These are legal since we can do a 96-bit load on some subtargets, but
this is only for vector loads. If we can't widen the load, it needs to
be broken down once known scalar. For 16-byte alignment, widen to a
128-bit load.
The logic is written for what loads/stores should be selectable. There
are a set of cases that should be selectable, but due to missing MVTs
and/or selection patterns, will fail to select. I think eventually
load/store select patterns should ignore the type and only look at the
value size, but until that happens, bitcast these to equivalent i32
vectors.
Current implementation of division estimation isn't correct for some
cases like 1.0/0.0 (result is nan, not expected inf).
And this change exposes a potential infinite loop: we use
isConstOrConstSplatFP in combineRepeatedFPDivisors to look up if the
divisor is some constant. But it doesn't work after legalized on some
platforms. This patch restricts the method to act before LegalDAG.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D80542
This was implicitly assuming the branch instruction was the next after
the pseudo. It's possible for another non-terminator instruction to be
inserted between the intrinsic and the branch, so adjust the insertion
point. Fixes a non-terminator after terminator verifier error (which
without the verifier, manifested itself as an infinite loop in
analyzeBranch much later on).
Put AND before ADD in LegalizerHelper::lowerFPTRUNC_F64_TO_F16
in order to match algorithm from AMDGPUTargetLowering::LowerFP_TO_FP16.
Differential Revision: https://reviews.llvm.org/D81666
The baffling thing is this passed the OpenCL conformance test for
32-bit integer divisions, but only failed in the 32-bit path of
BypassSlowDivisions for the 64-bit tests.
There is an invalid cast produced when a pointee is a pointer
and the alloca type is cast to a pointer to int.
Differential Revision: https://reviews.llvm.org/D81606
Summary:
Make use of both the - (1) clustered bytes and (2) cluster length, to decide on
the max number of mem ops that can be clustered. On an average, when loads
are dword or smaller, consider `5` as max threshold, otherwise `4`. This heuristic
is purely based on different experimentation conducted, and there is no analytical
logic here.
Reviewers: foad, rampitec, arsenm, vpykhtin
Reviewed By: foad, rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, Anastasia, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D81085
In two instances of CreateStackTemporary we are sometimes promoting
alignments beyond the stack alignment. I have introduced a new function
called getReducedAlign that will return the alignment for the broken
down parts of illegal vector types. For example, on NEON a <32 x i8>
type is made up of two <16 x i8> types - in this case the sensible
alignment is 16 bytes, not 32.
In the legalization code wherever we create stack temporaries I have
started using the reduced alignments instead for illegal vector types.
I added a test to
CodeGen/AArch64/build-one-lane.ll
that tries to insert an element into an illegal fixed vector type
that involves creating a temporary stack object.
Differential Revision: https://reviews.llvm.org/D80370
The update_*test_checks scripts miss new stuff added at the end of
lines. Regenerate checks so the new mode register operands don't show
up in the diff of a future patch.
The annoying behavior where the output is different due to the
legality check struck again, plus the subtarget predicate wasn't
really correctly set for DS FP atomics.
Some of the FP min/max instructions seem to be in the gfx6/gfx7
manuals, but IIRC this might have been one of the cases where the
manual got ahead of the actual hardware support, but I've left these
as-is for now since the assembler tests seem to expect them.
The current set is an incomprehensible mess riddled with ordering
hacks for various limitations in the legalizer at the time of writing,
many of which have been fixed. This takes a very small step in
correcting this.
The core first change is to start checking for fully legal cases
first, rather than trying to figure out all of the actions that could
need to be performed. It's recommended to check the legal cases first
for faster legality checks in the common case. This still has a table
listing some common cases, but it needs measuring whether this really
helps or not.
More significantly, stop trying to allow any arbitrary type with a
legal bitwidth as a legal memory type, and start using the bitcast
legalize action for them. Allowing loads of these weird vector types
produced new burdens we don't need for handling all of the
legalization artifacts. Unlike the SelectionDAG handling, this is
still not casting 64 or 16-bit element vectors to 32-bit
vectors. These cases should still be handled by increasing/decreasing
the number of 16-bit elements. This is primarily to fix 8-bit element
vectors.
Another change is to stop trying to handle the load-widening based on
a higher alignment. We should still do this, but the way it was
handled wasn't really correct. We really need to modify the MMO's size
at the same time, and not just increase the result type. The
LegalizerHelper does not do this, and I think this would really
require a separate WidenMemory action (or to add a memory action
payload to the LegalizeMutation). These will now fail to legalize.
The structure of the legalizer rules makes writing concise rules here
difficult. It would be easier if the same function could answer the
query the query, and report the action to perform at the same
time. Instead these two are split into distinct predicate and action
functions. This is mostly tolerable for other cases, but the
load/store rules get pretty complicated so it's difficult to keep two
versions of these functions in sync.
