It's dangerous to knowingly create an illegal vector type
no matter what stage of combining we're in.
This prevents the missed folding/scalarization seen in:
https://bugs.llvm.org/show_bug.cgi?id=40146
llvm-svn: 350034
trunc (add X, C ) --> add (trunc X), C'
If we're throwing away the top bits of an 'add' instruction, do it in the narrow destination type.
This makes the truncate-able opcode list identical to the sibling transform done in IR (in instcombine).
This change used to show regressions for x86, but those are gone after D55494.
This gets us closer to deleting the x86 custom function (combineTruncatedArithmetic)
that does almost the same thing.
Differential Revision: https://reviews.llvm.org/D55866
llvm-svn: 350006
This saves materializing the immediate. The additional forms are less
common (they don't usually show up for bitfield insert/extract), but
they're still relevant.
I had to add a new target hook to prevent DAGCombine from reversing the
transform. That isn't the only possible way to solve the conflict, but
it seems straightforward enough.
Differential Revision: https://reviews.llvm.org/D55630
llvm-svn: 349857
Now that SimplifyDemandedBits/SimplifyDemandedVectorElts is simplifying vector elements, we're seeing more constant BUILD_VECTOR containing undefs.
This patch provides opt-in support for UNDEF elements in matchBinaryPredicate, passing NULL instead of the result ConstantSDNode* argument.
I've updated the (or (and X, c1), c2) -> (and (or X, c2), c1|c2) fold to demonstrate its use, which I believe is safe for undef cases.
Differential Revision: https://reviews.llvm.org/D55822
llvm-svn: 349629
As described on PR40091, we have several places where zext (and zext_vector_inreg) fold an undef input into an undef output. For zero extensions this is incorrect as the output should guarantee to least have the new upper bits set to zero.
SimplifyDemandedVectorElts is the worst offender (and its the most likely to cause new undefs to appear) but DAGCombiner's tryToFoldExtendOfConstant has a similar issue.
Thanks to @dmgreen for catching this.
Differential Revision: https://reviews.llvm.org/D55883
llvm-svn: 349625
The transform performs a bitwise logic op in a wider type followed by
truncate when both inputs are truncated from the same source type:
logic_op (truncate x), (truncate y) --> truncate (logic_op x, y)
There are a bunch of other checks that should prevent doing this when
it might be harmful.
We already do this transform for scalars in this spot. The vector
limitation was shared with a check for the case when the operands are
extended. I'm not sure if that limit is needed either, but that would
be a separate patch.
Differential Revision: https://reviews.llvm.org/D55448
llvm-svn: 349303
Also exposes an issue in DAGCombiner::visitFunnelShift where we were assuming the shift amount had the result type (after legalization it'll have the targets shift amount type).
llvm-svn: 349298
Summary:
If the setcc already has the target desired type we can reach the getSetCC/getSExtOrTrunc after the MatchingVecType check with the exact same types as the nodes we started with. This causes those causes VsetCC to be CSEd to N0 and the getSExtOrTrunc will CSE to N. When we return N, the caller will think that meant we called CombineTo and did our own worklist management. But that's not what happened. This prevents target hooks from being called for the node.
To fix this, I've now returned SDValue if the setcc is already the desired type. But to avoid some regressions in X86 I've had to disable one of the target combines that wasn't being reached before in the case of a (sext (setcc)). If we get vector widening legalization enabled that entire function will be deleted anyway so hopefully this is only for the short term.
Reviewers: RKSimon, spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D55459
llvm-svn: 349137
This isn't quite NFC, but I don't know how to expose
any outward diffs from these changes. Mostly, this
was confusing because it used 'VT' to refer to the
operand type rather the usual type of the input node.
There's also a large block at the end that is dedicated
solely to matching loads, but that wasn't obvious. This
could probably be split up into separate functions to
make it easier to see.
It's still not clear to me when we make certain transforms
because the legality and constant conditions are
intertwined in a way that might be improved.
llvm-svn: 349095
This is a retry of rL349051 (reverted at rL349056). I changed the check for dead-ness from
number of uses to an opcode test for DELETED_NODE based on existing similar code.
Differential Revision: https://reviews.llvm.org/D55655
llvm-svn: 349058
As discussed on D55511, this caused an issue if the inner node deletes a node that the outer node depends upon. As it doesn't affect any lit-tests and I've only been able to expose this with the D55511 change I'm committing this now.
llvm-svn: 348781
This triggers an assert when combining concat_vectors of a bitcast of
merge_values.
With asserts disabled, it fails to select:
fatal error: error in backend: Cannot select: 0x7ff19d000e90: i32 = any_extend 0x7ff19d000ae8
0x7ff19d000ae8: f64,ch = CopyFromReg 0x7ff19d000c20:1, Register:f64 %1
0x7ff19d000b50: f64 = Register %1
In function: d
Differential Revision: https://reviews.llvm.org/D55507
llvm-svn: 348759
This is effectively re-committing the changes from:
rL347917 (D54640)
rL348195 (D55126)
...which were effectively reverted here:
rL348604
...because the code had a bug that could induce infinite looping
or eventual out-of-memory compilation.
