In AArch64ISelLowering.cpp this patch implements this fold:
GEP (%ptr, (splat(%offset) + stepvector(A)))
into GEP ((%ptr + %offset), stepvector(A))
The above transform simplifies the index operand so that it can be expressed
as i32 elements.
This allows using only one gather/scatter assembly instruction instead of two.
Patch by Paul Walker (@paulwalker-arm).
Depends on D118459
Differential Revision: https://reviews.llvm.org/D117900
D116804 proposes to alter codegen on this example based on
CPU tuning, so check a variety of models to confirm it works
as expected. We already have this test mixed in with several
others in another test file, but it seems wasteful to add so
many RUN lines to check this difference over and over again.
In some cases, when selecting a (trunc (slr)) pattern, the slr gets translated
to a v_lshrrev_b3e2_e64 instruction whereas the truncation gets selected to
a sequence of v_and_b32_e64 and v_cmp_eq_u32_e64. In the final ISA, this appears
as selecting the nth-bit:
v_lshrrev_b32_e32 v0, 2, v1
v_and_b32_e32 v0, 1, v0
v_cmp_eq_u32_e32 vcc_lo, 1, v0
However, when the value used in the right shift is known at compilation time, the
whole sequence can be reduced to two VALUs when the constant operand in the v_and is adjusted to (1 << lshrrev_operand):
v_and_b32_e32 v0, (1 << 2), v1
v_cmp_ne_u32_e32 vcc_lo, 0, v0
In the example above, the following pseudo-code:
v0 = (v1 >> 2)
v0 = v0 & 1
vcc_lo = (v0 == 1)
would be translated to:
v0 = v1 & 0b100
vcc_lo = (v0 == 0b100)
which should yield an equivalent result.
This is a little bit hard to test as one needs to force the SelectionDAG to
contain the nodes before instruction selection, but the test sequence was
roughly derived from a production shader.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D118461
This teaches AArch64TargetLowering::shouldSinkOperands to sink the
operands of aarch64_neon_pmull intrinsic.
Differential Revision: https://reviews.llvm.org/D117944
Always set uniform metadata on the pointer if it is an instruction, but
otherwise do not bother to create a trivial getelementptr instruction,
because AMDGPUInstrInfo::isUniformMMO can already detect that various
non-instruction pointers are uniform.
Most of the test case churn is from tests that used undef as a pointer,
which AMDGPUInstrInfo::isUniformMMO treats as uniform.
Differential Revision: https://reviews.llvm.org/D118909
Tweak some of the tests to demonstrate
AMDGPUAnnotateUniformValues::visitLoadInst inserting a trivial
getelementptr instruction, just to have somewhere to put amdgpu.uniform
metadata. NFC.
Add support for the 'pause' hint instruction as an alias for
'fence w, 0'. To do this allow the 'fence' operands pred and succ
to be set to 0 (the empty set). This will also allow future hints
to be encoded as 'fence 0, <x>' and 'fence <x>, 0'.
This patch revised from @mundaym's D93019.
Reviewed By: asb
Differential Revision: https://reviews.llvm.org/D117789
Given an (integer) vecreduce, we know the order of the inputs does not matter.
We can convert UADDV(add(zext(extract_lo(x)), zext(extract_hi(x)))) into
UADDV(UADDLP(x)). This can also happen through an extra add, where we transform
UADDV(add(y, add(zext(extract_lo(x)), zext(extract_hi(x))))).
This makes sure the same thing happens signed cases too, which requires adding
a new SADDLP node.
Differential Revision: https://reviews.llvm.org/D118107
Unfortunately, it seems we really do need to take the long route;
start from the "merge" block, find (all the) "dispatch" blocks,
and deal with each "dispatch" block separately, instead of simply
starting from each "dispatch" block like it would logically make sense,
otherwise we run into a number of other missing folds around
`switch` formation, missing sinking/hoisting and phase ordering.
This reverts commit 85628ce75b3084dc0f185a320152baf85b59aba7.
