Truncations lowered as shuffles of multiple (concatenated) vectors often leave us with lane-crossing shuffles that feed a PACKSS/PACKUS, if both shuffles are fed from the same 2 vector sources, then we can PACK the sources directly and shuffle the result instead.
This is currently limited to whole i128 lanes in a 256-bit vector, but we can extend this if the need arises (but I'm not seeing many examples in real world code).
This patch builds on 0d7286a652 by simplifying the code for detecting
splat values and adding new tests demonstrating the lowering of
splatted absolute value shift amounts, which are common in code
generated by Halide. The lowering is very bad right now, but
subsequent patches will improve it considerably. The tests will be
useful for evaluating the improvements in those patches.
Reviewed By: aheejin
Differential Revision: https://reviews.llvm.org/D83493
As requested, these tests check for specific debug locs on the output of the
legalizer. The only one that I couldn't write was for moreElementsVector, which
AFAICT we don't trigger on AArch64.
This is currently bare-bones; we aren't taking advantage of any of the
FMA variant instructions. But it's enough to at least generate
code.
Differential Revision: https://reviews.llvm.org/D83444
The following: https://reviews.llvm.org/D82552
fixed an assert in the SelectionDag ISel legalizer for some CCs on armv7.
I noticed that this fix also fixes the assert when using fastcc, so I am
adding a fastcc regression test here.
Differential Revision: https://reviews.llvm.org/D82443
If we don't immediately lower the vector shift, the splat
constant vector we created may get turned into a constant pool
load before we get around to lowering the shift. This makes it
a lot more difficult to create a shift by constant. Sometimes we
fail to see through the constant pool at all and end up trying
to lower as if it was a variable shift. This requires custom
handling and may create an unsupported vselect on pre-sse-4.1
targets. Since we're after LegalizeVectorOps we are unable to
legalize the unsupported vselect as that code is in LegalizeVectorOps
rather than LegalizeDAG.
So calling LowerShift immediately ensures that we get see the
splat constant.
Fixes PR46527.
Differential Revision: https://reviews.llvm.org/D83455
The unfortunate split between finalizeLowering and the selector pass
means there's a point where the verifier fails. The DAG selector pass
skips the verifier, but this seems to not work when using the
GlobalISel fallback.
If we're extracting a subvector from a shuffle that is shuffling entire subvectors we can peek through and extract the subvector from the shuffle source instead.
This helps remove some cases where concat_vectors(extract_subvector(),extract_subvector()) legalizations has resulted in BLEND/VPERM2F128 shuffles of the subvectors.
Since the `RISCVExpandPseudo` pass has been split from
`RISCVExpandAtomicPseudo` pass, it would be nice to run the former as
early as possible (The latter has to be run as late as possible to
ensure correctness). Running earlier means we can reschedule these pairs
as we see fit.
Running earlier in the machine pass pipeline is good, but would mean
teaching many more passes about `hasLabelMustBeEmitted`. Splitting the
basic blocks also pessimises possible optimisations because some
optimisations are MBB-local, and others are disabled if the block has
its address taken (which is notionally what `hasLabelMustBeEmitted`
means).
This patch uses a new approach of setting the pre-instruction symbol on
the AUIPC instruction to a temporary symbol and referencing that. This
avoids splitting the basic block, but allows us to reference exactly the
instruction that we need to. Notionally, this approach seems more
correct because we do actually want to address a specific instruction.
This then allows the pass to be moved much earlier in the pass pipeline,
before both scheduling and register allocation. However, to do so we
must leave the MIR in SSA form (by not redefining registers), and so use
a virtual register for the intermediate value. By using this virtual
register, this pass now has to come before register allocation.
Reviewed By: luismarques, asb
Differential Revision: https://reviews.llvm.org/D82988
Fixed length vector code generation for SVE does not yet custom
lower BUILD_VECTOR and instead relies on expansion. At the same
time custom lowering for VECTOR_SHUFFLE is also not available so
this patch updates isShuffleMaskLegal to reject vector types that
require SVE.
Related to this it also prevents the merging of stores after
legalisation because this only works when BUILD_VECTOR is either
legal or can be elminated. When this is not the case the code
generator enters an infinite legalisation loop.
Differential Revision: https://reviews.llvm.org/D83408
Summary:
When legalizing a biscast operation from an fp16 operand to an i16 on a
target that requires both input and output types to be promoted to
32-bits, an assertion can fail when building the new node due to a
mismatch between the the operation's result size and the type specified to
the node.
This patches fix the issue by making sure the bit width of the types
match for the FP_TO_FP16 node, covering the difference with an extra
ANYEXTEND operation.
Reviewers: ostannard, efriedma, pirama, jmolloy, plotfi
Reviewed By: efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82552
On PPC64, for a variadic function, if va_start is not called, it won't
access any variadic argument on stack, thus we can save stores of
registers used to pass arguments.