The bug was that this code did not guard against transforming
opaque constants. More details are in the post-commit mailing
list thread for r347917. A reduced test for that is included
in the x86 bool-math.ll file. (I wasn't able to reduce a PPC
backend test for this, but it was almost the same pattern.)
Original commit message for r347917:
The motivating case for this is shown in:
https://bugs.llvm.org/show_bug.cgi?id=32023
and the corresponding rot16.ll regression tests.
Because x86 scalar shift amounts are i8 values, we can end up with trunc-binop-trunc
sequences that don't get folded in IR.
As the TODO comments suggest, there will be regressions if we extend this (for x86,
we mostly seem to be missing LEA opportunities, but there are likely vector folds
missing too). I think those should be considered existing bugs because this is the
same transform that we do as an IR canonicalization in instcombine. We just need
more tests to make those visible independent of this patch.
llvm-svn: 348706
As discussed in the post-commit thread of r347917, this
transform is fighting with an existing transform causing
an infinite loop or out-of-memory, so this is effectively
reverting r347917 and its follow-up r348195 while we
investigate the bug.
llvm-svn: 348604
If this is not a valid way to assign an SDLoc, then we get this
wrong all over SDAG.
I don't know enough about the SDAG to explain this. IIUC, theoretically,
debug info is not supposed to affect codegen. But here it has clearly
affected 3 different targets, and the x86 change is an actual improvement.
llvm-svn: 348552
We shouldn't care about the debug location for a node that
we're creating, but attaching the root of the pattern should
be the best effort. (If this is not true, then we are doing
it wrong all over the SDAG).
This is no-functional-change-intended, and there are no
regression test diffs...and that's what I expected. But
there's a similar line above this diff, where those
assumptions apparently do not hold.
llvm-svn: 348550
This was probably organized as it was because bswap is a unary op.
But that's where the similarity to the other opcodes ends. We should
not limit this transform to scalars, and we should not try it if
either input has other uses. This is another step towards trying to
clean this whole function up to prevent it from causing infinite loops
and memory explosions.
Earlier commits in this series:
rL348501
rL348508
rL348518
llvm-svn: 348534
Unlike some of the folds in hoistLogicOpWithSameOpcodeHands()
above this shuffle transform, this has the expected hasOneUse()
checks in place.
llvm-svn: 348523
This patch introduces a new DAGCombiner rule to simplify concat_vectors nodes:
concat_vectors( bitcast (scalar_to_vector %A), UNDEF)
--> bitcast (scalar_to_vector %A)
This patch only partially addresses PR39257. In particular, it is enough to fix
one of the two problematic cases mentioned in PR39257. However, it is not enough
to fix the original test case posted by Craig; that particular case would
probably require a more complicated approach (and knowledge about used bits).
Before this patch, we used to generate the following code for function PR39257
(-mtriple=x86_64 , -mattr=+avx):
vmovsd (%rdi), %xmm0 # xmm0 = mem[0],zero
vxorps %xmm1, %xmm1, %xmm1
vblendps $3, %xmm0, %xmm1, %xmm0 # xmm0 = xmm0[0,1],xmm1[2,3]
vmovaps %ymm0, (%rsi)
vzeroupper
retq
Now we generate this:
vmovsd (%rdi), %xmm0 # xmm0 = mem[0],zero
vmovaps %ymm0, (%rsi)
vzeroupper
retq
As a side note: that VZEROUPPER is completely redundant...
I guess the vzeroupper insertion pass doesn't realize that the definition of
%xmm0 from vmovsd is already zeroing the upper half of %ymm0. Note that on
%-mcpu=btver2, we don't get that vzeroupper because pass vzeroupper insertion
%pass is disabled.
Differential Revision: https://reviews.llvm.org/D55274
llvm-svn: 348522
The PPC test with 2 extra uses seems clearly better by avoiding this transform.
With 1 extra use, we also prevent an extra register move (although that might
be an RA problem). The general rule should be to only make a change here if
it is always profitable. The x86 diffs are all neutral.
llvm-svn: 348518
The AVX512 diffs are neutral, but the bswap test shows a clear overreach in
hoistLogicOpWithSameOpcodeHands(). If we don't check for other uses, we can
increase the instruction count.
This could also fight with transforms trying to go in the opposite direction
and possibly blow up/infinite loop. This might be enough to solve the bug
noted here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20181203/608593.html
I did not add the hasOneUse() checks to all opcodes because I see a perf
regression for at least one opcode. We may decide that's irrelevant in the
face of potential compiler crashing, but I'll see if I can salvage that first.
llvm-svn: 348508
Because we're potentially peeking through a bitcast in this transform,
we need to use overall bitwidths rather than number of elements to
determine when it's safe to proceed.
Should fix:
https://bugs.llvm.org/show_bug.cgi?id=39893
llvm-svn: 348383
This is an initial patch to add a minimum level of support for funnel shifts to the SelectionDAG and to begin wiring it up to the X86 SHLD/SHRD instructions.
Some partial legalization code has been added to handle the case for 'SlowSHLD' where we want to expand instead and I've added a few DAG combines so we don't get regressions from the existing DAG builder expansion code.
Differential Revision: https://reviews.llvm.org/D54698
llvm-svn: 348353
Add support for ISD::*_EXTEND and ISD::*_EXTEND_VECTOR_INREG opcodes.