This reverts commit c5fff9095342a792bf4b9a077fe3c3a83c4e566c.
This reverts commit 34a98e1046e3aa55e5f26ab20a15e96b4034d25a.
This reverts commit 1e353f092288309d74d380367aa50bbd383780ed.
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
This change adds a new Codegen test with auto-generated checks and
updates divergence-driven-trunc-to-i1.ll with auto-generated checks.
This is in preparation to D118461 to visualize the Codegen changes.
* LABEL is important to give a better diagnostic in case a check pattern fails
* Some NOT negative patterns are not effective. NEXT is useful to ensure
a code sequence has the desired instructions and report a better
diagnostic if something goes off.
* Since the ABI says the first parameter is in RDI. Replacing the
pattern `[[REG16:.*]]` with `RDI` should not cause maintenance burden.
Since the test is pretty mechanical, just use update_llc_test_checks.py
to re-generate it. Most functions can use `nounwind` to avoid CFI
directives.
Reviewed By: kstoimenov, vitalybuka
Differential Revision: https://reviews.llvm.org/D118864
In a future change, we will sometimes use a VGPR offset for doing
spills to memory, in which case we need 2 free VGPRs to do the SGPR
spill. In most cases we could spill the VGPR along with the SGPR being
spilled, but we don't have any free lanes for SGPR_1024 in wave32 so
we could still potentially need a second scavenging slot.
setjmp can return twice, but PostDominatorTree is unaware of this. as
such, it overestimates postdominance, leaving some cases where memory
does not get untagged on return. this causes false positives later in
the program execution.
this is a workaround for now, in the longer term PostDominatorTree
should be made aware of returns_twice, as this may cause problems
elsewhere.
See D118647 for equivalent fix to HWASan.
Reviewed By: eugenis
Differential Revision: https://reviews.llvm.org/D118749
This was using the ugly tablegenerated register enum names, which are
really hideous for register tuples on AMDGPU. Use the prettier names
which are recognized by the asm parser.
Warn on inline assembly clobbering reserved registers. It should also
warn on at least some reserved register defs, but that isn't happening
right now. If you have a def and re-use of a register we reserve, the
register coalescer will eliminate the intermediate virtual
register. When the reserved reg def is introduced later by the
backend, it will end up clobbering the value the register coalescer
assumed was live through the range.
There is also isInlineAsmReadOnlyReg, although I don't understand what
the distinction really is. It's called in SelectionDAGBuilder, long
before the set of reserved registers is frozen so I'm not sure how
that can possibly work reliably.
Unfortunately this is also using the ugly tablegenerated names for the
registers.
This adds or reuses ISD opcodes for vadd.wv, vaddu.wv, vadd.vv, vaddu.vv
and a similar set for sub.
I've included support for narrowing scalar splats that have known
sign/zero bits similar to what was done for MUL_VL.
The conversion to vwadd.vv proceeds in two phases. First we'll form
a vwadd.wv by narrowing one of the operands. Then we'll visit the
vwadd.wv to try to narrow the other operand. This turned out to be
simpler than catching all the cases in one step. The forming of of
vwadd.wv can happen for either operand for add, but only the right
hand side for sub since sub isn't commutable.
An interesting quirk is that ADD_VL and VZEXT_VL/VSEXT_VL are formed
during vector op legalization, but VMV_V_X_VL isn't usually formed
until op legalization when BUILD_VECTORS are handled. This leads to
VWADD_W_VL forming in one DAG combine round, and then a later DAG combine
round sees the VMV_V_X_VL and needs to commute the operands to get the
splat in position. This alone necessitated a VWADD_W_VL combine function
which made forming vwadd.vv in two stages an easy choice.
I've left out trying hard to form vwadd.wx instructions for now. It would
only save an extend in the scalar domain which isn't as interesting.
Might need to review the test coverage a bit. Most of the vwadd.wv
instructions are coming from vXi64 tests on rv64. The tests were
copy pasted from the existing multiply tests.