Just computing the alignment makes sense without caring about the
general known bits, such as for non-integral pointers. Separate the
two and start calling into the TargetLowering hooks for frame indexes.
Start calling the TargetLowering implementation for FrameIndexes,
which improves the AMDGPU matching for stack addressing modes. Also
introduce a new hook for returning known alignment of target
instructions. For AMDGPU, it would be useful to report the known
alignment implied by certain intrinsic calls.
Also stop using MaybeAlign.
Global TableGen let override blocks are pretty dangerous and override
any local special cases. In this case, the broader HasFlatGlobalInsts
was overriding the more specific predicate for
FeatureAtomicFaddInsts. Make sure HasFlatGlobalInsts is implied by
FeatureAtomicFaddInsts, and make sure the right predicate is used.
One issue with independently setting the subtarget features on
incompatible targets is all of the encoding families do not define all
opcodes. This will hit an assert on gfx10 for example, since we set
the encoding independently based on the generation and not based on a
feature.
There are a number of MIR tests using instructions on subtargets where
they don't really exist. These are some of the easy cases that don't
require splitting up test functions.
The AMDGPU lowering for unconstrained G_FDIV sometimes needs to
introduce a mode switch in the middle, so it's helpful to have
constrained instructions available to legalize this. Right now nothing
is preventing reordering of the mode switch with the other
instructions in the expansion.
Record internal state based on register units. This is often more
efficient as there are typically fewer register units to update
compared to iterating over all the aliases of a register.
Original patch by Matthias Braun, but I've been rebasing and fixing it
for almost 2 years and fixed a few bugs causing intermediate failures
to make this patch independent of the changes in
https://reviews.llvm.org/D52010.
If we're only demanding the (shifted) sign bits of the shift source value, then we can use the value directly.
This handles SimplifyDemandedBits/SimplifyMultipleUseDemandedBits for both ISD::SHL and X86ISD::VSHLI.
Differential Revision: https://reviews.llvm.org/D80869
Explicitly set the exec mask for SGPR spills and reloads.
This fixes a bug where SGPR spills to memory could be incorrect
if the exec mask was 0 (or differed between spill and reload).
Additionally pack scalar subregisters (upto 16/32 per VGPR),
so that the majority of scalar types can be spilt or reloaded
with a simple memory access. This should amortize some of the
additional overhead of manipulating the exec mask.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D80282
These are scalar instructions that change vector instructions, so they
should not be executed without any active lanes.
The implementation of -amdgpu-skip-threshold also seem to be backwards
from expected, since decreasing it prevents removal.
The pass which infers when it's legal to load a global address space
as SMRD was only considering amdgpu_kernel, and ignoring the shader
entry type calling conventions.
Summary:
Combine unmerge(trunc) to enable other merge combines.
Without this combine, the scalar unmerge(trunc(merge))
pattern cannot be combined and easily lead to
hard-to-legalize merge/unmerge artifacts.
Reviewed By: arsenm
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79567
This is a custom inserter because it was less work than teaching
tablegen a way to indicate that it is sometimes OK to have a no side
effect instruction in the output of a side effecting pattern.
The asm is needed to look like a read of the mode register to prevent
it from being deleted. However, there seems to be a bug where the mode
register def instructions are moved across the asm sideeffect by the
post-RA scheduler.
Another oddity is the immediate is formatted differently between
s_denorm_mode and s_round_mode.
Summary:
PHIs result register class is set to VGPR or SGPR depending on the cross block value divergence.
In some cases uniform PHI need to be converted to return VGPR to prevent the oddnumber of moves values from VGPR to SGPR and back.
PHI should certainly return VGPR if it has at least one VGPR input. This change adds the exception.
We don't want to convert uniform PHI to VGPRs in case the only VGPR input is a VGPR to SGPR COPY and definition od the
source VGPR in this COPY is move immediate.
bb.0:
%0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
%2:sreg_32 = .....
bb.1:
%3:sreg_32 = PHI %1, %bb.3, %2, %bb.1
S_BRANCH %bb.3
bb.3:
%1:sreg_32 = COPY %0
S_BRANCH %bb.2
Reviewers: rampitec
Reviewed By: rampitec
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D80434
OpenMP emits these for some reason, so handle them. Assume these use
4096 bytes by default, with a flag to override this. Also change the
related stack assumption for calls to have a flag.
We do not have register classes for all possible vector
sizes, so round it up for extract vector element.
Also fixes selection of G_MERGE_VALUES when vectors are
not a power of two.
This has required to refactor getRegSplitParts() in way
that it can handle not just power of two vectors.
Ideally we would like RegSplitParts to be generated by
tablegen.