The extra broadcast in trunc-subvector.ll will be fixed in an upcoming patch.
llvm-svn: 348246
This is the smallest vector enhancement I could find to D54640.
Here, we're allowing narrowing to only legal vector ops because we'll see
regressions without that. All of the test diffs are wins from what I can tell.
With AVX/AVX512, we can shrink ymm/zmm ops to xmm.
x86 vector multiplies are the problem case that we're avoiding due to the
patchwork ISA, and it's not clear to me if we can dance around those
regressions using TLI hooks or if we need preliminary patches to plug those
holes.
Differential Revision: https://reviews.llvm.org/D55126
llvm-svn: 348195
Summary:
Under -x86-experimental-vector-widening-legalization, fp_to_uint/fp_to_sint with a smaller than 128 bit vector type results are custom type legalized by promoting the result to a 128 bit vector by promoting the elements, inserting an assertzext/assertsext, then truncating back to original type. The truncate will be further legalizdd to a pack shuffle. In the case of a v8i8 result type, we'll end up with a v8i16 fp_to_sint. This will need to be further legalized during vector op legalization by promoting to v8i32 and then truncating again. Under avx2 this produces good code with two pack instructions, but Under avx512 this will result in a truncate instruction and a packuswb instruction. But we should be able to get away with a single truncate instruction.
The other option is to promote all the way to vXi32 result type during the first type legalization. But in some experimentation that seemed to require more work to produce good code for other configurations.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54836
llvm-svn: 348158
This change prevents the crash noted in the post-commit comments
for rL347478 :
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20181119/605166.html
We can't guarantee that an oversized shift amount is folded away,
so we have to check for it.
Note that I committed an incomplete fix for that crash with:
rL347502
But as discussed here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20181126/605679.html
...we have to try harder.
So I'm not sure how to expose the bug now (and apparently no fuzzers have found
a way yet either).
On the plus side, we have discovered that we're missing real optimizations by
not simplifying nodes sooner, so the earlier fix still has value, and there's
likely more value in extending that so we can simplify more opcodes and simplify
when doing RAUW and/or putting nodes on the combiner worklist.
Differential Revision: https://reviews.llvm.org/D54954
llvm-svn: 348089
The motivating case for this is shown in:
https://bugs.llvm.org/show_bug.cgi?id=32023
and the corresponding rot16.ll regression tests.
Because x86 scalar shift amounts are i8 values, we can end up with trunc-binop-trunc
sequences that don't get folded in IR.
As the TODO comments suggest, there will be regressions if we extend this (for x86,
we mostly seem to be missing LEA opportunities, but there are likely vector folds
missing too). I think those should be considered existing bugs because this is the
same transform that we do as an IR canonicalization in instcombine. We just need
more tests to make those visible independent of this patch.
Differential Revision: https://reviews.llvm.org/D54640
llvm-svn: 347917
This should likely be adjusted to limit this transform
further, but these diffs should be clear wins.
If we have blendv/conditional move, then we should assume
those are cheap ops. The loads become independent of the
compare, so those can be speculated before we need to use
the values in the blend/mov.
llvm-svn: 347526
rL347502 moved the null sibling, so we should group all of these
together. I'm not sure why these aren't methods of the SDValue
class itself, but that's another patch if that's possible.
llvm-svn: 347523
...and use them to avoid creating obviously undef values as
discussed in the post-commit thread for r347478.
The diffs in vector div/rem show that we were missing real
optimizations by creating bogus shift nodes.
llvm-svn: 347502
We fail to canonicalize IR this way (prefer 'not' ops to arbitrary 'xor'),
but that would not matter without this patch because DAGCombiner was
reversing that transform. I think we need this transform in the backend
regardless of what happens in IR to catch cases where the shift-xor
is formed late from GEP or other ops.
https://rise4fun.com/Alive/NC1
Name: shl
Pre: (-1 << C2) == C1
%shl = shl i8 %x, C2
%r = xor i8 %shl, C1
=>
%not = xor i8 %x, -1
%r = shl i8 %not, C2
Name: shr
Pre: (-1 u>> C2) == C1
%sh = lshr i8 %x, C2
%r = xor i8 %sh, C1
=>
%not = xor i8 %x, -1
%r = lshr i8 %not, C2
https://bugs.llvm.org/show_bug.cgi?id=39657
llvm-svn: 347478
This transform needs to be limited.
We are converting to a constant pool load very early, and we
are turning loads that are independent of the select condition
(and therefore speculatable) into a dependent non-speculatable
load.
We may also be transferring a condition code from an FP register
to integer to create that dependent load.
llvm-svn: 347424
This is another step in vector narrowing - a follow-up to D53784
(and hoping to eventually squash potential regressions seen in
D51553).
The x86 test diffs are wins, but the AArch64 diff is probably not.
That problem already exists independent of this patch (see PR39722), but it
went unnoticed in the previous patch because there were no regression tests
that showed the possibility.
The x86 diff in i64-mem-copy.ll is close. Given the frequency throttling
concerns with using wider vector ops, an extra extract to reduce vector
width is the right trade-off at this level of codegen.
