In the same spirit as D73543 and in reply to https://reviews.llvm.org/D126768#3549920 this patch is adding support for `__builtin_memset_inline`.
The idea is to get support from the compiler to easily write efficient memory function implementations.
This patch could be split in two:
- one for the LLVM part adding the `llvm.memset.inline.*` intrinsics.
- and another one for the Clang part providing the instrinsic as a builtin.
Differential Revision: https://reviews.llvm.org/D126903
This should fix a number of shuffle regressions in D127115 where the re-ordered combines mean we fail to fold a EXTRACT_VECTOR_ELT/INSERT_VECTOR_ELT sequence into a BUILD_VECTOR if we extract from more than one vector source.
This matches what we do in IR. For the RISC-V test case, this allows
us to use -8 for the AND mask instead of materializing a constant in a register.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D127335
During lowering of memcmp/bcmp, the check for a size of 0 is done
in 2 different ways. In rare cases this can lead to a crash in
SystemZSelectionDAGInfo::EmitTargetCodeForMemcmp(). The root cause
is that SelectionDAGBuilder::visitMemCmpBCmpCall() checks for a
constant int value which is not yet evaluated. When the value is
turned into a SDValue, then the evaluation is done and results in
a ConstantSDNode. But EmitTargetCodeForMemcmp() expects the special
case of 0 length to be handled, which results in an assertion.
The fix is to turn the value into a SDValue, so that both functions
use the same check.
Reviewed By: uweigand
Differential Revision: https://reviews.llvm.org/D126900
Extend the TypeWidenVector case of PromoteIntRes_BITCAST to work
with TypeSize directly rather than silently casting to unsigned.
To accomplish this I've extended TypeSize with an interface that
essentially allows TypeSize division when both operands have the
same number of dimensions.
There still exists combinations of scalable vector bitcasts that
cause compiler crashes. I call these out by adding "is missing"
entries to sve-bitcast.
Depends on D126957.
Fixes: #55114
Differential Revision: https://reviews.llvm.org/D127126
Bitcasting between unpacked scalable vector types of different
element counts is not a NOP because the live elements are laid out
differently.
01234567
e.g. nxv2i32 = XX??XX??
nxv4f16 = X?X?X?X?
Differential Revision: https://reviews.llvm.org/D126957
As noticed on D127115 - we were missing this fold, instead just having the shuffle(shuffle(x,undef,splatmask),undef) fold. We should be able to merge these into one using SelectionDAG::isSplatValue, but we'll need to match the shuffle's undef handling first.
This also exposed an issue in SelectionDAG::isSplatValue which was incorrectly propagating the undef mask across a bitcast (it was trying to just bail with a APInt::isSubsetOf if it found any undefs but that was actually the wrong way around so didn't fire for partial undef cases).
I can't remove the function just yet as it is used in the generated .inc files.
I would also like to provide a way to compare alignment with TypeSize since it came up a few times.
Differential Revision: https://reviews.llvm.org/D126910
These assert that there are no "useless" assertzext/assertsext nodes
(that assert a wider width than a following trunc), but I don't think
there is anything preventing such nodes from reaching this code.
I don't think the assertion is relevant for correctness of this
transform either -- if such an assert is present, then the other
one will always be to a smaller width, and we'll pick that one.
The assertion dates back to D37017.
Fixes https://github.com/llvm/llvm-project/issues/55846.
Differential Revision: https://reviews.llvm.org/D126952
Move the code that was added for D126896 after the normal recursive calls
to computeKnownBits. This allows us to calculate trailing zeros.
Previously we would break out of the switch before the recursive calls.
Some cl::ZeroOrMore were added to avoid the `may only occur zero or one times!`
error. More were added due to cargo cult. Since the error has been removed,
cl::ZeroOrMore is unneeded.
Also remove cl::init(false) while touching the lines.
When promoting a shift, make sure we only fetch the second operand
after promoting the first. Load promotion may replace users of the
old load, and we don't want to be left with a dangling reference to
the old load instruction.
The crashing test case is from https://reviews.llvm.org/D126689#3553212.
Differential Revision: https://reviews.llvm.org/D126886
If C is non-negative, the result of the smax must also be
non-negative, so all sign bits of the result are 0.
This allows DAGCombiner to remove a zext_inreg in the modified test.
This zext_inreg started as a sext that became zext before type
legalization then was promoted to a zext_inreg.
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D126896
With a fix for an expensive checks build failure exposed by new RISC-V tests.
Something about expanding two rotates in type legalization caused a change
in the remapping tables that the expensive checks verifying wasn't expecting.
See comment in the code for how it was fixed.
Tests came from this commit that exposed the bug
[RISCV] Add test cases showing failure to remove mask on rotate amounts.
If the masking AND has multiple users we fail to remove it.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D126036
The second argument to is_fp_class specifies the set of floating-point
class to test against. It can be zero, in this case the intrinsic is
expected to return zero value.
Differential Revision: https://reviews.llvm.org/D112025
Followup to D125988 - FPOW is similar to FREM and will most likely scalarize to libcalls, so unroll before widening to prevent use making additional libcalls with UNDEF args.
Currently, two element vectors produced as the result of a binary op are
widened to four element vectors on x86 by
DAGTypeLegalizer::WidenVecRes_BinaryCanTrap. If the op still isn't legal
after widening it is unrolled into 4 scalar ops in SelectionDAG before
being converted into a libcall. This way we end up with 4 libcalls (two of
them on known undef elements) instead of the original two libcalls.
This patch modifies DAGTypeLegalizer::WidenVectorResult to ensure that if
it is known that a binary op will be tunred into a libcall, it is unrolled
instead of being widened. This prevents the creation of the extra scalar
instructions on known undef elements and (eventually) libacalls with known
undef parameters which would otherwise be created when the op gets expanded
post widening.
Differential Revision: https://reviews.llvm.org/D125988
If the VT is i2, then 2 is really -2.
Test has not been commited yet, but diff shows the change.
Fixes PR55644.
Differential Revision: https://reviews.llvm.org/D126213
I had initially assumed this was the problem with
https://github.com/llvm/llvm-project/issues/55271#issuecomment-1133426243
But it turns out that was a simpler issue. This patch is still
more correct than what we were doing before so figured I'd submit
it anyway.
No test case because I'm not sure how to get an undef around
until expansion.
Looking at the test deltas I wonder if it be valid to combine
(sext_inreg (freeze (aextload X))) -> (freeze (sextload X)).
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D126175
abs should only produce a positive value or the signed minimum
value. This means we can't fold abs(undef) to undef as that would
allow more values. Fold to 0 instead to match InstSimplify.
Fixes test mentioned in comment on pr55271.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D126174
Most clients only used these methods because they wanted to be able to
extend or truncate to the same bit width (which is a no-op). Now that
the standard zext, sext and trunc allow this, there is no reason to use
the OrSelf versions.
The OrSelf versions additionally have the strange behaviour of allowing
extending to a *smaller* width, or truncating to a *larger* width, which
are also treated as no-ops. A small amount of client code relied on this
(ConstantRange::castOp and MicrosoftCXXNameMangler::mangleNumber) and
needed rewriting.
Differential Revision: https://reviews.llvm.org/D125557
This function tries to match (a >> 8) | (a << 8) as (bswap a) >> 16.
If the SRL isn't masked and the high bits aren't demanded, we still
need to ensure that bits 23:16 are zero. After the right shift they
will be in bits 15:8 which is where the important bits from the SHL
end up. It's only a bswap if the OR on bits 15:8 only takes the bits
from the SHL.
Fixes PR55484.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D125641
The patch does not pass math flags to float VPCmpIntrinsics because LLParser
could not identify float VPCmpIntrinsics as FPMathOperators.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D125600
If we're using shift pairs to mask, then relax the one use limit if the shift amounts are equal - we'll only be generating a single AND node.
AArch64 has a couple of regressions due to this, so I've enforced the existing one use limit inside a AArch64TargetLowering::shouldFoldConstantShiftPairToMask callback.
Part of the work to fix the regressions in D77804
Differential Revision: https://reviews.llvm.org/D125607
This patch uses VP_REDUCE_AND and VP_REDUCE_OR to replace VP_REDUCE_SMAX,VP_REDUCE_SMIN,VP_REDUCE_UMAX and VP_REDUCE_UMIN for mask vector type.
Differential Revision: https://reviews.llvm.org/D125002
During early gather/scatter enablement two different approaches
were taken to represent scaled indices:
* A Scale operand whereby byte_offsets = Index * Scale
* An IndexType whereby byte_offsets = Index * sizeof(MemVT.ElementType)
Having multiple representations is bad as shown by this patch which
fixes instances where the two are out of sync. The dedicated scale
operand is more flexible and pervasive so this patch removes the
UNSCALED values from IndexType. This means all indices are scaled
but the scale can be one, hence unscaled. SDNodes now use the scale
operand to answer the "isScaledIndex" question.
I toyed with the idea of keeping the UNSCALED enums and helper
functions but because they will have no uses and force SDNodes to
validate the set of supported values I figured it's best to remove
them. We can re-add them if there's a real need. For similar
reasons I've kept the IndexType enum when a bool could be used as I
think being explicitly looks better.
Depends On D123347
Differential Revision: https://reviews.llvm.org/D123381
If we use multiply it would be with 0x0101 which is 1 more than a power
of 2. On some targets we would expand this to shl+add. By avoiding the
multiply earlier, we can generate better code.
Note, PowerPC doesn't do the shl+add expansion of multiply so one of
the tests increased in instruction count.
Limiting to scalars because it almost always increased the number of
instructions in vector tests.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D125638
The patch make users not need to know getNode with SDNodeFlags argument may not
pass its flags.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D125659
FunctionLoweringInfo::StatepointRelocationMaps map is used to pass GC pointer
lowering information from statepoint to gc.relocate which may appear ini
different block.
D124444 introduced different lowering for local and non-local relocates.
Local relocates use SDValue and non-local relocates use value exported to VReg.
But I overlooked the fact that StatepointRelocationMap is indexed not by
GCRelocate instruction, but by derived pointer. This works incorrectly when
we have two relocates (one local and another non-local) of the same value,
because they need different relocation records.
This patch fixes the problem by recording relocation information per relocate
instruction, not per derived pointer. This way, each gc.relocate can be lowered
differently.
Reviewed By: skatkov
Differential Revision: https://reviews.llvm.org/D125538
FastISel tries to fold loads into the single using instruction.
However, if the register has fixups, then there may be additional
uses through an alias of the register.
In particular, this fixes the problem reported at
https://reviews.llvm.org/D119432#3507087. The load register is
(at the time of load folding) only used in a single call instruction.
However, selection of the bitcast has added a fixup between the
load register and the cross-BB register of the bitcast result.
After fixups are applied, there would now be two uses of the load
register, so load folding is not legal.
Differential Revision: https://reviews.llvm.org/D125459
SelectionDAG::FoldConstantArithmetic determines if operands are foldable constants, so we don't need to bother with isConstantOrConstantVector / Opaque tests before calling it directly.
SelectionDAG::FoldConstantArithmetic determines if operands are foldable constants, so we don't need to bother with isConstantOrConstantVector / Opaque tests before calling it directly.
Pulled out of D77804 as its going to be easier to address the regressions individually.
This patch allows SimplifyDemandedBits to call SimplifyMultipleUseDemandedBits in cases where the source operand has other uses, enabling us to peek through the shifted value if we don't demand all the bits/elts.
The lost RISCV gorc2 fold shouldn't be a problem - instcombine would have already destroyed that pattern - see https://github.com/llvm/llvm-project/issues/50553
Differential Revision: https://reviews.llvm.org/D124839
As suggested from 02f8519502, this uses the
isAnyConstantBuildVector method in lieu of separate
isBuildVectorOfConstantSDNodes calls. It should
otherwise be an NFC.
This prevents an infinite loop from D123801, where code trying to reduce
the total number of bitcasts, but also handling constants, could create
the opposite transform. Prevent the transform in these case to let the
bitcast of a constant transform naturally.
Fixes#55345
Like other shifts, the type isn't required to match. We shouldn't
assume we can call ZExtPromotedInteger.
I tested the PromoteIntOp_FunnelShift locally by removing the promotion
of the shift amount from PromoteIntRes_FunnelShift. But with the final
version of this patch it is never executed on any tests.
Differential Revision: https://reviews.llvm.org/D125106
This is part of an ongoing effort toward making DAGCombine process the nodes in topological order.
This is able to discover a couple of new optimizations, but also causes a couple of regression. I nevertheless chose to submit this patch for review as to start the discussion with people working on the backend so we can find a good way forward.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D124743
Add helper functions to query the signed and scaled properties
of ISD::IndexType along with functions to change them.
Remove setIndexType from MaskedGatherSDNode because it only has
one usage and typically should only be changed alongside its
index operand.
Minimise the direct use of the enum values to lay the groundwork
for more refactoring.
Differential Revision: https://reviews.llvm.org/D123347
Something is going wrong with the BigEndian PowerPC bot. It is hard to
tell what is wrong from here, but attempt to fix it by disabling the
combineShuffleOfBitcast combine for bigendian.
Otherwise we have garbage in the upper bits that can affect the
results of the UREM.
Fixes PR55296.
Differential Revision: https://reviews.llvm.org/D125076
If the mask is made up of elements that form a mask in the higher type
we can convert shuffle(bitcast into the bitcast type, simplifying the
instruction sequence. A v4i32 2,3,0,1 for example can be treated as a
1,0 v2i64 shuffle. This helps clean up some of the AArch64 concat load
combines, along with helping simplify a number of other tests.
