The patch did not account for one corner case where cmp does not dominate
the loop latch. This patch adds this check, hopefully it's cheap because
the CFG does not change during the transform, so DT queries should be
executed quickly.
If you see compile time slowness from this, please revert.
Differential Revision: https://reviews.llvm.org/D96119
Move implementation of kill intrinsics to WQM pass. Add live lane
tracking by updating a stored exec mask when lanes are killed.
Use live lane tracking to enable early termination of shader
at any point in control flow.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D94746
With t2DoLoopDec we can be left with some extra MOV's in the preheaders
of tail predicated loops. This removes them, in the same way we remove
other dead variables.
Differential Revision: https://reviews.llvm.org/D91857
We were storing predicate registers, such as a <8 x i1>, in the opposite
order to how the rest of llvm expects. This actually turns out to be
correct for the one place that usually uses it - the
ScalarizeMaskedMemIntrin pass, but only because the pass was incorrect
itself. This fixes the order so that bits are stored in the opposite
order and bitcasts work as expected. This allows the Scalarization pass
to be fixed, as in https://reviews.llvm.org/D94765.
Differential Revision: https://reviews.llvm.org/D94867
Function `replaceMathCmpWithIntrinsic` artificially limits the scope
of the optimization, setting a requirement of two instructions be in
the same block, due to two reasons:
- usage of DT for more general check is costly in terms of compile time;
- risk of creating a new value that lives through multiple blocks.
Because of this, two semantically equivalent tests may be or not be the
subject of this opt depending on where the binary operation is located.
See `test/CodeGen/X86/usub_inc_iv.ll` for motivation
There is one important particular case where this limitation is too strict:
it is when the binary operation is the increment of the induction variable.
As result, the application of this opt becomes fragile and highly reliant on
where other passes decide to place IV increment. In most cases, they place
it in the end of the latch block, killing the opt opportunity (when in fact it
does not matter where to insert the actual instruction).
This patch handles this particular case separately.
- The detector does not use dom tree and has constant cost;
- The value of IV or IV.next lives through all loop in any case, so this should not
create a new unexpected long-living value.
As result, the transform becomes more robust. It also seems to lead to
better code generation in some cases (see `test/CodeGen/X86/lsr-loop-exit-cond.ll`).
Differential Revision: https://reviews.llvm.org/D96119
Reviewed By: spatel, reames
A G_MUL + G_PTR_ADD can also be folded into a madd. So, conservatively, we
shouldn't combine when the G_MUL is used by a G_PTR_ADD either.
Differential Revision: https://reviews.llvm.org/D96457
GlobalISel was only doing this with minsize. SDAG does this with optsize.
(See: `SelectionDAG::shouldOptForSize()`)
This is a 0.3% code size improvement for CTMark at -Os.
(Best: 1.1% improvements on lencod + pairlocalalign)
Differential Revision: https://reviews.llvm.org/D96451
When we have a G_ADD which is fed by a G_ICMP on one side, we can fold it into
the cset for the G_ICMP.
e.g. Given
```
%cmp = G_ICMP ... %x, %y
%add = G_ADD %cmp, %z
```
We would normally emit a cmp, cset, and add.
However, `%add` is either `%z` or `%z + 1`. So, we can just use `%z` as the
source of the cset rather than wzr, saving an instruction.
This would probably be cleaner in AArch64PostLegalizerLowering, but we'd need
to change the way we represent G_ICMP to do that, I think. For now, it's
easiest to implement in selection.
This is a 0.1% code size improvement on CTMark/pairlocalalign at -Os.
Example: https://godbolt.org/z/7KdrP8
Differential Revision: https://reviews.llvm.org/D96388
The test cases extract a fixed element from a vector and splat it
into a vector. This gets DAG combined into a splat shuffle.
I've used some very wide vectors in the test to make sure we have
at least a couple tests where the element doesn't fit into the
uimm5 immediate of vrgather.vi so we fall back to vrgather.vx.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D96186
Avoid doing the following combine for vector types:
```
copysign(x, fp_extend(y)) -> copysign(x, y)
copysign(x, fp_round(y)) -> copysign(x, y)
```
That combine seemed to impede the selection of vector instruction and cause
a mess in some circumstances.