Differential Revision: https://reviews.llvm.org/D54392
llvm-svn: 347356
This uncovered an off-by-one typo in SimplifyDemandedVectorElts's INSERT_SUBVECTOR handling as its bounds check was bailing on safe indices.
llvm-svn: 347313
Consistently use (!LegalOperations || isOperationLegalOrCustom) for all node pairs.
Differential Revision: https://reviews.llvm.org/D53478
llvm-svn: 347255
Sadly, this duplicates (twice) the logic from InstSimplify. There
might be some way to at least share the DAG versions of the code,
but copying the folds seems to be the standard method to ensure
that we don't miss these folds.
Unlike in IR, we don't run DAGCombiner to fixpoint, so there's no
way to ensure that we do these kinds of simplifications unless the
code is repeated at node creation time and during combines.
There were other tests that would become worthless with this
improvement that I changed as pre-commits:
rL347161
rL347164
rL347165
rL347166
rL347167
I'm not sure how to salvage the remaining tests (diffs in this patch).
So the x86 tests verify that the new code is working as intended.
The AMDGPU test is actually similar to my motivating case: we have
some undef value that has survived to machine IR in an x86 test, and
then it gets folded in some weird way, or we crash if we don't transfer
the undef flag. But we would have been better off never getting to that
point by doing these simplifications.
This will lead back to PR32023 someday...
https://bugs.llvm.org/show_bug.cgi?id=32023
llvm-svn: 347170
It should be ok to create a new build_vector after legal operations so long as it doesn't cause an infinite loop in DAG combiner.
Unfortunately, X86's custom constant folding in combineVSZext is hiding any test changes from this. But I'm trying to get to a point where that X86 specific code isn't necessary at all.
Differential Revision: https://reviews.llvm.org/D54285
llvm-svn: 346728
Summary:
Handle extra output from index loads in cases where we wish to
forward a load value directly from a preceeding store.
Fixes PR39571.
Reviewers: peter.smith, rengolin
Subscribers: javed.absar, hiraditya, arphaman, llvm-commits
Differential Revision: https://reviews.llvm.org/D54265
llvm-svn: 346654
This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more
opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs.
Apart from 2-3 strange cases, these are all wins.
I've structured this to be no-functional-change-intended for any target except for x86
because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those
targets have existing regression tests (4, 4, 10 files respectively) that would be
affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show
any regression test diffs. The trade-off is deciding if an extra vector load is better
than a single wide load + extract_subvector.
For x86, this is almost always better (on paper at least) because we often can fold
loads into subsequent ops and not increase the official instruction count. There's also
some unknown -- but potentially large -- benefit from using narrower vector ops if wide
ops are implemented with multiple uops and/or frequency throttling is avoided.
Differential Revision: https://reviews.llvm.org/D54073
llvm-svn: 346595
It's possible for vector op legalization to generate a shuffle. If that happens we should give a chance for DAG combine to combine that with a build_vector input.
I also fixed a bug in combineShuffleOfScalars that was considering the number of uses on a undef input to a shuffle. We don't care how many times undef is used.
Differential Revision: https://reviews.llvm.org/D54283
llvm-svn: 346530
The DAGCombiner tries to SimplifySelectCC as follows:
select_cc(x, y, 16, 0, cc) -> shl(zext(set_cc(x, y, cc)), 4)
It can't cope with the situation of reordered operands:
select_cc(x, y, 0, 16, cc)
In that case we just need to swap the operands and invert the Condition Code:
select_cc(x, y, 16, 0, ~cc)
Differential Revision: https://reviews.llvm.org/D53236
llvm-svn: 346484
FindBetterNeighborChains simulateanously improves the chain
dependencies of a chain of related stores avoiding the generation of
extra token factors. For chains longer than the GatherAllAliasDepths,
stores further down in the chain will necessarily fail, a potentially
significant waste and preventing otherwise trivial parallelization.
This patch directly parallelize the chains of stores before improving
each store. This generally improves DAG-level parallelism.
Reviewers: courbet, spatel, RKSimon, bogner, efriedma, craig.topper, rnk
Subscribers: sdardis, javed.absar, hiraditya, jrtc27, atanasyan, llvm-commits
Differential Revision: https://reviews.llvm.org/D53552
llvm-svn: 346432
The original code avoided creating a zero vector after type legalization, but if we're after type legalization the type we have is legal. The real hazard we need to avoid is creating a build vector after op legalization. tryFoldToZero takes care of checking for this.
llvm-svn: 346119
These methods were just wrappers around getNode with additional asserts (identical and repeated 3 times). But getNode already has a switch that can be used to hold these asserts that allows them to be shared for all 3 opcodes. This also enables checking on the places that create these nodes without using the wrappers.
The rest of the patch is just changing all callers to use getNode directly.
llvm-svn: 346087
We already have custom lowering for the AVX case in LegalizeVectorOps. So its better to keep the regular extend op around as long as possible.
I had to qualify one place in DAG combine that created illegal vector extending load operations. This change by itself had no effect on any tests which is why its included here.
I've made a few cleanups to the custom lowering. The sign extend code no longer creates an identity shuffle with undef elements. The zero extend code now emits a zero_extend_vector_inreg instead of an unpckl with a zero vector.
For the high half of the custom lowering of zero_extend/any_extend, we're now using an unpckh with a zero vector or undef. Previously we used used a pshufd to move the upper 64-bits to the lower 64-bits and then used a zero_extend_vector_inreg. I think the zero vector should require less execution resources and be smaller code size.