The PowerPC combine for v16i8 splat vector loads needed some fixes to
keep it working for v16i8 vectors. This improves the handling of v2i64
shuffles to match too, hopefully improving them in general.
Differential Revision: https://reviews.llvm.org/D123801
The result of sign_extend_inreg needs to have as many sign bits
as requested by the VT argument. The easiest way to guarantee this
is to fold it to 0.
SystemZ test was modified to avoid using undef.
Fixes https://github.com/llvm/llvm-project/issues/55178
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D124696
This extends the (X & ~Y) | Y to X | Y fold to also work if ~Y is
a truncated not (when taking into account the mask X). This is
done by exporting the infrastructure added in D124856 and reusing
it here.
I've retained the old value of AllowUndefs=false, though probably
this can be switched to true with extra test coverage.
Differential Revision: https://reviews.llvm.org/D124930
Demanded bits analysis may replace a full-width not with a
any_extend (not (truncate X)) pattern. This patch looks through
this kind of pattern in haveNoCommonBitsSet(). Of course, we can
only do this if we only need negated bits in the non-extended part,
as the other bits may now be arbitrary. For example, if we have
haveNoCommonBitsSet(~X & Y, X) then ~X only needs to actually
negate bits set in Y.
This is only a partial solution to the problem in that it allows
add -> or conversion, but the resulting or doesn't get folded yet.
(I guess that will involve exposing getBitwiseNotOperand() as a
more general helper and using that in the relevant transform.)
Differential Revision: https://reviews.llvm.org/D124856
Don't assume the rotation amounts have been correctly normalized - do it as part of the constant folding.
Also, the normalization should be performed with UREM not SREM.
This is the DAG variant of D124763. The code already handles the
general pattern, but not this degenerate case.
This allows folding A + (B&~A) to A | (B&~A) which further holds
to A | B.
Handling on the SDAG level is needed because in the motivating
case the add is actually a getelementptr, which only gets converted
into an add on the SDAG level. However, this patch is not quite
sufficient to handle the getelementptr case yet, because of an
interfering demanded bits simplification.
Differential Revision: https://reviews.llvm.org/D124772
When looking for memory uses,
reassociationCanBreakAddressingModePattern should check uses of
the outer ADD rather than the inner ADD. We want to know if the
two ops we're reassociating are used by a load/store.
In practice, the existing check usually works because CodeGenPrepare
will make one of the load/stores have an offset of 0 relative to
split GEP. That will make the inner add have a memory use.
To test this, I've manually split the GEPs so there is no 0 offset
store.
This issue was recently discussed in the original review D60294.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D124644
SIGN_EXTEND_INREG expansion can trigger a TypeSize error because
"VT.getSizeInBits() == 1" is used to detect for a boolean without
first verifying VT is a scalar.
We try to match as a disguised rotate by constant of these forms
(shl (X | Y), C1) | (srl X, C2) --> (rotl X, C1) | (shl Y, C1)
(shl X, C1) | (srl (X | Y), C2) --> (rotl X, C1) | (srl Y, C2)
We may have also looked through an AND to find the shift. If we
did, we need to apply a mask to the result.
I'll add an AArch64 test and pre-commit it and the RISC-V test
tomorrow.
Fixes PR55201.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D124711
When looking through extends of gather/scatter indices it's safe
to convert a known positive signed index to unsigned, but unsigned
indices must remain unsigned.
Depends On D123318
Differential Revision: https://reviews.llvm.org/D123326
This is an alternative to D124530. In getUniformBase() only create
scales that match the gather/scatter element size. If targets also
support other scales, then they can produce those scales in target
DAG combines. This is what X86 already does (as long as the
resulting scale would be 1, 2, 4 or 8).
This essentially restores the pre-opaque-pointer state of things.
Fixes https://github.com/llvm/llvm-project/issues/55021.
Differential Revision: https://reviews.llvm.org/D124605
refineUniformBase and selectGatherScatterAddrMode both attempt the
transformation:
base(0) + index(A+splat(B)) => base(B) + index(A)
However, this is only safe when index is not implicitly scaled.
Differential Revision: https://reviews.llvm.org/D123222
PowerPC supports `ppc_fp128`, which is not an IEEE floating point
type. The generic lowering of llvm.is_fpclass could not handle it
properly. This change extends the generic lowering code to
support `ppc_fp128`.
The change was tested on emulator using runtime tests from
https://reviews.llvm.org/D112933 and the patch for clang
https://reviews.llvm.org/D112932.
Differential Revision: https://reviews.llvm.org/D113908
Introduced masks where they are not added and improved target dependent
cost models to avoid returning of the incorrect cost results after
adding masks.
Differential Revision: https://reviews.llvm.org/D100486
The description of SETCC says
/// SetCC operator - This evaluates to a true value iff the condition is
/// true. If the result value type is not i1 then the high bits conform
/// to getBooleanContents.
Without this patch, we sign extended the i1 to the used larger type
regardless of getBooleanContents. This resulted in miscompiles, as
shown in the attached testcase that ended up returning -1 instead of
1 when using -mattr=+v.
Fixes https://github.com/llvm/llvm-project/issues/55168
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D124618
Introduced masks where they are not added and improved target dependent
cost models to avoid returning of the incorrect cost results after
adding masks.
Differential Revision: https://reviews.llvm.org/D100486
Cuurently we always export STATEPOINT results (GC pointers lowered via VRegs)
to virtual registers. When processing gc.relocate instructions we have to
generate CopyFromRegs node and then export it to VReg again if gc.relocate
is used in other basic blocks. This results in generation of extra COPY MIR
instruction if statepoint and its gc.relocate are in the same BB, but gc.relocate
result is used in other blocks.
This patch changes this behavior to export statepoint results only if used
in other basic blocks. For local uses StatepointLoweringState.(get|set)Location()
API is used to communicate appropriate statepoint result from `LowerStatepoint()`
to `visitGCRelocate()`
This is NFC and is purely compile time optimization. On big methids it can improve
codegen compile time up to 10%.
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D124444
This change introduces a new intrinsic, `llvm.is.fpclass`, which checks
if the provided floating-point number belongs to any of the the specified
value classes. The intrinsic implements the checks made by C standard
library functions `isnan`, `isinf`, `isfinite`, `isnormal`, `issubnormal`,
`issignaling` and corresponding IEEE-754 operations.
The primary motivation for this intrinsic is the support of strict FP
mode. In this mode using compare instructions or other FP operations is
not possible, because if the value is a signaling NaN, floating-point
exception `Invalid` is raised, but the aforementioned functions must
never raise exceptions.
Currently there are two solutions for this problem, both are
implemented partially. One of them is using integer operations to
implement the check. It was implemented in https://reviews.llvm.org/D95948
for `isnan`. It solves the problem of exceptions, but offers one
solution for all targets, although some can do the check in more
efficient way.
The other, implemented in https://reviews.llvm.org/D96568, introduced a
hook 'clang::TargetCodeGenInfo::testFPKind', which injects a target
specific code into IR to implement `isnan` and some other functions. It is
convenient for targets that have dedicated instruction to determine FP data
class. However using target-specific intrinsic complicates analysis and can
prevent some optimizations.
A special intrinsic for value class checks allows representing data class
tests with enough flexibility. During IR transformations it represents the
check in target-independent way and saves it from undesired transformations.
In the instruction selector it allows efficient lowering depending on the
used target and mode.
This implementation is an extended variant of `llvm.isnan` introduced
in https://reviews.llvm.org/D104854. It is limited to minimal intrinsic
support. Target-specific treatment will be implemented in separate
patches.
Differential Revision: https://reviews.llvm.org/D112025
This is a very specific fold to fix an upstream poor codegen issue.
InstCombine has the much more flexible pushFreezeToPreventPoisonFromPropagating but I don't think we're quite there with DAG/TLI handling for canCreateUndefOrPoison/isGuaranteedNotToBeUndefOrPoison value tracking yet.
Fixes#54911
Differential Revision: https://reviews.llvm.org/D124185
We can process the long shuffles (working across several actual
vector registers) in the best way if we take the actual register
represantion into account. We can build more correct representation of
register shuffles, improve number of recognised buildvector sequences.
Also, same function can be used to improve the cost model for the
shuffles. in future patches.
Part of D100486
Differential Revision: https://reviews.llvm.org/D115653
We can process the long shuffles (working across several actual
vector registers) in the best way if we take the actual register
represantion into account. We can build more correct representation of
register shuffles, improve number of recognised buildvector sequences.
Also, same function can be used to improve the cost model for the
shuffles. in future patches.
Part of D100486
Differential Revision: https://reviews.llvm.org/D115653
1. X%C to the equivalent of X-X/C*C is not always fastest path if there is no SDIV pair exist. So check target have faster for srem only first.
2. Add AArch64 faster path for SREM only pow2 case.
Fix https://github.com/llvm/llvm-project/issues/54649
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D122968
hasOneUse is not cheap on nodes with chain results that might have
many uses. By checking the opcode first, we can avoid a costly walk
of the use list on nodes we aren't interested in.
Found by investigating calls to hasNUsesOfValue from the example
provided in D123857.
For strict FP16 to work correctly needs some changes in lowering and
legalization:
* SelectionDAGLegalize::PromoteNode was missing handling for some
strict fp opcodes.
* Some of the custom lowering of strict fp operations needed to be
adjusted to work with FP16.
* Custom lowering needed to be added for round-to-int operations.
With this, and the previous patches for the rest of the strict fp
isel, we can set IsStrictFPEnabled = true.
Differential Revision: https://reviews.llvm.org/D115620
The lowering code did not use the scale operand of MGATHER/MSCATTER
nodes, but instead assumed scaled indices were always scaled based
on the element type of the memory type. This patch adds the missing
support by rewritting the nodes as unscaled variants.
Differential Revision: https://reviews.llvm.org/D123670
This patch adds support for inline assembly address operands using the "p"
constraint on X86 and SystemZ.
This was in fact broken on X86 (see example at
https://reviews.llvm.org/D110267, Nov 23).
These operands should probably be treated the same as memory operands by
CodeGenPrepare, which have been commented with "TODO" there.
Review: Xiang Zhang and Ulrich Weigand
Differential Revision: https://reviews.llvm.org/D122220
As far as I know getNode will never return a null SDValue.
I'm guessing this was modeled after the FoldConstantArithmetic
call earlier.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D123550
Materializing constants on RISCV is simpler if the constant is sign
extended from i32. By default i32 constant operands of phis are
zero extended.
This patch adds a hook to allow RISCV to override this for i32. We
have an existing isSExtCheaperThanZExt, but it operates on EVT which
we don't have at these places in the code.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D122951
We're just trying to canonicalize here and won't be using the constant
value returned.
The attached test changes are because we were previously commuting
a seteq X, (splat_vector 0) because we also have (sub 0, X). The
0 is larger than the element type so we don't detect it as a splat
without the AllowTruncation flag. By preventing the commute we are
able to match it to the vmseq.vx instruction during isel. We only
look for constants on the RHS in isel.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D123256
fshl (or X, Y), X, C ==/!= 0 --> or (shl Y, C), X ==/!= 0
fshl X, (or X, Y), C ==/!= 0 --> or (srl Y, BW-C), X ==/!= 0
This is similar to an existing setcc-of-rotate fold, but the
matching requires more checks for the more general funnel op:
https://alive2.llvm.org/ce/z/Ab2jDd
We are effectively decomposing the funnel shift into logical
shifts, reassociating, and removing a shift.
This should get us the final improvements for x86-64 that were
originally shown in D111530
( https://github.com/llvm/llvm-project/issues/49541 );
x86-32 still shows some SHLD/SHRD, so the pattern is not
matching there yet.
Differential Revision: https://reviews.llvm.org/D122919
arm64_32 guarantees the high 32 bits of pointer parameters are passed as 0, and
this is modelled in the IR by inserting an AssertZExt after the CopyFromReg.
The function deciding whether registers that need to be preserved actually are
wasn't expecting this so it banned perfectly legitimate tail calls.
This patch aims to overcome an issue in these mappings where, when an ISD
node was registered with BEGIN_REGISTER_VP_SDNODE but outwidth the scope
of a pair of BEGIN_REGISTER_VP_INTRINSIC/END_REGISTER_VP_INTRINSIC
macros, the switch cases fell apart. This in particular happened with
VP_SETCC, where we'd end up with something along the lines of:
case Intrinsic::vp_fcmp:
break;
case Intrinsic::vp_icmp:
break;
ResOpc = ISD::VP_SETCC;
case Intrinsic::vp_store:
...
To remedy this, we introduce a special-purpose mapping macro which can
map any number of VP intrinsic opcodes to an ISD opcode.
As a result, we no longer need to special-case the mapping from vp.icmp
and vp.fcmp to VP_SETCC, as the new helper macro does it for us.
Thanks to @craig.topper for noticing this and to @rogfer01 for the idea.
Reviewed By: rogfer01
Differential Revision: https://reviews.llvm.org/D123324
This patch adds the necessary infrastructure to lower vp.fcmp via
ISD::VP_SETCC to RVV instructions.
Most notably this patch adds cond-code legalization for VP_SETCC,
reusing the existing TargetLowering::LegalizeSetCCCondCode by passing in
additional SDValue parameters for the Mask and EVL. This method then
uses VP operations to legalize the condcode.
There is still a general lack of canonicalization on VP_SETCC as opposed
to SETCC which results in worse code than is theoretically possible.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D123051
The VP path was using the split source VTs instead of the split
destination VTs. This may not be a problem today because the VP
nodes going through this have the same source and dest VTs.