Differential Revision: https://reviews.llvm.org/D96037
This patch optimizes a build_vector "index sequence" and lowers it to
the existing custom RISCVISD::VID node. This pattern is common in
autovectorized code.
The custom node was updated to allow it to be used by both scalable and
fixed-length vectors, thus avoiding pattern duplication.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D96332
Similar to the case for G_ADD.
There was a function in CTMark/pairlocalalign which was missing this case,
causing GlobalISel to emit a add + csel when a csinc is all that is necessary.
https://godbolt.org/z/ax69E9
Minor code size improvements on CTMark at -Os.
Differential Revision: https://reviews.llvm.org/D96390
We need to avoid setting the kill flag on the CSR spill if there's an
additional use of the register after the spill.
This does rely on consistency between the entry block liveins and the
MRI's function live ins, which is not something the verifier checks
now.
As of the current draft these are no longer being considered
for the bitmanip spec. It wasn't clear what sub extension they
belonged in in the 0.93 spec.
So remove them. They can always be added back if something changes.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D96157
This reverts commit 4a64d8fe392449b205e59031aad5424968cf7446.
Makes clang crash when buildling trivial iOS programs, see comment
after https://reviews.llvm.org/D92808#2551401
Fold shufps(hop(x,y),hop(z,w)) -> permute(hop(x,z)) - this is very similar to the equivalent unpack fold.
I did start trying to convert foldShuffleOfHorizOp to handle generic shuffle masks but we're relying on a lot of special cases at the moment.
As of commit 284f2bffc9bc5, the DAG Combiner gets rid of the masking of the
input to this node if the mask only keeps the bottom 16 bits. This is because
the underlying library function does not use the high order bits. However, on
PowerPC's ELFv2 ABI, it is the caller that is responsible for clearing the bits
from the register. Therefore, the library implementation of __gnu_h2f_ieee will
return an incorrect result if the bits aren't cleared.
This combine is desired for ARM (and possibly other targets) so this patch adds
a query to Target Lowering to check if this zeroing needs to be kept.
Fixes: https://bugs.llvm.org/show_bug.cgi?id=49092
Differential revision: https://reviews.llvm.org/D96283
References to functions are in program memory and need a `pm()` fixup. This should fix trait objects for Rust on AVR.
Differential Revision: https://reviews.llvm.org/D87631
Patch by Alex Mikhalev.
As for SETCC, use a less expensive condition code when generating
STRICT_FSETCC if the node is known not to have Nan.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D91972
In vector v0.10, there are whole vector register load/store
instructions. I suggest to use the whole register load/store
instructions for generic load/store for scalable vector types. It could
save up vset{i}vl{i} for these load/store.
For fractional LMUL, I keep to use vle{eew}.v/vse{eew}.v instructions to
load/store partial vector registers.
Differential Revision: https://reviews.llvm.org/D95853
A One-Off Identity mask is a shuffle that is mostly an identity mask
from as single source but contains a single element out-of-place, either
from a different vector or from another position in the same vector. As
opposed to lowering this via a ARMISD::BUILD_VECTOR we can generate an
extract/insert pair directly. Under ARM with individually accessible
lane elements this often becomes a simple lane move.
This also alters the LowerVECTOR_SHUFFLEUsingMovs code to use v4f32 (not
v4i32), a more natural type for lane moves.
Differential Revision: https://reviews.llvm.org/D95551
On AArch64 (which seems to be the only target that supports it), this
attribute allows codegen to avoid saving/restoring the value in x0
across a call.
Gives a 0.1% geomean -Os code size improvement on CTMark.
Differential Revision: https://reviews.llvm.org/D96099
This reverts commit 0fc1738eb75d613b9e16143b83e7cb80512e84eb.
The test passes (unexpectedly, due to the XFAIL: *) when x86 isn't
the default triple (such as on an arm machine).