Differential Revision: https://reviews.llvm.org/D54024
llvm-svn: 346043
reduceBuildVecConvertToConvertBuildVec vectorizes int2float in the DAGCombiner, which means that even if the LV/SLP has decided to keep scalar code using the cost models, this will override this.
While there are cases where vectorization is necessary in the DAG (mainly due to legalization artefacts), I don't think this is the case here, we should assume that the vectorizers know what they are doing.
Differential Revision: https://reviews.llvm.org/D53712
llvm-svn: 345964
I'm having trouble creating a test case for the ISD::TRUNCATE part of this that shows any codegen differences. But I was able to test the setcc path which is what the test changes here cover.
llvm-svn: 345908
The test causes a crash because we were trying to extract v4f32 to v3f32, and the
narrowing factor was then 4/3 = 1 producing a bogus narrow type.
This should fix:
https://bugs.llvm.org/show_bug.cgi?id=39511
llvm-svn: 345842
Summary:
Normalize the offset for endianess before checking
if the store cover the load in ForwardStoreValueToDirectLoad.
Without this we missed out on some optimizations for big
endian targets. If for example having a 4 bytes store followed
by a 1 byte load, loading the least significant byte from the
store, the STCoversLD check would fail (see @test4 in
test/CodeGen/AArch64/load-store-forwarding.ll).
This patch also fixes a problem seen in an out-of-tree target.
The target has i40 as a legal type, it is big endian,
and the StoreSize for i40 is 48 bits. So when normalizing
the offset for endianess we need to take the StoreSize into
account (assuming that padding added when storing into
a larger StoreSize always is added at the most significant
end).
Reviewers: niravd
Reviewed By: niravd
Subscribers: javed.absar, kristof.beyls, llvm-commits, uabelho
Differential Revision: https://reviews.llvm.org/D53776
llvm-svn: 345636
Narrowing vector binops came up in the demanded bits discussion in D52912.
I don't think we're going to be able to do this transform in IR as a canonicalization
because of the risk of creating unsupported widths for vector ops, but we already have
a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is
currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression
test diffs).
This is artificially limited to not look through bitcasts because there are so many
test diffs already, but that's marked with a TODO and is a small follow-up.
Differential Revision: https://reviews.llvm.org/D53784
llvm-svn: 345602
Summary:
Changes all uses of minnan/maxnan to minimum/maximum
globally. These names emphasize that the semantic difference between
these operations is more than just NaN-propagation.
Reviewers: arsenm, aheejin, dschuff, javed.absar
Subscribers: jholewinski, sdardis, wdng, sbc100, jgravelle-google, jrtc27, atanasyan, llvm-commits
Differential Revision: https://reviews.llvm.org/D53112
llvm-svn: 345218
Until now, we've only checked whether merging stores would cause a cycle via
the value argument, but the address and indexed offset arguments are also
capable of creating cycles in some situations.
The addresses are all base+offset with notionally the same base, but the base
SDNode may still be different (e.g. via an indexed load in one case, and an
ISD::ADD elsewhere). This allows cycles to creep in if one of these sources
depends on another.
The indexed offset is usually undef (representing a non-indexed store), but on
some architectures (e.g. 32-bit ARM-mode ARM) it can be an arbitrary value,
again allowing dependency cycles to creep in.
llvm-svn: 345200
When implementing memset's today we often see this pattern:
$x0 = MOV 0xXYXYXYXYXYXYXYXY
store $x0, ...
$w1 = MOV 0xXYXYXYXY
store $w1, ...
We first create a 64bit constant in a 64bit register with all bytes the
same and then create a 32bit constant with all bytes the same in a 32bit
register. In many targets we could just access the lower byte of the
64bit register instead.
- Ideally this would be handled by the ConstantHoist pass but it runs
too early when memset isn't expanded yet.
- The memset expansion code already had this optimization implemented,
however SelectionDAG constantfolding would constantfold the
"trunc(bigconstnat)" pattern to "smallconstant".
- This patch makes the memset expansion mark the constant as Opaque and
stop DAGCombiner from constant folding in this situation. (Similar to
how ConstantHoisting marks things as Opaque to avoid folding
ADD/SUB/etc.)
Differential Revision: https://reviews.llvm.org/D53181
llvm-svn: 345102
I've included a fix to DAGCombiner::ForwardStoreValueToDirectLoad that I believe will prevent the previous miscompile.
Original commit message:
Theoretically this was done to simplify the amount of isel patterns that were needed. But it also meant a substantial number of our isel patterns have to match an explicit bitcast. By making the vXi32/vXi16/vXi8 types legal for loads, DAG combiner should be able to change the load type to rem
I had to add some additional plain load instruction patterns and a few other special cases, but overall the isel table has reduced in size by ~12000 bytes. So it looks like this promotion was hurting us more than helping.
I still have one crash in vector-trunc.ll that I'm hoping @RKSimon can help with. It seems to relate to using getTargetConstantFromNode on a load that was shrunk due to an extract_subvector combine after the constant pool entry was created. So we end up decoding more mask elements than the lo
I'm hoping this patch will simplify the number of patterns needed to remove the and/or/xor promotion.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, RKSimon
Differential Revision: https://reviews.llvm.org/D53306
llvm-svn: 344965
Introduce new versions that follow the IEEE semantics
to help with legalization that may need quieted inputs.