It will be a problem when we start using this function for legalizing
VP cast operations.
This patch adds the minimum required to successfully lower vp.icmp via
the new ISD::VP_SETCC node to RVV instructions.
Regular ISD::SETCC goes through a lot of canonicalization which targets
may rely on which has not hereto been ported to VP_SETCC. It also
supports expansion of individual condition codes and a non-boolean
return type. Support for all of that will follow in later patches.
In the case of RVV this largely isn't a problem as the vector integer
comparison instructions are plentiful enough that it can lower all
VP_SETCC nodes on legal integer vectors except for boolean vectors,
which regular SETCC folds away immediately into logical operations.
Floating-point VP_SETCC operations aren't as well supported in RVV and
the backend relies on condition code expansion, so support for those
operations will come in later patches.
Portions of this code were taken from the VP reference patches.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D122743
Promotion does not affect the base element type and so the original
index type will remain unchanged. This reflects the behaviour of
DAGTypeLegalizer::PromoteIntOp_MGATHER with no tests affected.
Accord the discussion in D122281, we missing an ISD::AND combine for MLOAD
because it relies on BuildVectorSDNode is fails for scalable vectors.
This patch is intend to handle that, so we can circle back the type MVT::nxv2i32
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D122703
E.g. in
```
%i0 = zext <2 x i8> to <2 x i16>
%i1 = bitcast <2 x i16> to <4 x i8>
```
the `%i0`'s zero bits are known to be `0xFF00` (upper half of every element is known zero),
but no elements are known to be zero, and for `%i1`, we don't know anything about zero bits,
but the elements under `0b1010` mask are known to be zero (i.e. the odd elements).
But, we didn't perform such a propagation.
Noticed while investigating more aggressive `vpmaddwd` formation.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D123163
Place PersistentId declaration under #if LLVM_ENABLE_ABI_BREAKING_CHECKS to
reduce memory usage when it is not needed.
Differential Revision: https://reviews.llvm.org/D120714
Variable locations now come in two modes, instruction referencing and
DBG_VALUE. At -O0 we pick DBG_VALUE to allow fast construction of variable
information. Unfortunately, SelectionDAG edits the optimisation level in
the presence of opt-bisect-limit, meaning different passes have different
views of what variable location mode we should use. That causes assertions
when they're mixed.
This patch plumbs through a boolean in SelectionDAG from start to
instruction emission, so that we don't rely on the current optimisation
level for correctness.
Differential Revision: https://reviews.llvm.org/D123033
As raised on PR52267, XOR(X,MIN_SIGNED_VALUE) can be treated as ADD(X,MIN_SIGNED_VALUE), so let these cases use the 'AddLike' folds, similar to how we perform no-common-bits OR(X,Y) cases.
define i8 @src(i8 %x) {
%r = xor i8 %x, 128
ret i8 %r
}
=>
define i8 @tgt(i8 %x) {
%r = add i8 %x, 128
ret i8 %r
}
Transformation seems to be correct!
https://alive2.llvm.org/ce/z/qV46E2
Differential Revision: https://reviews.llvm.org/D122754
https://alive2.llvm.org/ce/z/A_auBq
Remove limitation that wouldn't perform the fold if all the inverted bits are known zero
The thumb2 changes look to be benign, although it does show that the TEQ/TST isel patterns could probably be improved.
Fixes movmsk regression in D122754
Differential Revision: https://reviews.llvm.org/D123023
The reason why I am making this change is that before this commit,
EmitFuncArgumentDbgValue relied on a boolean flag IsDbgDeclare both to signal
that a DBG_VALUE should be made to be indirect /and/ that the original intrinsic
was a dbg.declare. This is no longer always true if we add support for handling
dbg.addr since we will have an indirect DBG_VALUE that is a different intrinsic
from dbg.declare.
With that in mind, in this NFC patch, we prepare for future fixes by introducing
a 3 case-enum argument to EmitFuncArgumentDbgValue that allows the caller to
explicitly specify how the argument's DBG_VALUE should be emitted. This then
allows us to turn the indirect checks into a != FuncArgumentDbgValueKind::Value
and prepare us for a future where we add support here for llvm.dbg.addr
directly.
rdar://83957028
Reviewed By: aprantl
Differential Revision: https://reviews.llvm.org/D122945
If we expand (uaddo X, 1) we previously expanded the overflow calculation
as (X + 1) <u X. This potentially increases the live range of X and
can prevent X+1 from reusing the register that previously held X.
Since we're adding 1, overflow only occurs if X was UINT_MAX in which
case (X+1) would be 0. So this patch adds a special case to expand
the overflow calculation to (X+1) == 0.
This seems to help with uaddo intrinsics that get introduced by
CodeGenPrepare after LSR. Alternatively, we could block the uaddo
transform in CodeGenPrepare for this case.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D122933
D122053 set the ExtendType for ConstantSDNodes in getCopyToRegs to
ZERO_EXTEND to match assumptions in ComputePHILiveOutRegInfo. PHIs
are probably not the only way ConstantSDNodeNodes can get to
getCopyToRegs.
This patch adds an ExtendType parameter to CopyValueToVirtualRegister and
has HandlePHINodesInSuccessorBlocks pass ISD::ZERO_EXTEND for ConstantInts.
This way we only affect ConstantSDNodes for PHIs.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D122171
This is an extension of D70965 to avoid creating a mathlib
call where it did not exist in the original source. Also see
D70852 for discussion about an alternative proposal that was
abandoned.
In the motivating bug report:
https://github.com/llvm/llvm-project/issues/54554
...we also have a more general issue about handling "no-builtin" options.
Differential Revision: https://reviews.llvm.org/D122610
When shifting by a byte-multiple:
bswap (shl X, C) --> lshr (bswap X), C
bswap (lshr X, C) --> shl (bswap X), C
This is the backend version of D122010 and an alternative
suggested in D120648.
There's an extra check to make sure the shift amount is
valid that was not in the rough draft.
I'm not sure if there is a larger motivating case for RISCV (bug report?),
but the ARM diffs show a benefit from having a late version of the
transform (because we do not combine the loads in IR).
Differential Revision: https://reviews.llvm.org/D122655
This patch fixes a (seemingly very rare) crash during vector constant
folding introduced in D113300.
Normally, during legalization, if we create an illegally-typed node during
a failed attempt at constant folding it's cleaned up before being
visited, due to it having no uses.
If, however, an illegally-typed node is created during one round of
legalization and isn't cleaned up, it's possible for a second round of
legalization to create new illegally-typed nodes which add extra uses to
the old illegal nodes. This means that we can end up visiting the old
nodes before they're known to be dead, at which point we crash.
I'm not happy about this fix. Creating illegal types at all seems like a
bad idea, but we all-too-often rely on illegal constants being
successfully folded and being fixed up afterwards. However, we can't
rely on constant folding actually happening, and we don't have a
foolproof way of peering into the future.
Perhaps the correct fix is to revisit the node-iteration order during
legalization, ensuring we visit all uses of nodes before the nodes
themselves. Or alternatively we could try and clean up dead nodes
immediately after failing constant folding.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D122382
Modified DAGCombiner to pass the shift the bittest input and the shift amount
to hasBitTest. This matches the other call to hasBitTest in TargetLowering.h
This is an alternative to D122454.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D122458
The alignment needs to be part of the folding set hash. This is
handled by getAssertAlign when nodes are created, but needs to repeated here.
No test case as I found it as part of a very early experimental patch.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D122279
Non-static class members declared under #ifndef NDEBUG should be declared
under #if LLVM_ENABLE_ABI_BREAKING_CHECKS to make headers library-friendly and
allow cross-linking, as discussed in D120714.
Differential Revision: https://reviews.llvm.org/D121549
Instead of using operator[], use DenseMap::find to prevent default
constructing an entry if it isn't already in the map.
Also simplify a condition to check for 0 instead of a virtual register.
I'm pretty sure we can only get 0 or a virtual register out of the value
map.
This is only called for instructions and the caller is already holding
an Instruction *. This makes the code more explicit and makes it
obvious the code doesn't make decisions about constants.
Trying to reduce the number of masked loads in favour of more unpklo/hi
instructions. Both ISD::ZEXTLOAD and ISD::SEXTLOAD are supported to extensions
from legal types.
Both of normal and masked loads test cases added to guard compile crash.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D120953
Add shl/srl/sra to the list of ops that we canonicalize with a select to expose an identity merge
Differential Revision: https://reviews.llvm.org/D122070
ComputePHILiveOutRegInfo assumes that constant incoming values to
Phis will be zero extended if they aren't a legal type. To guarantee
that we should zero_extend rather than any_extend constants.
This fixes a bug for RISCV where any_extend of constants can be
treated as a sign_extend.
Differential Revision: https://reviews.llvm.org/D122053
This fixes a reported bug that caused an infinite loop during the
SelectionDAG optimization phase in ISel, by creating an overridable hook
in `TargetLowering` that allows us to bail out from running
`SimplifyDemandedVectorElts`.
Reviewed By: tlively
Differential Revision: https://reviews.llvm.org/D121869
If we promote the ABS and then Expand in LegalizeDAG, then both the
sra and the xor will have their inputs sign extended. This generates
extra code on RISCV which lacks an i8 or i16 sign extend instructon.
If we expand during type legalization, then only the sra will get its
input sign extended. RISCV is able to combine this with the sra by
doing a shift left followed by an sra.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D121664
RISCV strong prefers i32 values be sign extended to i64. This combine
was always zero extending the constant using APInt methods.
This adjusts the code so that it calls getNode using ISD::ANY_EXTEND instead.
getNode will call TLI.isSExtCheaperThanZExt to decide how to handle
the constant.
Tests were copied from D121598 where I noticed that we were creating
constants that were hard to materialize.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D121650
We do not have general reassociation here (and probably
do not need it), but I noticed these were missing in
patches/tests motivated by D111530, so we can at
least handle the simplest patterns.
The VE test diff looks correct, but we miss that
pattern in IR currently:
https://alive2.llvm.org/ce/z/u66_PM
This patch introduces two new experimental IR intrinsics and SDAG nodes
to represent vector strided loads and stores.
Reviewed By: simoll
Differential Revision: https://reviews.llvm.org/D114884
SDNodes with different target flags may now be folded together
rightfully resulting in the assertion in the refineAlignment.
Folding nodes with different target flags may result in the
wrong load instructions produced at least on the AMDGPU.
Fixes: SWDEV-326805
Differential Revision: https://reviews.llvm.org/D121335
This is another fold generalized from D111530.
We can find a common source for a rotate operation hidden inside an 'or':
https://alive2.llvm.org/ce/z/9pV8hn
Deciding when this is profitable vs. a funnel-shift is tricky, but this
does not show any regressions: if a target has a rotate but it does not
have a funnel-shift, then try to form the rotate here. That is why we
don't have x86 test diffs for the scalar tests that are duplicated from
AArch64 ( 74a65e3834 ) - shld/shrd are available. That also makes it
difficult to show vector diffs - the only case where I found a diff was
on x86 AVX512 or XOP with i64 elements.
There's an additional check for a legal type to avoid a problem seen
with x86-32 where we form a 64-bit rotate but then it gets split
inefficiently. We might avoid that by adding more rotate folds, but
I didn't check to see what is missing on that path.
This gets most of the motivating patterns for AArch64 / ARM that are in
D111530.
We still need a couple of enhancements to setcc pattern matching with
rotate/funnel-shift to get the rest.
Differential Revision: https://reviews.llvm.org/D120933
Previously we used sra+add+xor if ADDCARRY is supported. This changes
to sra+xor+sub is SUBCARRY is available.
This is consistent with the recent change to the default expansion
in LegalizeDAG.
Differential Revision: https://reviews.llvm.org/D121039
When inserting undef into buildvectors created from shuffles of
buildvectors, we convert elements to the largest needed type. This had
the effect of converting undef into 0, which isn't needed as the
buildvector implicitly truncates and trunc(zext(undef)) == undef.
Differential Revision: https://reviews.llvm.org/D121002
This extends acb96ffd14 to 'and' and 'xor' opcodes.
Copying from that message:
LOGIC (LOGIC (SH X0, Y), Z), (SH X1, Y) --> LOGIC (SH (LOGIC X0, X1), Y), Z
https://alive2.llvm.org/ce/z/QmR9rR
This is a reassociation + factoring fold. The common shift operation is moved
after a bitwise logic op on 2 input operands.
We get simpler cases of these patterns in IR, but I suspect we would miss all
of these exact tests in IR too. We also handle the simpler form of this plus
several other folds in DAGCombiner::hoistLogicOpWithSameOpcodeHands().
When triggered during operation legalisation the affected combine
generates a splat_vector that when custom lowered for SVE fixed
length code generation, results in the original precombine sequence
and thus we enter a legalisation/combine hang.
NOTE: The patch contains no tests because I observed this issue
only when combined with other work that might never become public.
The current way AArch64 lowers ISD::SPLAT_VECTOR meant a specific
test was not possible so I'm hoping the DAGCombiner fix can be seen
as obvious. The AArch64ISelLowering change is requirted to maintain
existing code quality.
Differential Revision: https://reviews.llvm.org/D120735
https://alive2.llvm.org/ce/z/mJP7XP
This can be viewed as expanding the compare into and/or-of-compares:
https://alive2.llvm.org/ce/z/bkZYWE
followed by reduction of each compare.
This could be extended in several ways:
1. There's a (X & Y) == -1 sibling.
2. We can recurse through more than 1 'or'.
3. The fold could be generalized beyond rotates - any operation that
only changes the order of bits (bswap, bitreverse).