As the actual MSVC toolset doesn't use the GAS-style assembly that
Clang/LLVM produces and consumes, there's no reference for what
string to use for e.g. comments when building with a MSVC triple.
This frees up the use of semicolon as separator string, just like
was done for GNU targets in 23413195649d0cf6f3860ae8b5fb115b35032075.
(Previously, both the separator and comment strings were set to
the same, a semicolon.)
Compiler-rt extensively uses separator chars in its assembly,
and that assembly should be buildable with clang-cl for MSVC too.
Differential Revision: https://reviews.llvm.org/D96259
Building on the fixed vector support from D95705
I've added ISD nodes for vmv.v.x and vfmv.v.f and switched to
lowering the intrinsics to it. This allows us to share the same
isel patterns for both.
This doesn't handle splats of i64 on RV32 yet. The build_vector
gets converted to a vXi32 build_vector+bitcast during type
legalization. Not sure the best way to handle this at the moment.
Differential Revision: https://reviews.llvm.org/D96108
This is an alternative to D95563.
This is modeled after a similar feature for AArch64's SVE that uses
predicated scalable vector instructions.a
Rather than use predication, this patch uses an explicit VL operand.
I've limited it to always use LMUL=1 for now, but we can improve this
in the future.
This requires a bunch of new ISD opcodes to carry the VL operand.
I think we can probably lower intrinsics to these ISD opcodes to
cut down on the size of the isel table. Which is why I've added
patterns for all integer/float types and not just LMUL=1.
I'm only testing one vector width right now, but the width is
programmable via the command line.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D95705
This adds support for commuting operands and converting between
vfmadd and vfmacc to avoid register copies.
To avoid messing up intrinsic behavior, I've added new pseudo
instructions that have the isCommutable flag set. These pseudos also
force a tail agnostic policy. The intrinsic version still use
the tail undisturbed policy.
For best results it looks like we need to start with fmadd and only
pick fmacc if its beneficial. MachineCSE commutes without contraining
the operands and then commutes back if it didn't help with CSE. So
I've made sure that when the operand choice isn't constrained, we
will keep fmadd for MachineCSE and when it does the second commute,
we get back the original instruction.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D95800
More MachO madness for everyone. MachO relocations are only 32-bits, which
means the ARM64_RELOC_ADDEND one only actually has 24 (signed) bits for the
actual addend. This is a problem when calculating the address of a basic block;
because it has no symbol of its own, the sequence
adrp x0, Ltmp0@PAGE
add x0, x0, x0 Ltmp0@PAGEOFF
is represented by relocation with an addend that contains the offset from the
function start to Ltmp, and so the largest function where this is guaranteed to
work is 8MB. That's not quite big enough that we can call it user error (IMO).
So this patch puts the any blockaddress into a constant-pool, where the addend
is instead stored in the (x)word being relocated, which is obviously big enough
for any function.
Different targets might handle branch performance differently, so this patch allows for
targets to specify the TailDuplicateSize threshold. Said threshold defines how small a branch
can be and still be duplicated to generate straight-line code instead.
This patch also specifies said override values for the AArch64 subtarget.
Differential Revision: https://reviews.llvm.org/D95631
When running the tests on PowerPC and x86, the lit test GlobalISel/trunc.ll fails at the memory sanitize step. This seems to be due to wrong invalid logic (which matches even if it shouldn't) and likely missing variable initialisation."
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D95878
Because we mark all operations as expand for v2f64, scalar_to_vector
would end up lowering through a stack store/reload. But it is pretty
simple to implement, only inserting a D reg into an undef vector. This
helps clear up some inefficient codegen from soft calling conventions.
Differential Revision: https://reviews.llvm.org/D96153
This patch adds support for both the fadd reduction intrinsic, in both
the ordered and unordered modes.
The fmin and fmax intrinsics are not currently supported due to a
discrepancy between the LLVM semantics and the RVV ISA behaviour with
regards to signaling NaNs. This behaviour is likely fixed in version 2.3
of the RISC-V F/D/Q extension, but until then the intrinsics can be left
unsupported.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D95870