There are some regressions from inserting unnecessary
canonicalizes when these are matched from fast math
fcmp + select which should be fixed in a future commit.
llvm-svn: 344914
This is a late backend subset of the IR transform added with:
D52439
We can confirm that the conversion to a 'trunc' is correct by running:
$ opt -instcombine -data-layout="e"
(assuming the IR transforms are correct; change "e" to "E" for big-endian)
As discussed in PR39016:
https://bugs.llvm.org/show_bug.cgi?id=39016
...the pattern may emerge during legalization, so that's we are waiting for an
insertelement to become a scalar_to_vector in the pattern matching here.
The DAG allows for fun variations that are not possible in IR. Result types for
extracts and scalar_to_vector don't necessarily match input types, so that means
we have to be a bit more careful in the transform (see code comments).
The tests show that we don't handle cases that require a shift (as we did in the
IR version). I've left that as a potential follow-up because I'm not sure if
that's a real concern at this late stage.
Differential Revision: https://reviews.llvm.org/D53201
llvm-svn: 344872
I want to add another pattern here that includes scalar_to_vector,
so this makes that patch smaller. I was hoping to remove the
hasOneUse() check because it shouldn't be necessary for common
codegen, but an AMDGPU test has a comment suggesting that the
extra check makes things better on one of those targets.
llvm-svn: 344320
Summary:
Extend analysis forwarding loads from preceeding stores to work with
extended loads and truncated stores to the same address so long as the
load is fully subsumed by the store.
Hexagon's swp-epilog-phis.ll and swp-memrefs-epilog1.ll test are
deleted as they've no longer seem to be relevant.
Reviewers: RKSimon, rnk, kparzysz, javed.absar
Subscribers: sdardis, nemanjai, hiraditya, atanasyan, llvm-commits
Differential Revision: https://reviews.llvm.org/D49200
llvm-svn: 344142
We already do the following combines:
(bitcast int (and (bitcast fp X to int), 0x7fff...) to fp) -> fabs X
(bitcast int (xor (bitcast fp X to int), 0x8000...) to fp) -> fneg X
When the target has "bit preserving fp logic". This patch just extends it
to also combine:
(bitcast int (or (bitcast fp X to int), 0x8000...) to fp) -> fneg (fabs X)
As some targets have fnabs and even those that don't can efficiently lower
both the fabs and the fneg.
Differential revision: https://reviews.llvm.org/D44548
llvm-svn: 344093
This change is proposed as a part of D44548, but we
need this independently to avoid regressions from improved
undef propagation in SimplifyDemandedVectorElts().
llvm-svn: 343940
And use that to transform fsub with zero constant operands.
The integer part isn't used yet, but it is proposed for use in
D44548, so adding both enhancements here makes that
patch simpler.
llvm-svn: 343865
This fixes a case of bad index calculation when merging mismatching
vector types. This changes the existing code to just use the existing
extract_{subvector|element} and a bitcast (instead of bitcast first and
then newly created extract_xxx) so we don't need to adjust any indices
in the first place.
rdar://44584718
Differential Revision: https://reviews.llvm.org/D52681
llvm-svn: 343493
The SINT_TO_FP<->UINT_TO_FP combines for non-negative integers should only occur for legal ops once LegalOperations = true
No test case to hand, noticed when investigating PR38226 + PR38970
llvm-svn: 343405
DAGCombine will try to fold two loads that feed a SELECT or SELECT_CC
after the select, resulting in a select of an address and a single
load after.
If either of the loads depend on the other, this is not legal as it
could introduce cycles. However, it only checked this if the opcode
was a SELECT, and not for a SELECT_CC.
Unfortunately, the only reproducer I have for this is for our
downstream target. I've tried getting it to trigger on an upstream one
but haven't been successful.
Patch thanks to Bevin Hansson.
llvm-svn: 342980
This is a preliminary step towards solving PR14613:
https://bugs.llvm.org/show_bug.cgi?id=14613
If we have an 'add' instruction that sets flags, we can use that to eliminate an
explicit compare instruction or some other instruction (cmn) that sets flags for
use in the later select.
As shown in the unchanged tests that use 'icmp ugt %x, %a', we're effectively
reversing an IR icmp canonicalization that replaces a variable operand with a
constant:
https://rise4fun.com/Alive/V1Q
But we're not using 'uaddo' in those cases via DAG transforms. This happens in
CGP after D8889 without checking target lowering to see if the op is supported.
So AArch already shows 'uaddo' codegen for the i8/i16/i32/i64 test variants with
"using_cmp_sum" in the title. That's the pattern that CGP matches as an unsigned
saturated add and converts to uaddo without checking target capabilities.
This patch is gated by isOperationLegalOrCustom(ISD::UADDO, VT), so we see only
see AArch diffs for i32/i64 in the tests with "using_cmp_notval" in the title
(unlike x86 which sees improvements for all sizes because all sizes are 'custom').
But the AArch code (like x86) looks better when translated to 'uaddo' in all cases.