This is a transform noted in D111530.
Instead of emitting 0 > Hi, emit Hi < 0. If Hi needs to be expanded again
this will allow the special case for sign bit tests in ExpandIntOp_SETCC
to trigger.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D120761
Currently we only check for splat shuffles, this extends it to see if the source operand is a splat across the demanded elts based upon the shuffle mask
This patch adds support for recognising vector splats by peeking through bitcasts to vectors with smaller element types - if all the offset subelements are splats then the bitcasted vector is a splat as well.
We don't have great coverage for isSplatValue so I've made this pretty specific to the use case I'm trying to fix - regressions in some vXi64 vector shift by splat cases that 32-bit x86 doesn't recognise because the shift amount buildvector has been type legalised to v2Xi32.
We can add further support (floats, bitcast from larger element types, undef elements) when we have actual test coverage.
Differential Revision: https://reviews.llvm.org/D120553
If the types aren't legal, the expansions may get type legalized in a
different way preventing code sharing. If the type is legal, we will
share some instructions between the two expansions, but we will need an
extra register.
Since we don't appear to fold (neg (sub A, B)) if the sub has an
additional user, I think it makes sense not to expand NABS.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D120513
This is the SDAG equivalent of an instcombine transform added with:
fd807601a7
This is another step towards solving #49541 and part of an alternative
set of more general transforms than what is proposed in D111530.
https://alive2.llvm.org/ce/z/ToxaE8
LOGIC (LOGIC (SH X0, Y), Z), (SH X1, Y) --> LOGIC (SH (LOGIC X0, X1), Y), Z
https://alive2.llvm.org/ce/z/QmR9rR
This is a reassociation + factoring fold. The common shift operation is moved
after a bitwise logic op on 2 input operands.
We get simpler cases of these patterns in IR, but I suspect we would miss all
of these exact tests in IR too. We also handle the simpler form of this plus
several other folds in DAGCombiner::hoistLogicOpWithSameOpcodeHands().
This is a partial implementation of a transform suggested in D111530
(only handles 'or' bitwise logic as a first step - need to stamp out more
tests for other opcodes).
Several of the same tests added for D111530 are altered here (but not
fully optimized). I'm not sure yet if this would help/hinder that patch,
but this should be an improvement for all tests added with ecf606cb43
since it removes a shift operation in those examples.
Differential Revision: https://reviews.llvm.org/D120516
If the shl is at least half the bitwidth (i.e. the lower half of the bswap source is zero), then we can reduce the shift and perform the bswap at half the bitwidth and just zero extend.
Based off PR51391 + PR53867
Differential Revision: https://reviews.llvm.org/D120192
This is the SDAG translation of D120253 :
https://alive2.llvm.org/ce/z/qHpmNn
The SDAG nodes can have different operand types than the result value.
We can see an example of that with AArch64 - the funnel shift amount
is an i64 rather than i32.
We may need to make that match even more flexible to handle
post-legalization nodes, but I have not stepped into that yet.
Differential Revision: https://reviews.llvm.org/D120264
Internally to DAGCombiner the SDValues were passed by non-const
reference despite not being modified. They were then passed by
const reference to TLI.
This patch passes them by value which is consistent with the vast
majority of code.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D120420
In combineCarryDiamond() use getAsCarry() to find more candidates for being a carry flag.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D118362
This is a fix for a regression discussed in:
https://github.com/llvm/llvm-project/issues/53829
We cleared more high multiplier bits with 995d400,
but that can lead to worse codegen because we would fail
to recognize the now disguised multiplication by neg-power-of-2
as a shift-left. The problem exists independently of the IR
change in the case that the multiply already had cleared high
bits. We also convert shl+sub into mul+add in instcombine's
negator.
This patch fills in the high-bits to see the shift transform
opportunity. Alive2 attempt to show correctness:
https://alive2.llvm.org/ce/z/GgSKVX
The AArch64, RISCV, and MIPS diffs look like clear wins. The
x86 code requires an extra move register in the minimal examples,
but it's still an improvement to get rid of the multiply on all
CPUs that I am aware of (because multiply is never as fast as a
shift).
There's a potential follow-up noted by the TODO comment. We
should already convert that pattern into shl+add in IR, so
it's probably not common:
https://alive2.llvm.org/ce/z/7QY_GaFixes#53829
Differential Revision: https://reviews.llvm.org/D120216
Previous we used sra (X, size(X)-1); xor (add (X, Y), Y).
By placing sub at the end, we allow RISCV to combine sign_extend_inreg
with it to form subw.
Some X86 tests for Z - abs(X) seem to have improved as well.
Other targets look to be a wash.
I had to modify ARM's abs matching code to match from sub instead of
xor. Maybe instead ISD::ABS should be made legal. I'll try that in
parallel to this patch.
This is an alternative to D119099 which was focused on RISCV only.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D119171
This code was detecting whether the value returned by getShiftAmountTy
can represent all shift amounts. If not, it would use MVT::i32 as a
placeholder. getShiftAmountTy was updated last year to return i32
if the type returned by the target couldn't represent all values.
This means the MVT::i32 case here is dead and can the logic can
be simplified.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D120164
The code was considering shifts by an about larger than the number of
bits in the original VT to be out of range. Shifts exactly equal to
the original bit width are also out of range.
I don't know how to test this. DAGCombiner should usually fold this
away. I just noticed while looking for something else in this code. The
llvm-cov report shows that we don't have coverage for out of range shifts here.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D120170
getShiftAmountTy will return MVT::i32 if the shift amount
coming from the target's getScalarShiftAmountTy can't reprsent
all possible values. That should eliminate the need to use the
pointer type which is what we do when LegalTypes is false.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D120165
This fold is done in IR:
https://alive2.llvm.org/ce/z/jWyFrP
There is an x86 test that shows an improvement
from the added flexibility of using add (commutative).
The other diffs are presumed neutral.
Note that this could also be folded to an 'xor',
but I'm not sure if that would be universally better
(eg, x86 can convert adds more easily into LEA).
This helps prevent regressions from a potential fold for
issue #53829.
The current ABD combine doesn't quite work for SVE because only a
single scalable vector per scalar integer type is legal (e.g. for
i32, <vscale x 4 x i32> is the only legal scalable vector type).
This patch extends the combine to also trigger for the cases when
operand extension must be retained.
Differential Revision: https://reviews.llvm.org/D115739
When doing SelectionDAG::ReplaceAllUsesOfValuesWith a worklist is
prepared containing all users that should be updated. Then we use
the RemoveNodeFromCSEMaps/AddModifiedNodeToCSEMaps helpers to handle
recursive CSE updates while doing the replacements.
This patch aims at solving a problem that could arise if the recursive
CSE updates would result in an SDNode present in the worklist is being
removed as a side-effect of morphing a prio user in the worklist.
To examplify such a scenario, imagine that we have these nodes in
the DAG
t12: i64 = add t8, t11
t13: i64 = add t12, t8
t14: i64 = add t11, t11
t15: i64 = add t14, t8
t16: i64 = sub t13, t15
and that the t8 uses should be replaced by t11. An initial worklist
(listing the users that should be morphed) could be [t12, t13, t15].
When updating t12 we get
t12: i64 = add t11, t11
which results in a CSE update that replaces t14 by t12, so we get
t15: i64 = add t12, t8
which results in a CSE update that replaces t13 by t12, so we get
t16: i64 = sub t12, t15
and then t13 is removed given that it was the last use of t13.
So when being done with the updates triggered by rewriting the use
of t8 in t12 the t13 node no longer exist. And we used to end up
hitting an assertion when continuing with the worklist aiming at
replacing the t8 uses in t13.
The solution is based on using a DAGUpdateListener, making sure that
we prune a user from the worklist if it is removed during the
recursive CSE updates.
The bug was found using an OOT target. I think the problem is quite
old, even if the particular intree target reproducer added in this
patch seem to pass when using LLVM 13.0.0.
Differential Revision: https://reviews.llvm.org/D119088
This moves the matching of AVGFloor and AVGCeil into a place where
demand bit are available, so that it can detect more cases for more
folds. It changes the transform to start from a shift, not from a
truncate. We match the pattern shr(add(ext(A), ext(B)), 1), transforming
to ext(hadd(A, B)).
For signed values, because only the bottom bits are demanded llvm will
transform the above to use a lshr too, as opposed to ashr. In order to
correctly detect the hadd we need to know the demanded bits to turn it
back. Depending on whether the shift is signed (ashr) or logical (lshr),
and the extensions are signed or unsigned we can create different nodes.
If the shift is signed:
Needs >= 2 sign bits. https://alive2.llvm.org/ce/z/h4gQAW generating signed rhadd.
Needs >= 2 zero bits. https://alive2.llvm.org/ce/z/B64DUA generating unsigned rhadd.
If the shift is unsigned:
Needs >= 1 zero bits. https://alive2.llvm.org/ce/z/ByD8sj generating unsigned rhadd.
Needs 1 demanded bit zero and >= 2 sign bits https://alive2.llvm.org/ce/z/hvPGxX and
https://alive2.llvm.org/ce/z/32P5n1 generating signed rhadd.
Differential Revision: https://reviews.llvm.org/D119072
This adds very basic combines for AVG nodes, mostly for constant folding
and handling degenerate (zero) cases. The code performs mostly the same
transforms as visitMULHS, adjusted for AVG nodes.
Constant folding extends to a higher bitwidth and drops the lowest bit.
For undef nodes, `avg undef, x` is transformed to x. There is also a
transform for `avgfloor x, 0` transforming to `shr x, 1`.
Differential Revision: https://reviews.llvm.org/D119559
The current FastISel code reuses the register for a bitcast that
doesn't change the IR type, but uses a reg-to-reg copy if it
changes the IR type without changing the MVT. However, we can
simply reuse the register in that case as well.
In particular, this avoids unnecessary reg-to-reg copies for pointer
bitcasts. This was found while inspecting O0 codegen differences
between typed and opaque pointers.
Differential Revision: https://reviews.llvm.org/D119432
This enables fshl to be matched earlier on X86
%6 = lshr i32 %3, 1
%7 = select i1 %4, i32 -2147483648, i32 0
%8 = or i32 %6, %7
X86 uses i8 for shift amounts. SelectionDAGBuilder creates the
ISD::SRL with an i8 shift type. DAGCombiner turns the select into
an ISD::SHL. Prior to this patch it would use i32 for the shift
amount. fshl matching failed because the shift amounts have different
types. LegalizeDAG fixes the ISD::SHL shift amount to i8. This
allowed fshl matching to succeed.
With this patch, the ISD::SHL will be created with an i8 shift
amount. This allows the fshl to match immediately.
No test case beause we still end up with a fshl either way.
I have not found a way to expose a difference for this patch in a test
because it only triggers for a one-use load, but this is the code that
was adapted into D118376 and caused miscompiles. The new code pattern
is the same as what we do in narrowExtractedVectorLoad() (reduces load
width for a subvector extract).
This removes seemingly unnecessary manual worklist management and fixes
the chain updating via "SelectionDAG::makeEquivalentMemoryOrdering()".
Differential Revision: https://reviews.llvm.org/D119549
This ports the aarch64 combines for HADD and RHADD over to DAG combine,
so that they can be used in more architectures (notably MVE in a
followup patch). They are renamed to AVGFLOOR and AVGCEIL in the
process, to avoid confusion with instructions such as X86 hadd. The code
was also rewritten slightly to remove the AArch64 idiosyncrasies.
The general pattern for a AVGFLOORS is
%xe = sext i8 %x to i32
%ye = sext i8 %y to i32
%a = add i32 %xe, %ye
%r = lshr i32 %a, 1
%t = trunc i32 %r to i8
An AVGFLOORU is equivalent with zext. Because of the truncate
lshr==ashr, as the top bits are not demanded. An AVGCEIL also includes
an extra rounding, so includes an extra add of 1.
Differential Revision: https://reviews.llvm.org/D106237
Add a new llvm.fptrunc.round intrinsic to precisely control
the rounding mode when converting from f32 to f16.
Differential Revision: https://reviews.llvm.org/D110579
We're hitting a pathological compile-time case, profiled to be in
DagCombiner::visitTokenFactor and many inserts into a SmallPtrSet.
It looks like one of the paths around findBetterNeighborChains is not
capped and leads to this.
This patch resolves the issue. Looking for feedback if this solution
looks reasonable.
Differential Revision: https://reviews.llvm.org/D118877
The test diffs are identical to D119111.
This only affects x86 currently because no other target
has an override for the TLI hook that controls this transform.
This is no-functional-change-intended because only the
x86 target enables the TLI hook currently.
We can add fmul/fdiv opcodes to the switch similar to the
proposal D119111, but we don't need to make other changes
like enabling target-specific combines.
We can also add integer opcodes (add, or, shl, etc.) to
the switch because this function is called from all of the
generic binary opcodes.
The goal is to incrementally enable the profitable diffs
from D90113 while avoiding regressions.
Differential Revision: https://reviews.llvm.org/D119150
In many cases, calls to isShiftedMask are immediately followed with checks to determine the size and position of the bitmask.
This patch adds variants of APInt::isShiftedMask, isShiftedMask_32 and isShiftedMask_64 that return these values as additional arguments.
I've updated a number of cases that were either performing seperate size/position calculations or had created their own local wrapper versions of these.
Differential Revision: https://reviews.llvm.org/D119019
rv64izbb has a RORW/ROLW instructions that operate on the lower
32-bits of a 64-bit value and sign extend bit 31 of the result.