Reviewed By: rogfer01
Differential Revision: https://reviews.llvm.org/D117954
This allows us to set the noclobber flag on (the MMO of) a load
instruction instead of on the pointer. This fixes a bug where noclobber
was being applied to all loads from the same pointer, even if some of
them were clobbered.
Differential Revision: https://reviews.llvm.org/D118775
This patch introduces the conversions from math function calls
to MASS library calls. To resolves calls generated with these conversions, one
need to link libxlopt.a library. This patch is tested on PowerPC Linux and AIX.
Differential: https://reviews.llvm.org/D101759
Reviewer: bmahjour
The current `FoldTwoEntryPHINode()` is not quite designed correctly.
It starts from the merge point, and then tries to detect
the 'divergence' point.
Because of that, it is limited to the simple two-predecessor case,
where the PHI completely goes away. but that is rather pessimistic,
and it doesn't make much sense from the costmodel side of things.
For example if there is some other unrelated predecessor of
the merge point, we could split the merge point so that
the then/else blocks first branch to an empty block
and then to the merge point, and then we'd be able to speculate
the then/else code.
But if we'd instead simply start at the divergence point,
and look for the merge point, then we'll just natively support this case.
There's also the fact that `SpeculativelyExecuteBB()` already does
just that, but only if there is a single block to speculate,
and with a much more restrictive cost model.
But that also means we have code duplication.
Now, sadly, while this is as much NFCI as possible,
there is just no way to cleanly migrate to
the proper implementation. The results *are* going to be different
somewhat because of various phase ordering effects and SimplifyCFG
block iteration strategy.
LLVM has a couple of ways of producing ccmp - either from chains in isel
or from a later ifcvt style pass. This adds a simple DAG combine to
capture more cases, converting and(csel(0, 1, cc0), csel(0, 1, cc1))
into a csel(ccmp(.., cc0)), depending on cc1 (a SUBS in this case).
Differential Revision: https://reviews.llvm.org/D118327
This is an intentionally limited/different form of D90113.
That patch bravely tries to generalize folds where we pull
a binop into the arms of a select:
N0 + (Cond ? 0 : FVal) --> Cond ? N0 : (N0 + FVal)
...but it is not universally profitable.
This is the inverse of IR canonicalization as discussed in
D113442.
We know that this transform is not entirely profitable even
within x86, so we only handle x86 vector fadd/fsub as a 1st
step. The intent is to prevent AVX512 regressions as mentioned
in D113442.
The plan is to port this to DAGCombiner (so it will eventually
look more like D90113) and add more types/cases in pieces with
many more tests to verify that we are seeing improvements.
Differential Revision: https://reviews.llvm.org/D118644
This patch adds custom lowering support for ISD::SDIV and ISD::UDIV
when SVE is enabled, regardless of the minimum SVE vector length. We do
this because NEON simply does not have vector integer divide support, so
we want to take advantage of these instructions in SVE.
As part of this patch I've also simplified LowerToPredicatedOp to avoid
re-asking the same question about whether we should be using SVE for
fixed length vectors. Once we've made the decision to call
LowerToPredicatedOp, then we should simply assert we should be using SVE.
I've updated the 128-bit min SVE vector bits tests here:
CodeGen/AArch64/sve-fixed-length-int-div.ll
CodeGen/AArch64/sve-fixed-length-int-rem.ll
Differential Revision: https://reviews.llvm.org/D117871
Use the same mnemonics in the tests that are used in the AtomicLoadOp
pattern ($rd, $rr) but use RR1 instead of $operand. This matches similar
tests in load8.ll.
Differential Revision: https://reviews.llvm.org/D117991
This patch fixes the atomicrmw result value to be the value before the
operation instead of the value after the operation. This was a bug, left
as a FIXME in the code (see https://reviews.llvm.org/D97127).
From the LangRef:
> The contents of memory at the location specified by the <pointer>
> operand are atomically read, modified, and written back. The original
> value at the location is returned.
Doing this expansion early allows the register allocator to arrange
registers in such a way that commutable operations are simply swapped
around as needed, which results in shorter code while still being
correct.