So someone that is involved with AArch may want to set i8/i16 to 'custom' for UADDO,
so this patch will fire on those tests.
Another possibility given the existing behavior: we could remove the legal-or-custom
check altogether because we're assuming that a UADDO sequence is canonical/optimal
before we ever reach here. But that seems like a bug to me. If the target doesn't
have an add-with-flags op, then it's not likely that we'll get optimal DAG combining
using a UADDO node. This is similar justification for why we don't canonicalize IR to
the overflow math intrinsic sibling (llvm.uadd.with.overflow) for UADDO in the first
place.
Differential Revision: https://reviews.llvm.org/D51929
llvm-svn: 342886
This code handled SCALAR_TO_VECTOR being returned by the recursion, but the code that used to return SCALAR_TO_VECTOR was removed in 2015.
llvm-svn: 342856
This comment was misleading about why we were restricting to before legalize types. The reason given would only apply to before legalize ops. But there is a before legalize types reason that should also be listed.
llvm-svn: 342851
This is an alternative to https://reviews.llvm.org/D37896. We can't decompose
multiplies generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some existing code that overlaps with
this transform.
This extends D52195 and may resolve PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
(still an open question about transforming legal vector multiplies, but we
could open another bug report for those)
llvm-svn: 342844
x86 had 2 versions of peekThroughBitcast. DAGCombiner had 1. Plus, it had a 1-off implementation for the one-use variant.
Move the x86 versions of the code to SelectionDAG, so we don't have different copies of the code.
No functional change intended.
I'm putting this next to isBitwiseNot() because I am planning to use it in there. Another option is next to the
helpers in the ISD namespace (eg, ISD::isConstantSplatVector()). But if there's no good reason for those to be
there, I'd prefer to pull other helpers over to SelectionDAG in follow-up steps.
Differential Revision: https://reviews.llvm.org/D52285
llvm-svn: 342669
The test diff in not-and-simplify.ll is from a use in SimplifyDemandedBits,
and the test diff in add.ll is from a DAGCombiner transform.
llvm-svn: 342594
This is an alternative to D37896. I don't see a way to decompose multiplies
generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some duplicate code that overlaps with
this transform.
As a first step, we're only getting the most clear wins on the vector examples
requested in PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
As noted in the code comment, it's likely that the x86 constraints are tighter
than necessary, but it may not always be a win to replace a pmullw/pmulld.
Differential Revision: https://reviews.llvm.org/D52195
llvm-svn: 342554
This is a follow-up suggested in D51630 and originally proposed as an IR transform in D49040.
Copying the motivational statement by @evandro from that patch:
"This transformation helps some benchmarks in SPEC CPU2000 and CPU2006, such as 188.ammp,
447.dealII, 453.povray, and especially 300.twolf, as well as some proprietary benchmarks.
Otherwise, no regressions on x86-64 or A64."
I'm proposing to add only the minimum support for a DAG node here. Since we don't have an
LLVM IR intrinsic for cbrt, and there are no other DAG ways to create a FCBRT node yet, I
don't think we need to worry about DAG builder, legalization, a strict variant, etc. We
should be able to expand as needed when adding more functionality/transforms. For reference,
these are transform suggestions currently listed in SimplifyLibCalls.cpp:
// * cbrt(expN(X)) -> expN(x/3)
// * cbrt(sqrt(x)) -> pow(x,1/6)
// * cbrt(cbrt(x)) -> pow(x,1/9)
Also, given that we bail out on long double for now, there should not be any logical
differences between platforms (unless there's some platform out there that has pow()
but not cbrt()).
Differential Revision: https://reviews.llvm.org/D51753
llvm-svn: 342348
Add support for bitcasts from float type to an integer type of the same element bitwidth.
There maybe cases where we need to support different widths (e.g. as SSE __m128i is treated as v2i64) - but I haven't seen cases of this in the wild yet.
llvm-svn: 341652
This was proposed as an IR transform in D49306, but it was not clearly justifiable as a canonicalization.
Here, we only do the transform when the target tells us that sqrt can be lowered with inline code.
This is the basic case. Some potential enhancements are in the TODO comments:
1. Generalize the transform for other exponents (allow more than 2 sqrt calcs if that's really cheaper).
2. If we have less fast-math-flags, generate code to avoid -0.0 and/or INF.
3. Allow the transform when optimizing/minimizing size (might require a target hook to get that right).
Note that by default, x86 converts single-precision sqrt calcs into sqrt reciprocal estimate with
refinement. That codegen is controlled by CPU attributes and can be manually overridden. We have plenty
of test coverage for that already, so I didn't bother to include extra testing for that here. AArch uses
its full-precision ops in all cases (not sure if that's the intended behavior or not, but that should
also be covered by existing tests).
Differential Revision: https://reviews.llvm.org/D51630
llvm-svn: 341481
Summary:
I'm not sure if this patch is correct or if it needs more qualifying somehow. Bitcast shouldn't change the size of the load so it should be ok? We already do something similar for stores. We'll change the type of a volatile store if the resulting store is Legal or Custom. I'm not sure we should be allowing Custom there...
I was playing around with converting X86 atomic loads/stores(except seq_cst) into regular volatile loads and stores during lowering. This would allow some special RMW isel patterns in X86InstrCompiler.td to be removed. But there's some floating point patterns in there that didn't work because we don't fold (f64 (bitconvert (i64 volatile load))) or (f32 (bitconvert (i32 volatile load))).