DAGCombiner won't match rotate idioms because the i32 type isn't Legal
on riscv64.
This patch teaches DAGCombiner to allow it if the type is going to
be promoted and the target has Custom type legalization for ISD::ROTL
or ISD::ROTR. I've restricted this to scalar types. It doesn't appear
any in tree targets other than riscv64 have custom type legalization
for rotates.
If this patch isn't acceptable, I guess I can match SRLW, SLLW, and OR
after type legalization, but I'd like to avoid that if possible.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D119062
When the shift amount is known and a known sign bit analysis of
the shiftee indicates that no saturation will occur, then we can
replace SSHLSAT/USHLSAT by SHL.
Differential Revision: https://reviews.llvm.org/D118765
These operations are scalarized but the result type v1i1 isn't which
needs special handling (the same as is done for the non-strict
versions of these operations).
Differential Revision: https://reviews.llvm.org/D118258
This header is very large (3M Lines once expended) and was included in location
where dwarf-specific information were not needed.
More specifically, this commit suppresses the dependencies on
llvm/BinaryFormat/Dwarf.h in two headers: llvm/IR/IRBuilder.h and
llvm/IR/DebugInfoMetadata.h. As these headers (esp. the former) are widely used,
this has a decent impact on number of preprocessed lines generated during
compilation of LLVM, as showcased below.
This is achieved by moving some definitions back to the .cpp file, no
performance impact implied[0].
As a consequence of that patch, downstream user may need to manually some extra
files:
llvm/IR/IRBuilder.h no longer includes llvm/BinaryFormat/Dwarf.h
llvm/IR/DebugInfoMetadata.h no longer includes llvm/BinaryFormat/Dwarf.h
In some situations, codes maybe relying on the fact that
llvm/BinaryFormat/Dwarf.h was including llvm/ADT/Triple.h, this hidden
dependency now needs to be explicit.
$ clang++ -E -Iinclude -I../llvm/include ../llvm/lib/Transforms/Scalar/*.cpp -std=c++14 -fno-rtti -fno-exceptions | wc -l
after: 10978519
before: 11245451
Related Discourse thread: https://llvm.discourse.group/t/include-what-you-use-include-cleanup
[0] https://llvm-compile-time-tracker.com/compare.php?from=fa7145dfbf94cb93b1c3e610582c495cb806569b&to=995d3e326ee1d9489145e20762c65465a9caeab4&stat=instructions
Differential Revision: https://reviews.llvm.org/D118781
In the aftermath of D116895 a problem was found in the analysis of
dependencies between store merge candidates in
checkMergeStoreCandidatesForDependencies, that is needed to avoid
the cycles are introduced in the DAG.
In the past it has been enough (or assumed to be enough) to start
scanning from non-chain operands when analysing the store merge
candidates for dependencies, assuming that the analysis of chain
dependencies performed when finding the candidates would cover
up for potential dependencies that exist involving the chain operands.
It was however discovered that one could end up with scenarios such
as descibed in the aarch64-checkMergeStoreCandidatesForDependencies.ll
test case, when the dependency between two stores is given by a mix
of chain operand dependencies and non-chain operand dependencies.
The fix in this patch make sure that we also account for chain operand
dependencies when doing the more elaborate analysis in
checkMergeStoreCandidatesForDependencies, no longer relying on that
the earlier check involving chain operands is enough.
Differential Revision: https://reviews.llvm.org/D118943
This helps recognise patterns where we're trying to match STEP_VECTOR
patterns to INDEX instructions that take a GPR for the Start/Step.
The reason for canonicalising this operation to the LHS is
because it will already be canonicalised to the LHS if the RHS
is a constant splat vector.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D118459
None of the external users actual touch these (they're purely used internally down the recursive call) - its trivial to add another wrapper if anything ever does want to track known elements.
The new LEGALAVL node annotates that the AVL refers to packs of 64bit.
We use a two-stage lowering approach with LEGALAVL:
First, standard SDNodes are translated into illegal VVP layer nodes.
Regardless of source (VP or standard), all VVP nodes have a mask and AVL
parameter. The AVL parameter refers to the element position (just as in
VP intrinsics).
Second, we legalize the AVL usage in VVP layer nodes. If the element
size is < 64bit, the EVL parameter has to be adjusted to refer to packs
of 64bits. We wrap the legalized AVL in a LEGALAVL node to track this.
Reviewed By: kaz7
Differential Revision: https://reviews.llvm.org/D118321
Do "simplifyShift" and "FoldConstantArithmetic" folds for the SSHLSAT
and USHLSAT DAG nodes.
This includes folds such as:
(shlsat undef/poison, x) -> 0
(shlsat x, undef/poison) -> undef
(shlsat x, too_large_shamt) -> undef
(shlsat 0, x) -> 0
(shlsat x, 0) -> x
(shlsat c1, c2) -> c3
Differential Revision: https://reviews.llvm.org/D118603
I have updated TargetLowering::isConstTrueVal to also consider
SPLAT_VECTOR nodes with constant integer operands. This allows the
optimisation to also work for targets that support scalable vectors.
Differential Revision: https://reviews.llvm.org/D117210
For the cross block gc.result projection case, we only care about the return type if there is a cross block gc.result, and if there is one, we can take the type from the gc.result.
At the moment, this makes little difference, but for opaque pointers we need a means to get result typing without relying on pointee types.
When lowering a gc.result, we can assume that the result type of the gc.result matches the type of the underlying call. This is explicitly required in LangRef.
At the moment, this makes little difference, but for opaque pointers we need a means to get result typing without relying on pointee types.
Fixes a crash ('Invalid size request on a scalable vector') in visitAlloca()
when we call this function for a scalable alloca instruction, caused
by the implicit conversion of TySize to uint64_t.
This patch changes TySize to a TypeSize as returned by getTypeAllocSize()
and ensures the allocation size is multiplied by vscale for scalable vectors.
Reviewed By: sdesmalen, david-arm
Differential Revision: https://reviews.llvm.org/D118372
We already call SimplifyDemandedVectorElts using whether each vector mask element is zero/nonzero, this just extends this to also try SimplifyDemandedBits using the demanded bits mask generated from the nonzero elements.
This also requires an additional TargetLowering::SimplifyDemandedBits DemandedBits/DemandedElts wrapper.
If we have a vector FP division with a splatted divisor, use
getVectorMinNumElements when scaling the num of uses by splat factor.
For AArch64 the combine kicks in for the <vscale x 4 x float> case since it's
above the fdiv threshold (3) when scaling num uses by splat factor, but the
codegen is worse (splat + vector fdiv + vector fmul) than the <vscale x 2 x
double> case (splat + vector fdiv).
If the combine could be converted into a scalar FP division by
scalarizeBinOpOfSplats it may be cheaper, but it looks like this is predicated
on the isExtractVecEltCheap TLI function which is implemented for x86 but not
AArch64. Perhaps for now combineRepeatedFPDivisors should only scale num uses
by splat if the division can be converted into scalar op.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D118343
Use the llvm flag `-pgo-function-entry-coverage` to create single byte "counters" to track functions coverage. This mode has significantly less size overhead in both code and data because
* We mark a function as "covered" with a store instead of an increment which generally requires fewer assembly instructions
* We use a single byte per function rather than 8 bytes per block
The trade off of course is that this mode only tells you if a function has been covered. This is useful, for example, to detect dead code.
When combined with debug info correlation [0] we are able to create an instrumented Clang binary that is only 150M (the vanilla Clang binary is 143M). That is an overhead of 7M (4.9%) compared to the default instrumentation (without value profiling) which has an overhead of 31M (21.7%).
[0] https://groups.google.com/g/llvm-dev/c/r03Z6JoN7d4
Reviewed By: kyulee
Differential Revision: https://reviews.llvm.org/D116180
We already perform some basic folds (add/sub with zero etc.) on scalar types, this patch adds some basic support for constant splats as well in a few cases (we can add more with future test coverage).
In the cases I've enabled, we can handle buildvector implicit truncation as we're not creating new constant nodes from the vector types - we're just returning existing nodes. This allows us to get a number of extra cases in the aarch64 tests.
I haven't enabled support for undefs in buildvector splats, as we're often checking for zero/allones patterns that return the original constant and we shouldn't be returning undef elements in some of these cases - we can enable this later if we're OK with creating new constants.
Differential Revision: https://reviews.llvm.org/D118264
This patch adds support for expanding VP_MERGE through a sequence of
vector operations producing a full-length mask setting up the elements
past EVL/pivot to be false, combining this with the original mask, and
culminating in a full-length vector select.
This expansion should work for any data type, though the only use for
RVV is for boolean vectors, which themselves rely on an expansion for
the VSELECT.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D118058
This reverts commit ef82063207.
- It conflicts with the existing llvm::size in STLExtras, which will now
never be called.
- Calling it without llvm:: breaks C++17 compat
The loop below the changed line assumes that the element
width of the target constant is the same as the element
width of the loaded value, but that is not always true.
We could try harder to do some kind of min/max calc even
if the sizes don't match, but that can be another patch
if needed. This fixes#53401 (miscompile) and does not
change the motivating cases added when this analysis
was introduced:
ad298f86b7
Currently not (xor_one_use) pattern is always selected to S_XNOR irrelative od the node divergence.
This relies on further custom selection pass which converts to VALU if necessary and replaces with V_NOT_B32 ( V_XOR_B32)
on those targets which have no V_XNOR.
Current change enables the patterns which explicitly select the not (xor_one_use) to appropriate form.
We assume that xor (not) is already turned into the not (xor) by the combiner.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D116270
This is the unsigned variant of D111976, where we convert a clamped
fptoui to a fptoui.sat. Because we are unsigned, the condition this time
is only UMIN of UINT_MAX. Similarly to D111976 it handles ISD::UMIN,
ISD::SETCC/ISD::SELECT, ISD::VSELECT or ISD::SELECT_CC nodes.
This especially helps on ARM/AArch64 where the vcvt instructions
naturally saturate the result.
Differential Revision: https://reviews.llvm.org/D114964
An sra is basically sign-extending a narrower value. Fold away the
shift by doing a sextload of a narrower value, when it is legal to
reduce the load width accordingly.
Differential Revision: https://reviews.llvm.org/D116930
This patch adds widening support for ISD::VP_MERGE, which widens
identically to VP_SELECT and similarly to other select-like nodes.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D118030
This patch adds splitting support for ISD::VP_MERGE, which splits
identically to VP_SELECT and similarly to other select-like nodes.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D118032
Split these nodes in a similar way as their masked versions.
Reviewed By: frasercrmck, craig.topper
Differential Revision: https://reviews.llvm.org/D117760
Given that step_vector is practically a constant, doing this early
helps with DAGCombine folds that happen before type legalization.
There is currently no way to test this happens earlier, although existing
tests for step_vector folds continue protect the folds happening at all.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D117863
If the bitreverse gets expanded, it will introduce a new bswap. By
putting a bswap before the bitreverse, we can ensure it gets cancelled
out when this happens.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D118012
This can show up during when bitreverse is expanded to bswap and
swap of bits within a byte. If the input is already a bswap, we
should cancel them out before we further transform them in a way
that makes it harder to see the redundancy.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D118007
In code review for D117104 two slightly weird checks were found
in DAGCombiner::reduceLoadWidth. They were typically checking
if BitsA was a mulitple of BitsB by looking at (BitsA & (BitsB - 1)),
but such a comparison actually only make sense if BitsB is a power
of two.
The checks were related to the code that attempted to shrink a load
based on the fact that the loaded value would be right shifted.
Afaict the legality of the value types is checked later (typically in
isLegalNarrowLdSt), so the existing checks were both overly
conservative as well as being wrong whenever ExtVTBits wasn't a
power of two. The latter was a situation triggered by a number of
lit tests so we could not just assert on ExtVTBIts being a power of
two).
When attempting to simply remove the checks I found some problems,
that seems to have been guarded by the checks (maybe just out of
luck). A typical example would be a pattern like this:
t1 = load i96* ptr
t2 = srl t1, 64
t3 = truncate t2 to i64
When DAGCombine is visiting the truncate reduceLoadWidth is called
attempting to narrow the load to 64 bits (ExtVT := MVT::i64). Then
the SRL is detected and we set ShAmt to 64.
In the past we've bailed out due to i96 not being a multiple of 64.
If we simply remove that check then we would end up replacing the
load with a new load that would read 64 bits but with a base pointer
adjusted by 64 bits. So we would read 32 bits the wasn't accessed by
the original load.
This patch will instead utilize the fact that the logical left shift
can be folded away by using a zextload. Thus, the pattern above will
now be combined into
t3 = load i32* ptr+offset, zext to i64
Another case is shown in the X86/shift-folding.ll test case:
t1 = load i32* ptr
t2 = srl i32 t1, 8
t3 = truncate t2 to i16
In the past we bailed out due to the shift count (8) not being a
multiple of 16. Now the narrowing kicks in and we get
t3 = load i16* ptr+offset
Differential Revision: https://reviews.llvm.org/D117406
EmitSchedule() shouldn't be touching instructions after the provided
insertion point. The change introduced in D83561 performs a scan to
the end of the block, and thus may move unrelated instructions. In
particular, this ends up moving instructions that have been produced
by FastISel and will later be deleted. Moving them means that more
instructions than intended are removed.
Fix this by stopping the iteration when the insertion point is
reached.
Fixes https://github.com/llvm/llvm-project/issues/53243.