Differential Revision: https://reviews.llvm.org/D82361
Updated the AArch64 tests the best I could with my vague, inferred
understanding of AArch64 register banks. As far as I can tell, there
is only one 32-bit/64-bit type which will use the gpr register bank,
so we have to use the fpr bank for the other operand.
GFX10 image instructions use one or more address operands starting at
vaddr0, instead of a single vaddr operand, to allow for NSA forms.
Differential Revision: https://reviews.llvm.org/D81675
Fix the division/remainder algorithm by adding a second quotient
refinement step, which is required in some cases like
0xFFFFFFFFu / 0x11111111u (https://bugs.llvm.org/show_bug.cgi?id=46212).
Also document, rewrite and simplify it by ensuring that we always have a
lower bound on inv(y), which simplifies the UNR step and the quotient
refinement steps.
Differential Revision: https://reviews.llvm.org/D83381
ExpandVectorBuildThroughStack is also used for CONCAT_VECTORS.
However, when calculating the offsets for each of the operands we
incorrectly use the element size rather than actual size and thus
the stores overlap.
Differential Revision: https://reviews.llvm.org/D83303
Summary:
The following combine currently breaks in the DAGCombiner:
```
extract_vector_elt (concat_vectors v4i16:a, v4i16:b), x
-> extract_vector_elt a, x
```
This happens because after we have combined these nodes we have inserted nodes
that use individual instances of the vector element type. In the above example
i16. However this isn't a legal type on all backends, and when the combining pass calls
the legalizer it breaks as it expects types to already be legal. The type legalizer has
already been run, and running it again would make a mess of the nodes.
In the example code at least, the generated code is still efficient after the change.
Reviewers: miyuki, arsenm, dmgreen, lebedev.ri
Reviewed By: miyuki, lebedev.ri
Subscribers: lebedev.ri, wdng, hiraditya, steven.zhang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83231
vselect ((X & Pow2C) == 0), LHS, RHS --> vselect ((shl X, C') < 0), RHS, LHS
Follow-up to D83073 - the non-splat mask cases where we actually see an
improvement are quite limited from what I can tell. AVX1 needs multiply
and blend capabilities and AVX2 needs vector shift and blend capabilities.
The intersection of those 2 constraints is only vectors with 32-bit or
64-bit elements.
XOP is/was better.
Differential Revision: https://reviews.llvm.org/D83181
On SKX targets we end up loading a v16i8 PSHUFB mask from a v32i8 constant and scaling incorrectly indexes the demanded elts mask - we're missing a check that the constant pool is the same size as the loaded mask.
Test case from D81791 post-commit review.
We use extact_subvector and insert_subvector to "cast" between
fixed length and scalable vectors. This patch adds custom c++
based ISel for the following cases:
fixed_vector = ISD::EXTRACT_SUBVECTOR scalable_vector, 0
scalable_vector = ISD::INSERT_SUBVECTOR undef(scalable_vector), fixed_vector, 0
Which result in either EXTRACT_SUBREG/INSERT_SUBREG for NEON sized
vectors or COPY_TO_REGCLASS otherwise.
Differential Revision: https://reviews.llvm.org/D82871
For the GetElementPtr case in function
AddressingModeMatcher::matchOperationAddr
I've changed the code to use the TypeSize class instead of relying
upon the implicit conversion to a uint64_t. As part of this we now
check for scalable types and if we encounter one just bail out for
now as the subsequent optimisations doesn't currently support them.
This changes fixes up all warnings in the following tests:
llvm/test/CodeGen/AArch64/sve-ld1-addressing-mode-reg-imm.ll
llvm/test/CodeGen/AArch64/sve-st1-addressing-mode-reg-imm.ll
Differential Revision: https://reviews.llvm.org/D83124
`__stack_chk_fail` does not return, but `unreachable` was not generated
following `call __stack_chk_fail`. This had a possibility to generate an
invalid binary for functions with a return type, because
`__stack_chk_fail`'s return type is void and `call __stack_chk_fail` can
be the last instruction in the function whose return type is non-void.
Generating `unreachable` after it makes sure CFGStackify's
`fixEndsAtEndOfFunction` handles it correctly.
Reviewed By: tlively
Differential Revision: https://reviews.llvm.org/D83277
For an addition with an immediate in specific ranges, a pair of
addi-addi can be generated instead of the ordinary lui-addi-add serial.
Reviewed By: MaskRay, luismarques
Differential Revision: https://reviews.llvm.org/D82262
... to shift/add or shift/sub.
Do not enable it on riscv32 with the M extension where decomposeMulByConstant
may not be an optimization.
Reviewed By: luismarques, MaskRay
Differential Revision: https://reviews.llvm.org/D82660
handleAssignments was assuming every argument type is an MVT, and
assignArg would always fail. This fixes one of the hacks in the
current AMDGPU calling convention code that pre-processes the
arguments.