Reviewers: efriedma, atanasyan, arsenm
Reviewed By: efriedma
Subscribers: jvesely, arsenm, sdardis, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, arichardson, jrtc27, atanasyan, jfb, llvm-commits
Differential Revision: https://reviews.llvm.org/D50491
llvm-svn: 340797
I noticed this along with the patterns in D51125, but when the index is variable,
we don't convert insertelement into a build_vector.
For x86, that means these get expanded at legalization time into the loading/spilling
code that we see in the tests. I think it's always better to avoid going to memory on
these, and we get the optimal 'broadcast' if it's available.
I suspect other targets may want to look at enabling the hook. AArch64 and AMDGPU have
regression tests that would be affected (although I did not check what would happen in
those cases). In the most basic cases shown here, AArch64 would probably do much
better with a splat.
Differential Revision: https://reviews.llvm.org/D51186
llvm-svn: 340705
Previously we allowed the store to be Custom. But without knowing for sure that the Custom handling won't split the store, we shouldn't convert a volatile store. We also probably shouldn't be creating a store the requires custom handling after LegalizeOps. This could lead to an infinite loop if the custom handling was to insert a bitcast. Though I guess isStoreBitCastBeneficial could be used to block such a loop.
The test changes here are due to the volatile part of this. The stores in the test are all volatile and i32 stores are marked custom, So we are no longer converting them
This is related to D50491 where I was trying to allow some bitcasting of volatile loads
Differential Revision: https://reviews.llvm.org/D50578
llvm-svn: 340626
During combining, ReduceLoadWdith is used to combine AND nodes that
mask loads into narrow loads. This patch allows the mask to be a
shifted constant. This results in a narrow load which is then left
shifted to compensate for the new offset.
Differential Revision: https://reviews.llvm.org/D50432
llvm-svn: 340261
Summary:
I believe this restores the behavior we had before r339147.
Fixes PR38622.
Reviewers: RKSimon, chandlerc, spatel
Reviewed By: chandlerc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D50936
llvm-svn: 340120
Add support for cases where only some c1+c2 results exceed the max bitshift, clamping accordingly.
Differential Revision: https://reviews.llvm.org/D35722
llvm-svn: 340010
When nodes are reassociated the vector-reduction flag gets lost.
The test case is here is what would happen if you had a sum of absolute differences loop that started with a non-zero but contant sum and that loop was unrolled. The vectorizer will generate a constant vector for the initial value. And DAGCombiner reassociate tries to move it down the addition tree erasing the vector-reduction flag. Interestingly this moves constants the opposite direction of the reassociate IR pass.
I've chosen to just punt on the reassociate, but I suppose we could maybe preserve the flag if both nodes have it set.
Differential Revision: https://reviews.llvm.org/D50827
llvm-svn: 339946
Intentionally excluding nodes from the DAGCombine worklist is likely to
lead to weird optimizations and infinite loops, so it's generally a bad
idea.
To avoid the infinite loops, fix DAGCombine to use the
isDesirableToCommuteWithShift target hook before performing the
transforms in question, and implement the target hook in the ARM backend
disable the transforms in question.
Fixes https://bugs.llvm.org/show_bug.cgi?id=38530 . (I don't have a
reduced testcase for that bug. But we should have sufficient test
coverage for PerformSHLSimplify given that we're not playing weird
tricks with the worklist. I can try to bugpoint it if necessary,
though.)
Differential Revision: https://reviews.llvm.org/D50667
llvm-svn: 339734
Summary: This change provides a common optimization path for both Unsafe and FMF driven optimization for this fsub fold adding reassociation, as it the flag that most closely represents the translation
Reviewers: spatel, wristow, arsenm
Reviewed By: spatel
Subscribers: wdng
Differential Revision: https://reviews.llvm.org/D50195
llvm-svn: 339357
isNegatibleForFree() should not matter here (as the test diffs show)
because it's always a win to replace an fsub+fadd with fneg. The
problem in D50195 persists because either (1) we are doing these
folds in the wrong order or (2) we're missing another fold for fadd.
llvm-svn: 339299
I don't know if it's possible to expose this diff in a test,
but we should always try simplifications (no new nodes created)
before more complicated transforms for efficiency (similar to
what we do in IR).
llvm-svn: 339298
The isConstOrConstSplat result is only used in a ISD::matchUnaryPredicate call which can perform the equivalent iteration just as quickly.
llvm-svn: 339262
Scatter could have multiple identical indices. We need to maintain sequential order. We get this right in LegalizeVectorTypes, but not in this code.
Differential Revision: https://reviews.llvm.org/D50374
llvm-svn: 339157
This was missed in D50185.
NFC until we add actual non-uniform support to BuildSDIV (similar BuildUDIV support in D49248) - for now it just early outs.
llvm-svn: 339147
This patch refactors the existing TargetLowering::BuildUDIV base implementation to support non-uniform constant vector denominators.
It also includes a fold for MULHU by pow2 constants to SRL which can now more readily occur from BuildUDIV.
Differential Revision: https://reviews.llvm.org/D49248
llvm-svn: 339121
Craig Topper