Differential Revision: https://reviews.llvm.org/D117489
SelectionDAG::getNode() canonicalises constants to the RHS if the
operation is commutative, but it doesn't do so for constant splat
vectors. Doing this early helps making certain folds on vector types,
simplifying the code required for target DAGCombines that are enabled
before Type legalization.
Somewhat to my surprise, DAGCombine doesn't seem to traverse the
DAG in a post-order DFS, so at the time of doing some custom fold where
the input is a MUL, DAGCombiner::visitMUL hasn't yet reordered the
constant splat to the RHS.
This patch leads to a few improvements, but also a few minor regressions,
which I traced down to D46492. When I tried reverting this change to see
if the changes were still necessary, I ran into some segfaults. Not sure
if there is some latent bug there.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D117794
Fixes parity codegen issue where we know all but the lowest bit is zero, we can replace the ICMPNE with 0 comparison with a ext/trunc
Differential Revision: https://reviews.llvm.org/D117983
Fixes parity codegen issue where we know all but the lowest bit is zero, we can replace the ICMPNE with 0 comparison with a ext/trunc
Differential Revision: https://reviews.llvm.org/D117983
Pulled out of D106237, this folds truncstore(extend(x)) back to store(x)
if the original store was legal. This can come up due to the order we
fold nodes. A fold from X86 needs to be adjusted to prevent infinite
loops, to have it pick the operand of a trunc more directly.
Differential Revision: https://reviews.llvm.org/D117901
This extends the code in SearchForAndLoads to be able to look through
ANY_EXTEND nodes, which can be created from mismatching IR types where
the AND node we begin from only demands the low parts of the register.
That turns zext and sext into any_extends as only the low bits are
demanded. To be able to look through ANY_EXTEND nodes we need to handle
mismatching types in a few places, potentially truncating the mask to
the size of the final load.
Recommitted with a more conservative check for the type of the extend.
Differential Revision: https://reviews.llvm.org/D117457
This patch fixes a case where the 'align' parameter attribute on the
pointer operands to llvm.vp.gather and llvm.vp.scatter was being dropped
during the conversion to the SelectionDAG. The default alignment equal
to the ABI type alignment of the vector type was kept. It also updates
the documentation to reflect the fact that the parameter attribute is
now properly supported.
The default alignment of these intrinsics was previously documented as
being equal to the ABI alignment of the *scalar* type, when in fact that
wasn't the case: the ABI alignment of the vector type was used instead.
This has also been fixed in this patch.
Reviewed By: simoll, craig.topper
Differential Revision: https://reviews.llvm.org/D114423
This was noted as a potential cleanup in D117508.
getShiftAmountTy() has checks for vector, phase, etc. so it should
handle anything that the caller was trying to account for.
When widening these intrinsics, we do not have to insert neutral
elements at the end of the vector as when widening vector.reduce.*
intrinsics, thanks to vector predication semantics.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D117467
This caused builds to fail with
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:5638:
bool (anonymous namespace)::DAGCombiner::BackwardsPropagateMask(llvm::SDNode *):
Assertion `NewLoad && "Shouldn't be masking the load if it can't be narrowed"' failed.
See the code review for a link to a reproducer.
> This extends the code in SearchForAndLoads to be able to look through
> ANY_EXTEND nodes, which can be created from mismatching IR types where
> the AND node we begin from only demands the low parts of the register.
> That turns zext and sext into any_extends as only the low bits are
> demanded. To be able to look through ANY_EXTEND nodes we need to handle
> mismatching types in a few places, potentially truncating the mask to
> the size of the final load.
>
> Differential Revision: https://reviews.llvm.org/D117457
This reverts commit 578008789f.
Split vp.reduction.* intrinsics by splitting the vector to reduce in
two halves, perform the reduction operation in each one of them and
accumulate the results of both operations.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D117469
A possible codegen regression for PowerPC is noted in D117406
because we don't recognize a pattern that demands only 1 byte
from a bswap.
This fold has existed in IR since close to the beginning of LLVM:
https://github.com/llvm/llvm-project/blame/main/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp#L794
...so this patch copies that code as much as possible and adapts
it for SDAG.
The test for PowerPC that would change in D117406 is over-reduced
with undefs, so I recreated it for AArch64 and x86 by passing in
pointer args and renamed the values to make the logic clearer.
Differential Revision: https://reviews.llvm.org/D117508
This extends the code in SearchForAndLoads to be able to look through
ANY_EXTEND nodes, which can be created from mismatching IR types where
the AND node we begin from only demands the low parts of the register.
That turns zext and sext into any_extends as only the low bits are
demanded. To be able to look through ANY_EXTEND nodes we need to handle
mismatching types in a few places, potentially truncating the mask to
the size of the final load.
Differential Revision: https://reviews.llvm.org/D117457
When we know the value we're extending is a negative constant then it
makes sense to use SIGN_EXTEND because this may improve code quality in
some cases, particularly when doing a constant splat of an unpacked vector
type. For example, for SVE when splatting the value -1 into all elements
of a vector of type <vscale x 2 x i32> the element type will get promoted
from i32 -> i64. In this case we want the splat value to sign-extend from
(i32 -1) -> (i64 -1), whereas currently it zero-extends from
(i32 -1) -> (i64 0xFFFFFFFF). Sign-extending the constant means we can use
a single mov immediate instruction.
New tests added here:
CodeGen/AArch64/sve-vector-splat.ll
I believe we see some code quality improvements in these existing
tests too:
CodeGen/AArch64/reduce-and.ll
CodeGen/AArch64/unfold-masked-merge-vector-variablemask.ll
The apparent regressions in CodeGen/AArch64/fast-isel-cmp-vec.ll only
occur because the test disables codegen prepare and branch folding.
Differential Revision: https://reviews.llvm.org/D114357
Update code comments in DAGCombiner::ReduceLoadWidth and refactor
the handling of SRL a bit. The refactoring is done with the intent
of adding support for folding away SRA by using SEXTLOAD in a
follow-up patch.
The function is also renamed as DAGCombiner::reduceLoadWidth.
Differential Revision: https://reviews.llvm.org/D117104
Use the AttributeSet constructor instead. There's no good reason
why AttrBuilder itself should exact the AttributeSet from the
AttributeList. Moving this out of the AttrBuilder generally results
in cleaner code.
Original patch by @hussainjk.
This patch was split off from D109377 to keep vector legalization
(widening/splitting) separate from vector element legalization
(promoting).
While the original patch added a third overload of
SelectionDAG::getVPStore, this patch takes the liberty of collapsing
those all down to 1, as three overloads seems excessive for a
little-used node.
The original patch also used ModifyToType in places, but that method
still crashes on scalable vector types. Seeing as the other VP
legalization methods only work when all operands need identical
widening, this patch follows in that vein.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D117235
This seems to be a leftover from a long time ago when there was
an ISD::VBIT_CONVERT and a MVT::Vector. It looks like in those days
the vector type was carried in a VTSDNode.
As far as I know, these days ComputeValueTypes would have already
assigned "Result" the same type we're getting from TLI.getValueType
here. Thus the BITCAST is always a NOP. Verified by adding an assert
and running check-llvm.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D117335
This commit sometimes causes a crash when compiling a vtable thunk. E.g.:
clang '--target=aarch64-grtev4-linux-gnu' -xc++ - -c -o /dev/null <<EOF
struct a {
virtual int f();
};
struct c {
virtual int &g() const;
};
struct d : a, c {
int &g() const;
};
int &d::g() const {}
EOF
Some follow-up commits have been reverted as well:
Revert "IR: Make getRetAlign check callee function attributes"
Revert "Fix MSVC "32-bit shift implicitly converted to 64 bits" warning. NFC."
Revert "Fix MSVC "32-bit shift implicitly converted to 64 bits" warning. NFC."
This reverts commit 4f414af6a7.
This reverts commit a5507d2e25.
This reverts commit 3d2d208f6a.
This reverts commit 07ddfa95e3.
When we know the value we're extending is a negative constant then it
makes sense to use SIGN_EXTEND because this may improve code quality in
some cases, particularly when doing a constant splat of an unpacked vector
type. For example, for SVE when splatting the value -1 into all elements
of a vector of type <vscale x 2 x i32> the element type will get promoted
from i32 -> i64. In this case we want the splat value to sign-extend from
(i32 -1) -> (i64 -1), whereas currently it zero-extends from
(i32 -1) -> (i64 0xFFFFFFFF). Sign-extending the constant means we can use
a single mov immediate instruction.
New tests added here:
CodeGen/AArch64/sve-vector-splat.ll
I believe we see some code quality improvements in these existing
tests too:
CodeGen/AArch64/dag-numsignbits.ll
CodeGen/AArch64/reduce-and.ll
CodeGen/AArch64/unfold-masked-merge-vector-variablemask.ll
The apparent regressions in CodeGen/AArch64/fast-isel-cmp-vec.ll only
occur because the test disables codegen prepare and branch folding.
Differential Revision: https://reviews.llvm.org/D114357
This adds support for STRICT_FSETCC(quiet) and STRICT_FSETCCS(signaling).
FEQ matches well to STRICT_FSETCC oeq.
FLT/FLE matches well to STRICT_FSETCCS olt/ole.
Others require commuting operands or multiple instructions.
STRICT_FSETCC olt/ole/ogt/oge/ult/ule/ugt/uge uses FLT/FLE,
but we need to save/restore FFLAGS around them to avoid spurious
exceptions. I've implemented pseudo instructions with a
CustomInserter to insert the save/restore CSR instructions.
Unfortunately, this doesn't honor exceptions for signaling NANs
but I'm not sure if signaling nans are really supported by the
constrained intrinsics.
STRICT_FSETCC one and ueq expand to a pair of FLT instructions
with a save/restore of fflags around each. This could be improved
in the future.
There may be some opportunities to generate better code for strict
comparisons mixed with nonans fast math flags. I've left FIXMEs in
the .td files for that.
Co-Authored-by: ShihPo Hung <shihpo.hung@sifive.com>
Reviewed By: arcbbb
Differential Revision: https://reviews.llvm.org/D116694
Completely rework how we handle X constrained labels for inline asm.
X should really be treated as i. Then existing tests can be moved to use
i D115410 and clang can just emit i D115311. (D115410 and D115311 are
callbr, but this can be done for label inputs, too).
Coincidentally, this simplification solves an ICE uncovered by D87279
based on assumptions made during D69868.
This is the third approach considered. See also discussions v1 (D114895)
and v2 (D115409).
Reported-by: kernel test robot <lkp@intel.com>
Fixes: https://github.com/ClangBuiltLinux/linux/issues/1512
Reviewed By: void, jyknight
Differential Revision: https://reviews.llvm.org/D115688
I've changed the definition of the experimental.vector.splice
instrinsic to reject indices that are known to be or possibly
out-of-bounds. In practice, this means changing the definition so that
the index is now only valid in the range [-VL, VL-1] where VL is the
known minimum vector length. We use the vscale_range attribute to
take the minimum vscale value into account so that we can permit
more indices when the attribute is present.
The splice intrinsic is currently only ever generated by the vectoriser,
which will never attempt to splice vectors with out-of-bounds values.
Changing the definition also makes things simpler for codegen since we
can always assume that the index is valid.
This patch was created in response to review comments on D115863
Differential Revision: https://reviews.llvm.org/D115933
This commit fixes a missed opportunity in merging consecutive stores.
The code that searches for stores skipped the case of stores that
directly connect to the root. The comment above the implementation lists
this case but the code did not handle it. I found this pattern when
looking into the shared_ptr destructor. GCC generates the right
sequence. Here is a small repo:
int foo(int* buff) {
buff[0] = 0;
int x = buff[1];
buff[1] = 0;
return x;
}
Differential Revision: https://reviews.llvm.org/D116895
These nodes should saturate to their saturating VT. We can use this
information to know the bits past the VT are all zeros or all sign bits.
I think we might only have test coverage for the unsigned case. I'll
verify and add tests.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D116870
This is the last part of D116531. Fetch the type of the indirect
inline asm operand from the elementtype attribute, rather than
the pointer element type.
Fixes https://github.com/llvm/llvm-project/issues/52928.
Split vp.select in a similar way as vselect, splitting also the length
parameter.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D116651
We can either check the opcode or number of operands or use
ISD::isVPOpcode inside the methods.
In some places I've used number of operands figuring that it is
cheaper than isVPOpcode. I've included isVPOpcode in an assert to
verify.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D116578
Unsigned compares work with either zero extended or sign extended
inputs just like equality comparisons. I didn't allow this when
I refactored the code in D116421 due to lack of tests. But I've
since found a simple C test case that demonstrates when this can be
useful.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D116617
Use the VPIntrinsics.def's LEGALPOS that is specified with every VP
SDNode to determine which return or operand value type shall be used to
infer the legalization action.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D116594
This function returns an upper bound on the number of bits needed
to represent the signed value. Use "Max" to match similar functions
in KnownBits like countMaxActiveBits.
Rename APInt::getMinSignedBits->getSignificantBits. Keeping the old
name around to keep this patch size down. Will do a bulk rename as
follow up.
Rename KnownBits::countMaxSignedBits->countMaxSignificantBits.
Reviewed By: lebedev.ri, RKSimon, spatel
Differential Revision: https://reviews.llvm.org/D116522
This is similar to what is done for targets that prefer zero extend
where we avoid using a zero extend if the promoted values are sign
extended.
We'll also check for zero extended operands for ugt, ult, uge, and ule when the
target prefers sign extend. This is different than preferring zero extend, where
we only check for sign bits on equality comparisons.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D116421
The 'New' only makes sense in the context of these being
output arguments, but they are also used as inputs first.