The tests in a5b9ad7e9aca1329ba310e638dafa58c47468a58 actually failed
the verifier, which for some reason is not the default. Also add tests
for 0-sized function arguments, which do not add entries to the
expected register lists.
This is inspired by D81648. The basic idea is to have the set of SDValues which are lowered as either constants or direct frame references explicit in one place, and to separate them clearly from the spilling logic.
This is not NFC in that the handling of constants larger than > 64 bit has changed. The old lowering would crash on values which could not be encoded as a sign extended 64 bit value. The new lowering just spills all constants > 64 bits. We could be consistent about doing the sext(Con64) optimization, but I happen to know that this code path is utterly unexercised in practice, so simple is better for now.
This patch creates a clang flag to enable SESES. This flag also ensures that
lvi-cfi is on when using seses via clang.
SESES should use lvi-cfi to mitigate returns and indirect branches.
The flag to enable the SESES functionality only without lvi-cfi is now
-x86-seses-enable-without-lvi-cfi to warn users part of the mitigation is not
enabled if they use this flag. This is useful in case folks want to see the
cost of SESES separate from the LVI-CFI.
Reviewed By: sconstab
Differential Revision: https://reviews.llvm.org/D79910
We were checking the VBROADCAST_LOAD element size against the extraction destination size instead of the extracted vector element size - PEXTRW/PEXTB have implicit zext'ing so have i32 destination sizes for v8i16/v16i8 vectors, resulting in us extracting from the wrong part of a load.
This patch bails from the fold if the vector element sizes don't match, and we now use the target constant extraction code later on like the pre-AVX2 targets, fixing the test case.
Found by internal fuzzing tests.
Use SESES as the fallback at O0 where the optimized LVI pass isn't desired due
to its effect on build times at O0.
I updated the LVI tests since this changes the code gen for the tests touched in the parent revision.
This is a follow up to the comments I made here: https://reviews.llvm.org/D80964
Hopefully we can continue the discussion here.
Also updated SESES to handle LFENCE instructions properly instead of adding
redundant LFENCEs. In particular, 1) no longer add LFENCE if the current
instruction being processed is an LFENCE and 2) no longer add LFENCE if the
instruction right before the instruction being processed is an LFENCE
Reviewed By: sconstab
Differential Revision: https://reviews.llvm.org/D82037
Since WebAssembly's vector shift instructions take a scalar shift
amount rather than a vector shift amount, we have to check in ISel
that the vector shift amount is a splat. Previously, we were checking
explicitly for splat BUILD_VECTOR nodes, but this change uses the
standard utilities for detecting splat values that can handle more
complex splat patterns. Since the C++ ISel lowering is now more
general than the ISel patterns, this change also simplifies shift
lowering by using the C++ lowering for all SIMD shifts rather than
mixing C++ and normal pattern-based lowering.
This change improves ISel for shifts to the point that the
simd-shift-unroll.ll regression test no longer tests the code path it
was originally meant to test. The bug corresponding to that regression
test is no longer reproducible with its original reported reproducer,
so rather than try to fix the regression test, this change just
removes it.
Differential Revision: https://reviews.llvm.org/D83278
On AVX2 we tend to lower BUILD_VECTOR of constants as broadcasts if we can, in this case a <2 x i16> non-uniform constant has been lowered as a <4 x i32> broadcast.
The test case shows that the extraction folding code has incorrectly extracted the wrong part (lower WORD) of the resulting i32 memory source.
Found by internal fuzzing tests.
Provide the LLVM intrinsics needed to implement vector replace element
builtins in altivec.h which will be added in a subsequent patch.
Differential Revision: https://reviews.llvm.org/D83308
In the test based on PR46586:
https://bugs.llvm.org/show_bug.cgi?id=46586
...we are inserting 16-bits into the high element of the vector, shuffling it
to element 0, and extracting 32-bits. But xmm1 was never initialized, so the
top 16-bits of the extract are undef without this patch.
(It seems like we could do better than this by recognizing that we only demand
a subsection of the build vector, but I want to make sure we fix the
miscompile 1st.)
This path is only used for pre-SSE4.1, and simpler patterns get squashed
somewhere along the way, so the test still includes a 'urem' as it did in the
original test from the bug report.
Differential Revision: https://reviews.llvm.org/D83319
When an argument has 'byval' attribute and should be
passed on the stack according calling convention,
a stack copy would be emitted twice. This will cause
the real value will be put into stack where the pointer
should be passed.
Differential Revision: https://reviews.llvm.org/D83175
Summary:
When splitting a store of a scalable type, the new address is
calculated in SplitVecOp_STORE using a vscale and an add instruction.
Reviewers: sdesmalen, efriedma, david-arm
Reviewed By: david-arm
Subscribers: tschuett, hiraditya, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83041