Differential Revision: https://reviews.llvm.org/D117725
Previously, all children would be checked to see if any were an
explicit Register. If anywhere no commutable patterns would be
generated. This patch loosens the restriction to only check the
children that are being commuted.
Digging back through history, this code predates the existence of
commutable intrinsics and commutable SDNodes with more than 2
operands. At that time the loop would count the number of children that
weren't registers and if that was equal to 2 it would allow commuting.
I don't think this loop was re-considered when commutable
intrinsics were added or when we allowed SDNodes with more than 2
operands.
This important for RISCV were our isel patterns have a V0 mask
operand after the commutable operands on some RISCVISD opcodes.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D117955
MemorySSA considers any atomic a def to any operation it dominates
just like a barrier or fence. That is correct from memory state
perspective, but not required for the no-clobber metadata since
we are not using it for reordering. Skip such atomics during the
scan just like a barrier if it does not alias with the load.
Differential Revision: https://reviews.llvm.org/D118661
Currently we cannot convert a vector load into scalar if there
is dominating barrier or fence. It is considered a clobbering
memory access to prevent memory operations reordering. While
reordering is not possible the actual memory is not being clobbered
by a barrier or fence and we can still use a scalar load for a
uniform pointer.
The solution is not to bail on a first clobbering access but
traverse MemorySSA to the root excluding barriers and fences.
Differential Revision: https://reviews.llvm.org/D118419
These test cases all rely on a default target being specified. Adding
the requirement gets the tests properly skipped when
LLVM_DEFAULT_TARGET_TRIPLE is unset.
The first phase of the analysis can avoid a vsetvli if an earlier
instruction in the block used an SEW and LMUL that when combined with
the EEW of the load/store would produce the desired EMUL. If we
avoided a vsetvli this will affect the global analysis we do in the
second phase.
The third phase where we really insert the vsetvlis needs to agree
with the first phase. If it doesn't we can insert vsetvlis that
invalidate the global analysis.
In the test case there is a VSETVLI in the preheader that sets
SEW=64 and LMUL=1. Inside the loop there is a VADD with SEW=64 and LMUL=1.
This VADD is followed by a store that wants wants SEW=32 LMUL=1/2.
Because it has EEW=32 as part of the opcode the SEW=64 LMUL=1 from the
VADD can be become EMUL=1 for the store. So the first phase determines no
vsetvli is needed.
The third phase manages CurInfo differently than BBInfo.Change from the
first phase. CurInfo is only updated when we see a vsetvli or insert
a vsetvli. This was done to allow predecessor block information from
the global analysis to be applied to multiple instructions. Since the
loop body has no vsetvli we won't update CurInfo for either the VADD
or the VSE. This prevented us from checking the store vsetvli elision
for the VSE resulting in a vsetvli SEW=32 LMUL=1/2 being emitted which
invalidated the global analysis.
To mitigate this, I've added a BBLocalInfo variable that more closely
matches the first phase propagation. This gets updated based on the
VADD and prevents emitting a vsetvli for the store like we did in the
first phase.
I wonder if we should do an earlier phase to handle the load/store case
by adding more pseudo opcodes and changing the SEW/LMUL for those
instructions before the insertion analysis. That might be more robust
than trying to guarantee two phases make the same decision.
Fixes the test from D118629.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D118667
This patch updates the P10 patterns with a load feeding into an insertelt to
utilize the refactored load and store infrastructure, as well as updating any
tests that exhibit any codegen changes.
Furthermore, custom legalization is added for v4f32 on Power9 and above to not
only assist with adjusting the refactored load/stores for P10 vector insert,
but also it enables the utilization of direct moves.
Differential Revision: https://reviews.llvm.org/D115691
Pointer element types do not imply that the pointer is ABI aligned.
We should be using either an explicit align attribute here, or fall
back to an alignment of 1. This fixes a new element type access
introduced in D117764.
I don't think this makes any practical difference though, as the
lowering does not depend on alignment.
Differential Revision: https://reviews.llvm.org/D118681