Drop the 'New' and just call them LHS/RHS.
Factored out of D116421.
This patch adds isel support for STRICT_LRINT/LLRINT/LROUND/LLROUND.
It also adds test cases for f32 and f64 constrained intrinsics that
correspond to the intrinsics in float-intrinsics.ll and
double-intrinsics.ll. Support for promoting the integer argument of
STRICT_FPOWI was added.
I've skipped adding tests for f16 intrinsics, since we don't have libcalls
for them and we have inconsistent support for promoting them in LegalizeDAG.
This will need to be examined more closely.
Reviewed By: asb
Differential Revision: https://reviews.llvm.org/D116323
getShiftAmountTy used to directly return the shift amount type from
the target which could be too small for large illegal types. For
example, X86 always returns i8.
The code here detected this and used i32 instead if it won't fit. This
behavior was added to getShiftAmountTy in D112469 so we no longer need
this workaround.
Fix issue in TargetLowering::expandROT where we only attempt to flip a rotation if the other direction has better support - this matches TargetLowering::expandFunnelShift
This allows us to enable ISD::ROTR lowering on SSE targets, which particularly simplifies/improves codegen for splat amount and AVX2 per-element shifts.
When the source has a series of assignments, users reasonably want to
have the debugger step through each one individually. Turn off the combine
for adjacent stores so we get this behavior at -O0.
Similar to D7181.
Reviewed By: spatel, xgupta
Differential Revision: https://reviews.llvm.org/D115808
With Control-Flow Integrity (CFI), the LowerTypeTests pass replaces
function references with CFI jump table references, which is a problem
for low-level code that needs the address of the actual function body.
For example, in the Linux kernel, the code that sets up interrupt
handlers needs to take the address of the interrupt handler function
instead of the CFI jump table, as the jump table may not even be mapped
into memory when an interrupt is triggered.
This change adds the no_cfi constant type, which wraps function
references in a value that LowerTypeTestsModule::replaceCfiUses does not
replace.
Link: https://github.com/ClangBuiltLinux/linux/issues/1353
Reviewed By: nickdesaulniers, pcc
Differential Revision: https://reviews.llvm.org/D108478
When the source has a series of assignments, users reasonably want to
have the debugger step through each one individually. Turn off the combine
for adjacent stores so we get this behavior at -O0.
Similar to D7181.
Differential Revision: https://reviews.llvm.org/D115808
Merge the node combines into a common DAGCombiner::visitFMinMax (like we do for IMINMAX).
Move the constant folding into SelectionDAG::foldConstantFPMath.
This allows us to fold the vecreduce-propagate-sd-flags.ll test as it reduces constants - so I've refactored it to take variables instead.
Differential Revision: https://reviews.llvm.org/D115952
We were using a function attribute to indicate a non-standard FP mode,
but now we can use intrinsics for that job as shown in the new tests.
Presumably the x86 asm could be improved for that IR with intrinsics,
but I have not worked out exactly how to do that. Note that the
transform to FTRUNC still requires a hacky check for "nsz" (because
FMF are not applied to FP casts).
This is a cleanup based on the clang change in D115804 / 8c7f2a4f87 .
This is effectively a revert of 5a90285bd9 + D46237 .
Differential Revision: https://reviews.llvm.org/D115885
Replace custom constant scalar/splat folding with FoldConstantArithmetic call and canonicalize commutative constant ops to the RHS before the SimplifyVBinOp call
This extends the custom lowering for truncating stores on
fixed length vectors in SVE to support masked truncating stores.
It also adds a DAG combine for truncates followed by masked
stores.
Reviewed By: peterwaller-arm, paulwalker-arm
Differential Revision: https://reviews.llvm.org/D108115
SimplifyVBinOp still has a FoldConstantArithmetic call, which now it isn't vector specific we should be able to remove (once fp binops are tidied up); but we can at least clean up the integer opcodes to perform the basic constant/undef handling in common code first.
These were detected by the new -Wauto-by-value-copy (D114989) warning, these by-value
constant copies need only be references.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D114990
-(Za + Zm * Zn) != (-Za + Zm * (-Zn))
when the FMA produces a zero output (e.g. all zero inputs can produce -0
output)
Add a PatFrag to check presence of nsz on the fneg, add tests which
ensure the combine does not fire in the absense of nsz.
See https://reviews.llvm.org/D90901 for a similar discussion on X86.
Differential Revision: https://reviews.llvm.org/D109525
This patch fixes an issue during SelectionDAG construction. When the
target is unable to lower the function's return value, a hidden sret
parameter is created. It is initialized and copied to a stored variable
(DemoteRegister) with CopyToReg and is later fetched with
CopyFromReg. The bug is that the chains used for each copy are
inconsistent, and thus in rare cases the scheduler may issue them out of
order.
The fix is to ensure that the CopyFromReg uses the DAG root which is set
as the chain corresponding to the initial CopyToReg.
Fixes https://llvm.org/PR52475
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D114795
Previously we were using UADDO to generate a two-result value with
the unsigned addition and the overflow mask. We then combined the
overflow mask with the trip count comparison to get a result.
However, we don't need to do this - we can simply use a UADDSAT
saturating add node to add the vector index splat and the stepvector
together. Then we can just compare this to a splat of the trip count.
This results in overall better code quality for both Thumb2 and AArch64.
Differential Revision: https://reviews.llvm.org/D115354
There is a pointer to the DataLayout in SelectionDAGBuilder called
'DL' that is hardly ever used. In most cases the code seems to just
use `DAG.getDataLayout()` instead. Given that DL is also often used
as a shadowed variable for the debug location it seems sensible to
just kill off the few remaining uses and be consistent with the rest
of the code.
Differential Revision: https://reviews.llvm.org/D114451
After D113888 / 32b6c17b29 the MMO size of a masked loads/store is
unknown. When we are converting back to a standard load/store because
the mask is known all ones, we can refine that to the correct size from
the size of the vector being loaded/stored.
Differential Revision: https://reviews.llvm.org/D114582
In the style of D113888, this patch updates the various VP memory
operations (load, store, gather, scatter) to use UnknownSize. This is
for the same reason as for masked loads and stores: the number of
elements accessed is not generally known at compile time.
This is somewhat pessimistic in the sense that we may still find
un-canonicalized intrinsics featuring both an all-true mask and an EVL
equal to the vector size. Arguably those should be canonicalized before
the SelectionDAG, so those have been left for future work.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D115036
This patch fixes a case where the 'align' parameter attribute on the
pointer operands to llvm.vp.load and llvm.vp.store was being dropped
during the conversion to the SelectionDAG. The default alignment
equal to the ABI type alignment of the vector type was kept. It also
updates the documentation to reflect the fact that the parameter
attribute is now properly supported.
Reviewed By: simoll
Differential Revision: https://reviews.llvm.org/D114422
As an extension to D111976, this converts clamp fptosi, clamped between
0 and (2^n)-1 to a fptoui.sat. This can greatly help on targets with
conversions that naturally saturate, such as Arm.
X86 disables the transform as some of the test cases increases in size.
A fptoui.sat necessitates a fp clamp without native support, so there is
little use in converting if the instruction is just going to be
expanded.
Differential Revision: https://reviews.llvm.org/D112428
MVE can treat v16i1, v8i1, v4i1 and v2i1 as different views onto the
same 16bit VPR.P0 register, with v2i1 holding two 8 bit values for the
two halves. This was never treated as a legal type in llvm in the past
as there are not many 64bit instructions and no 64bit compares. There
are a few instructions that could use it though, notably a VSELECT (as
it can handle any size using the underlying v16i8 VPSEL), AND/OR/XOR for
similar reasons, some gathers/scatter and long multiplies and VCTP64
instructions.
This patch goes through and makes v2i1 a legal type, handling all the
cases that fall out of that. It also makes VSELECT legal for v2i64 as a
side benefit. A lot of the codegen changes as a result - usually in way
that is a little better or a little worse, but still expensive. Costs
can change a little too in the process, again in a way that expensive
things remain expensive. A lot of the tests that changed are mainly to
ensure correctness - the code can hopefully be improved in the future
where it comes up in practice.
The intrinsics currently remain using the v4i1 they previously did to
emulate a v2i1. This will be changed in a followup patch but this one
was already large enough.
Differential Revision: https://reviews.llvm.org/D114449
This patch begins extending handling for peeking through bitcast nodes to big-endian targets as well as the existing little-endian case.
Differential Revision: https://reviews.llvm.org/D114676
combinePMULH currently only truncates vXi32/vXi64 multiplies to PMULHW/PMULUW if the source operands are SEXT/ZEXT instructions for a 'free' truncation.
But we can generalize this to any source operand with sufficient leading sign/zero bits that would allow PACKS/PACKUS to be used as a 'cheap' truncation.
This helps us avoid the wider multiplies, in exchange for truncation on both source operands instead of the result.
Differential Revision: https://reviews.llvm.org/D113371
The REM DAG combine uses the visitDivLike functions to try and get an
optimized DIV node to provide better codegen, however in some cases this
visitDivLike call ends up in the BuildSDIVPow2 target hook, which in
turn sometimes will return the same node passed in to indicate not to
change it. The REM DAG combine does not anticipate this and creates a
cycle in the DAG because of it.
Fix this by ensuring any such optimized div node returned is distinct
from the node being combined.
Differential Revision: https://reviews.llvm.org/D114716
This adds a fold in DAGCombine to create fptosi_sat from sequences for
smin(smax(fptosi(x))) nodes, where the min/max saturate the output of
the fp convert to a specific bitwidth (say INT_MIN and INT_MAX). Because
it is dealing with smin(/smax) in DAG they may currently be ISD::SMIN,
ISD::SETCC/ISD::SELECT, ISD::VSELECT or ISD::SELECT_CC nodes which need
to be handled similarly.
A shouldConvertFpToSat method was added to control when converting may
be profitable. The original fptosi will have a less strict semantics
than the fptosisat, with less values that need to produce defined
behaviour.
This especially helps on ARM/AArch64 where the vcvt instructions
naturally saturate the result.
Differential Revision: https://reviews.llvm.org/D111976
It causes builds to fail with this assert:
llvm/include/llvm/ADT/APInt.h:990:
bool llvm::APInt::operator==(const llvm::APInt &) const:
Assertion `BitWidth == RHS.BitWidth && "Comparison requires equal bit widths"' failed.
See comment on the code review.
> This adds a fold in DAGCombine to create fptosi_sat from sequences for
> smin(smax(fptosi(x))) nodes, where the min/max saturate the output of
> the fp convert to a specific bitwidth (say INT_MIN and INT_MAX). Because
> it is dealing with smin(/smax) in DAG they may currently be ISD::SMIN,
> ISD::SETCC/ISD::SELECT, ISD::VSELECT or ISD::SELECT_CC nodes which need
> to be handled similarly.
>
> A shouldConvertFpToSat method was added to control when converting may
> be profitable. The original fptosi will have a less strict semantics
> than the fptosisat, with less values that need to produce defined
> behaviour.
>
> This especially helps on ARM/AArch64 where the vcvt instructions
> naturally saturate the result.
>
> Differential Revision: https://reviews.llvm.org/D111976
This reverts commit 52ff3b0093.
This adds a fold in DAGCombine to create fptosi_sat from sequences for
smin(smax(fptosi(x))) nodes, where the min/max saturate the output of
the fp convert to a specific bitwidth (say INT_MIN and INT_MAX). Because
it is dealing with smin(/smax) in DAG they may currently be ISD::SMIN,
ISD::SETCC/ISD::SELECT, ISD::VSELECT or ISD::SELECT_CC nodes which need
to be handled similarly.
A shouldConvertFpToSat method was added to control when converting may
be profitable. The original fptosi will have a less strict semantics
than the fptosisat, with less values that need to produce defined
behaviour.
This especially helps on ARM/AArch64 where the vcvt instructions
naturally saturate the result.
Differential Revision: https://reviews.llvm.org/D111976
Usually dbg.declares get translated into either entries in an MF
side-table, or a DBG_VALUE on entry to the function with IsIndirect set
(including in instruction referencing mode). Much rarer is a dbg.declare
attached to a non-argument value, such as in the test added in this patch
where there's a variable-length-array. Such dbg.declares become SDDbgValue
nodes with InIndirect=true.
As it happens, we weren't correctly emitting DBG_INSTR_REFs with the
additional indirection. This patch adds the extra indirection, encoded as
adding an additional DW_OP_deref to the expression.
Differential Revision: https://reviews.llvm.org/D114440
This allows the generic DAG combine to fold fp_extend/fp_trunc into
loads/stores which we can then lower into a integer extending
load/truncating store plus an FP_EXTEND/FP_ROUND.
The nuance here is that fixed-type FP_EXTEND/FP_ROUND require unpacked
types hence lowering them introduces an unpack/zip. By allowing these
nodes to be combined with loads/store we make it much easier to have
this unpack/zip combined into the load/store by our custom lowering.
Differential Revision: https://reviews.llvm.org/D114580
In most common cases the @llvm.get.active.lane.mask intrinsic maps directly
to the SVE whilelo instruction, which already takes overflow into account.
However, currently in SelectionDAGBuilder::visitIntrinsicCall we always lower
this immediately to a generic sequence of instructions that explicitly
take overflow into account. This makes it very difficult to then later
transform back into a single whilelo instruction. Therefore, this patch
introduces a new TLI function called shouldExpandGetActiveLaneMask that asks if
we should lower/expand this to a sequence of generic ISD nodes, or instead
just leave it as an intrinsic for the target to lower.