Differential Revision: https://reviews.llvm.org/D80457
- This allow us to specify the (minimal) alignment on an intrinsic's
arguments and, more importantly, the return value.
Differential Revision: https://reviews.llvm.org/D80422
This is the groundwork required to implement strictfp. For now, this
should be NFC for regular instructoins (many instructions just gain an
extra use of a reserved register). Regalloc won't rematerialize
instructions with reads of physical registers, but we were suffering
from that anyway with the exec reads.
Should add it for all the related FP uses (possibly with some
extras). I did not add it to either the gpr index mode instructions
(or every single VALU instruction) since it's a ridiculous feature
already modeled as an arbitrary side effect.
Also work towards marking instructions with FP exceptions. This
doesn't actually set the bit yet since this would start to change
codegen. It seems nofpexcept is currently not implied from the regular
IR FP operations. Add it to some MIR tests where I think it might
matter.
All 3 passes that change instruction encodings were dropping MI
flags. This avoids scheduling regressions caused by setting
mayRaiseFPExceptions on FP instructions for non-strictfp functions.
The vector equivalent has backwards operands, but the scalar version
does not. The passes that use these hooks aren't enabled by default,
so this doesn't really change anything.
I'm guessing this was a holdover from when 0 was an invalid stack
pointer, but surprised nobody has discovered this before.
Also don't allow offset folding for -1 pointers, since it looks weird
to partially fold this.
I consider this to be a hack, since we probably should not mark any
16-bit extract as legal, and require all extracts to be done on
multiples of 32. There are quite a few more battles to fight in the
legalizer for sub-dword vectors, so just select this for now so we can
pass OpenCL conformance without crashing.
Also fix the same assert for G_INSERTs. Unlike G_EXTRACT there's not a
trivial way to select this so just fail on it.
Confusingly, these were unrelated and had different semantics. The
G_PTR_MASK instruction predates the llvm.ptrmask intrinsic, but has a
different format. G_PTR_MASK only allows clearing the low bits of a
pointer, and only a constant number of bits. The ptrmask intrinsic
allows an arbitrary mask. Replace G_PTR_MASK to match the intrinsic.
Only selects the cases that look like the old instruction. More work
is needed to select the general case. Also new legalization code is
still needed to deal with the case where the incoming mask size does
not match the pointer size, which has a specified behavior in the
langref.
This is currently missing most of the hard parts to lower correctly,
so disable it for now. This fixes at least one OpenCL conformance test
and allows it to pass with fallback. Hide this behind an option for
now.
Summary: 'A' constraint requires an immediate int or fp constant that can be inlined in an instruction encoding.
Reviewers: arsenm, rampitec
Differential Revision: https://reviews.llvm.org/D78494
EarlyCSE was added with D75145, but the motivating test is
not regressed by removing the extra pass now. That might be
because VectorCombine altered the way it processes instructions,
or it might be from (re)moving VectorCombine in the pipeline.
The extra round of EarlyCSE appears to cost approximately
0.26% in compile-time as discussed in D80236, so we need some
evidence to justify its inclusion here, but we do not have
that (yet).
I suspect that between SLP and VectorCombine, we are creating
patterns that InstCombine and/or codegen are not prepared for,
but we will need to reduce those examples and include them as
PhaseOrdering and/or test-suite benchmarks.
As noted in D80236, moving the pass in the pipeline exposed this
shortcoming. Extra work to recalculate the alias results showed
up as a compile-time slowdown.
There are 2 known problem patterns shown in the test diffs here:
vector horizontal ops (an x86 specialization) and vector reductions.
SLP has greater ability to match and fold those than vector-combine,
so let SLP have first chance at that.
This is a quick fix while we continue to improve vector-combine and
possibly canonicalize to reduction intrinsics.
In the longer term, we should improve matching of these patterns
because if they were created in the "bad" forms shown here, then we
would miss optimizing them.
I'm not sure what is happening with alias analysis on the addsub test.
The old pass manager now shows an extra line for that, and we see an
improvement that comes from SLP vectorizing a store. I don't know
what's missing with the new pass manager to make that happen.
Strangely, I can't reproduce the behavior if I compile from C++ with
clang and invoke the new PM with "-fexperimental-new-pass-manager".
Differential Revision: https://reviews.llvm.org/D80236
Unlike SelectionDAGBuilder, IRTranslator omits the unconditional
branch in fallthrough cases. Confusingly, the control flow pseudos
function in the opposite way the intrinsics are used, and the branch
targets always need to be swapped. We're inverting the target blocks,
so we need to figure out the old fallthrough block and insert a branch
to the original unconditional branch target.
This only affects assembly and -filetype=asm codegen of PAL metadata.
Differential Revision: https://reviews.llvm.org/D78860
Change-Id: I7b822e1917bf7b403486820d31afc483be207652
Matt Arsenault