use to skip.
The AArch64 test reveals another (existing) ordering problem
though - we may try to create an estimate for plain sqrt(x)
before we see that it is part of a 1/sqrt(x) expression.
In general, we probably want to try the multi-use reciprocal
transform before sqrt transforms, but x/sqrt(x) is a special-case
because that will always reduce to plain sqrt(x) or an estimate.
The AArch64 tests show that the transform is limited by TLI
hook to patterns where there are 3 or more uses of the divisor.
So this change can result in an extra division compared to
what we had, but that's the intended behvior based on the
current setting of that hook.
Current `v:t = zext(setcc x,y,cc)` will be transformed to `select x, y, 1:t, 0:t, cc`. It misses some opportunities if x's type size is less than `t`'s size. This patch enhances the above transformation.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D86687
This patch changes ElementCount so that the Min and Scalable
members are now private and can only be accessed via the get
functions getKnownMinValue() and isScalable(). In addition I've
added some other member functions for more commonly used operations.
Hopefully this makes the class more useful and will reduce the
need for calling getKnownMinValue().
Differential Revision: https://reviews.llvm.org/D86065