You can see the significant improvement in code quality for some of the
tests in this file:
CodeGen/AArch64/active_lane_mask.ll
Differential Revision: https://reviews.llvm.org/D114542
Currently the generic lowering of llvm.get.active.lane.mask is done
in SelectionDAGBuilder::visitIntrinsicCall and currently assumes
only fixed-width vectors are used. This patch changes the code to be
more generic and support scalable vectors too. I have added tests
for SVE here:
CodeGen/AArch64/active_lane_mask.ll
although the code quality leaves a lot to be desired. The code will
be improved significantly in a later patch that makes use of the
SVE whilelo instruction.
Differential Revision: https://reviews.llvm.org/D114541
If we only demand bits from one half of a rotation pattern, see if we can simplify to a logical shift.
For the ARM/AArch64 rev16/32 patterns, I had to drop a fold to prevent srl(bswap()) -> rotr(bswap) -> srl(bswap) infinite loops. I've replaced this with an isel PatFrag which should do the same task.
Reapplied with fix for AArch64 rev patterns to matching the ARM fix.
https://alive2.llvm.org/ce/z/iroxki (rol -> shl by amt iff demanded bits has at least as many trailing zeros as the shift amount)
https://alive2.llvm.org/ce/z/4ez_U- (ror -> shl by revamt iff demanded bits has at least as many trailing zeros as the reverse shift amount)
https://alive2.llvm.org/ce/z/cD7dR- (ror -> lshr by amt iff demanded bits has at least as many leading zeros as the shift amount)
https://alive2.llvm.org/ce/z/_XGHtQ (rol -> lshr by revamt iff demanded bits has at least as many leading zeros as the reverse shift amount)
Differential Revision: https://reviews.llvm.org/D114354
If we only demand bits from one half of a rotation pattern, see if we can simplify to a logical shift.
For the ARM rev16 patterns, I had to drop a fold to prevent srl(bswap()) -> rotr(bswap) -> srl(bswap) infinite loops. I've replaced this with an isel PatFrag which should do the same task.
https://alive2.llvm.org/ce/z/iroxki (rol -> shl by amt iff demanded bits has at least as many trailing zeros as the shift amount)
https://alive2.llvm.org/ce/z/4ez_U- (ror -> shl by revamt iff demanded bits has at least as many trailing zeros as the reverse shift amount)
https://alive2.llvm.org/ce/z/cD7dR- (ror -> lshr by amt iff demanded bits has at least as many leading zeros as the shift amount)
https://alive2.llvm.org/ce/z/_XGHtQ (rol -> lshr by revamt iff demanded bits has at least as many leading zeros as the reverse shift amount)
Differential Revision: https://reviews.llvm.org/D114354
In quite a few places we were calling getCurSDLoc() to get the debug
location, but this is already a local variable `sdl`.
Differential Revision: https://reviews.llvm.org/D114447
Usage and naming of macros in VPIntrinsics.def has been inconsistent. Rename all property macros to VP_PROPERTY_<name>. Use BEGIN/END scope macros to attach properties to vp intrinsics and SDNodes (instead of specifying either directly with the property macro).
A follow-up patch has documentation on how the macros are (intended) to be used.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D114144
A masked load or store will load a potentially unknown number of bytes
from a memory location - that is not generally known at compile time.
They do not necessarily load/store the entire vector width, and treating
them as such can lead to incorrect aliasing information (for example, if
the underlying object is smaller than the size of the vector).
This makes sure that the MMO is given an unknown size to represent this.
which is less accurate that "may load/store from up to 16 bytes", but
less incorrect that "will load/store from 16 bytes".
Differential Revision: https://reviews.llvm.org/D113888
Patch to fix some of the regressions in D77804.
By folding to rotate/funnel-shift by constant amounts for illegal types, we prevent SimplifyDemandedBits from destroying the patterns prematurely, allowing us to use the rotate/funnel-shift legalization that was added in D112443.
Differential Revision: https://reviews.llvm.org/D113192
Fixed the vector type issue that where we used getVectorNumElements()
should be replaced by getVectorElementCount() when lowering these
intrinsics.
This is similar to D94149
Signed-off-by: Eric Tang <tangxingxin1008@gmail.com>
Reviewed By: craig.topper, frasercrmck
Differential Revision: https://reviews.llvm.org/D109809
Don't expand CTTZ if CTPOP or CTLZ is supported on the promoted type.
We have special handling for CTTZ expansion to use those ops with a
small conversion. The setup for that doesn't generate extra code or
large constants so we don't gain anything from expanding early and we
make CTTZ_ZERO_UNDEF codegen worse.
Follow up from post commit feedback on D112268. We don't seem to have
any in tree tests that care about this.
If we have a large enough floating point type that can exactly
represent the integer value, we can convert the value to FP and
use the exponent to calculate the leading/trailing zeros.
The exponent will contain log2 of the value plus the exponent bias.
We can then remove the bias and convert from log2 to leading/trailing
zeros.
This doesn't work for zero since the exponent of zero is zero so we
can only do this for CTLZ_ZERO_UNDEF/CTTZ_ZERO_UNDEF. If we need
a value for zero we can use a vmseq and a vmerge to handle it.
We need to be careful to make sure the floating point type is legal.
If it isn't we'll continue using the integer expansion. We could split the vector
and concatenate the results but that needs some additional work and evaluation.
Differential Revision: https://reviews.llvm.org/D111904
If we've only demanded the 0'th element, and it comes from a (one-use) AND, try to convert the zero_extend_vector_inreg into a mask and constant fold it with the AND.
This change make WidenVecRes_SELECT work for scalable vectors.
This patch is split from [D110319](https://reviews.llvm.org/D110319)
Signed-off-by: Eric Tang <tangxingxin1008@gmail.com>
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D110388
Fixes PR#48678. `X86TargetLowering::getRegForInlineAsmConstraint()` can adjust the register class to match the type, e.g. change `VR128X` to `VR256X` if the type needs 256 bits. However, the function currently returns the unadjusted register and the adjusted register class, e.g. `xmm15` and `VR256X`, which then causes an assertion failure later because the register class does not contain that register. This patch fixes this behavior.
Reviewed By: pengfei
Differential Revision: https://reviews.llvm.org/D113834
It's possible that the mask is already a NOT. At least if InstCombine
hasn't canonicalized the input. In that case we will form an ANDN with
X instead of with Y. So we don't need to worry about Y being a constant.
We might need to check that X isn't a constant instead, but we don't
have a test case for that yet.
This fixes a size regression found when trying to enable this combine
for RISCV in D113937.
Differential Revision: https://reviews.llvm.org/D113948
This was noted as a follow-up to D113212 / D113426:
4fc1fc40057e30404c3b11522cfcadhttps://alive2.llvm.org/ce/z/e4o96b
The canonicalization rules for these IR patterns are complicated,
and we were not matching the expected forms in 2 out of the 3
cases. We can make codegen more robust by matching the swapped
forms (and that will also work if these patterns are created late).
(Cond0 s> -1) ? N1 : 0 --> ~(Cond0 s>> BW-1) & N1
https://alive2.llvm.org/ce/z/mGCBrd
This was suggested as a potential enhancement in D113212 (also 7e30404c3b ).
There's an improvement for AArch that could be generalized ( X > -1 --> X >= 0 ).
For x86, we have a counter-acting fold for most cases that turns the shift+not
back into a setcc, so that needs a work-around to get more cases to use "pandn":
D113603
Note that this pattern (and a previous one) are not currently canonical forms
in IR:
https://alive2.llvm.org/ce/z/e4o96b
Adding swapped variants is left as a TODO item here, but is planned as
a near-term follow-up patch.
Differential Revision: https://reviews.llvm.org/D113426
Enable FoldConstantArithmetic to constant fold bitcasted constant build vectors. These have typically been bitcasted for type legalization purposes.
By extracting the raw constant bit data, performing the constant fold, and then casting the constant bit data back to the (legalized) type, we can perform constant folding on integer types after legalization.
This in particular helps 32-bit targets which need to handle vXi64 build vectors - during legalization the (unsupported) i64 elements are split to create a bitcasted v2Xi32 build vector.
Addresses some regressions in D113192.
Differential Revision: https://reviews.llvm.org/D113564
The introduction of this legalization, D111248, forgot to replace the
old chain with the new. This could manifest itself in the old
(illegally-typed) value remaining in the DAG, though the simple test
cases didn't catch this.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D113561
NFC refactor of D113351, pulling out the APInt split/merge code from the BuildVectorSDNode bits extraction into a BuildVectorSDNode::recastRawBits helper. This is to allow us to reuse the code when we're packing constant folded APInt data back together.
This patch fixes a compiler crash when widening scalable-vector loads
and stores which end up breaking down to element-wise store operations.
It does so by providing a way for targets with support for
vector-predicated loads and stores to use those instead. By widening the
operation but maintaining the original effective operation length via
the EVL, only the intended vector elements are loaded or stored.
This method should in theory be possible and even preferred for
fixed-length vector types, but all fixed-length types can be broken down
into their elements, and regardless I have observed regressions in the
generated code when doing so. I believe this is simply due to
VP_LOAD/VP_STORE not being up to par with LOAD/STORE in terms of
optimization. It does improve performance on smaller self-contained
examples, however, so the potential is there.
While the only target that benefits from this is RISCV, the legalization
is generic and so was placed centrally.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D111248
This patch merges FoldConstantVectorArithmetic back into FoldConstantArithmetic.
Like FoldConstantVectorArithmetic we now handle vector ops with any operand count, but we currently still only handle binops for scalar types - this can be improved in future patches - in particular some common unary/trinary ops still have poor constant folding.
There's one change in functionality causing test changes - FoldConstantVectorArithmetic bails early if the build/splat vector isn't all constant (with some undefs) elements, but FoldConstantArithmetic doesn't - it instead attempts to fold the scalar nodes and bails if they fail to regenerate a constant/undef result, allowing some additional identity/undef patterns to be handled.
Differential Revision: https://reviews.llvm.org/D113300
As suggested on D113371, this adds a wrapper to SelectionDAG::ComputeNumSignBits, similar to the llvm::ComputeMinSignedBits wrapper.
I've included some usage, its not exhaustive, just the more obvious cases where the intention is obvious.
Differential Revision: https://reviews.llvm.org/D113396
We have several places where we need to extract the raw bits data from a BUILD_VECTOR node, so consolidate this to a single helper function that handles Undefs and Integer/FP constants, including implicit truncation.
This should make it easier to extend D113202 to handle more constant folding of bitcasted constant data.
Differential Revision: https://reviews.llvm.org/D113351
Currently FoldConstantArithmetic only handles binops, so replacing other uses of FoldConstantVectorArithmetic (in particular for SETCC nodes), still require more work.
This is the 'or' sibling for the fold added with:
D113212
https://alive2.llvm.org/ce/z/tgnp7K
Note that neither of these transforms is poison-safe,
but it does not seem to matter at this level. We have
had the scalar version of D113212 for a long time, so
this is just making optimizer behavior consistent.
We do not have the scalar version of *this* fold,
however, so that is another follow-up.
Another minor step towards merging FoldConstantVectorArithmetic into FoldConstantArithmetic.
We don't use SDNodeFlags in any constant folding inside DAG, so passing the Flags argument is a waste of time - an alternative would be to wire up FoldConstantArithmetic to take SDNodeFlags just-in-case we someday start using it, but we don't have any way to test it and I'd prefer to avoid dead code.
Differential Revision: https://reviews.llvm.org/D113276
(X s< 0) ? Y : 0 --> (X s>> BW-1) & Y
We canonicalize to the icmp+select form in IR, and we already have this fold
for scalar select in SDAG, so I think it's an oversight that we don't have
the fold for vectors. It seems neutral for AArch64 and saves some instructions
on x86.
Whether we should also have the sibling folds for the inverse condition or
all-ones true value may depend on target-specific factors such as whether
there's an "and-not" instruction.
Differential Revision: https://reviews.llvm.org/D113212
NumOps represents the number of elements for vector constant folding, rename this NumElts so in future we can the consistently use NumOps to represent the number of operands of the opcode.
Minor cleanup before trying to begin generalizing FoldConstantArithmetic to support opcodes other than binops.
To constant fold bitwise logic ops where we've legalized constant build vectors to a different type (e.g. v2i64 -> v4i32), this patch adds a basic ability to peek through the bitcasts and perform the constant fold on the inner operands.
The MVE predicate v2i64 regressions will be addressed by future support for basic v2i64 type support.
One of the yak shaving fixes for D113192....
Differential Revision: https://reviews.llvm.org/D113202
Fold (srl (mul (zext i32:$a to i64), i64:c), 32) -> (mulhu $a, $b),
if c can truncate to i32 without loss.
Reviewed By: frasercrmck, craig.topper, RKSimon
Differential Revision: https://reviews.llvm.org/D108129
If the type of a funnel shift needs to be expanded, expand it to two funnel shifts instead of regular shifts. For constant shifts, this doesn't make much difference, but for variable shifts it allows a more optimal lowering.
Also use the optimized funnel shift lowering for rotates.
Alive2: https://alive2.llvm.org/ce/z/TvHDB- / https://alive2.llvm.org/ce/z/yzPept
(Branched from D108058 as getting this completed should help unlock some other WIP patches).
Original Patch: @efriedma (Eli Friedman)
Differential Revision: https://reviews.llvm.org/D112443
Rely on the hasOperation() instead - as commented on D77804, the mid-term intention is to recognise rotate/funnel-by-constant pre-legalization to help avoid SimplifyDemandedBits regressions.
Guillaume Chatelet