The way fallback to SelectionDAG works is somewhat surprising to
me. When the fallback path is enabled, the entire set of SelectionDAG
selector passes is added to the pass pipeline, and each one needs to
check if the function was selected. This results in the surprising
behavior of running SIFixSGPRCopies for example, but only if
-global-isel-abort=2 is used.
SIAddIMGInitPass is also added in addInstSelector, but I'm not sure
why we have this pass or if it should be added somewhere else for
GlobalISel.
Produce an unmerge to a narrower type and introduce a narrower shift
if needed. I wasn't sure if there was a better way to parameterize the
target's preferred shift type for the GICombineRule, so manually call
the combine helper.
While looking at the output on real sized programs, there is a lot of
extra SGPR spilling compared to the DAG path. This seems to largely be
from all constants being SGPRs in the entry block.
Try to handle arbitrary scalar BFEs by packing the operands. The DAG
gives up on non-constant arguments. We're still missing any constant
folding, so we end up with pretty ugly code most of the time. Also
handle the 64-bit scalar case, which the DAG doesn't try to do.
The division expansions in AMDGPUCodeGenPrepare can't be relied on for
correctness, since they punt to later optimization and possibly
legalization in some cases. We still need a way to be able to write
tests for the legalizer versions of the expansion. This is mostly for
GlobalISel, since the expected optimzations is expecting aren't
implemented.
The interaction with the flag to expand 64-bit division in the IR is
pretty confusing, but these flags have different purposes.
I didn't realize we were already expanding 24/32-bit division here
already. Use the available IntegerDivision utilities. This uses loops,
so produces significantly smaller code than the inline DAG expansion.
This now requires width reductions of 64-bit divisions before
introducing the expanded loops.
This helps work around missing legalization in GlobalISel for
division, which are the only remaining core instructions that didn't
work at all.
I think this is plausibly a better implementation than exists in the
DAG, although turning it on by default misses out on the constant
value optimizations and also needs benchmarking.
This is more or less directly ported from the AMDGPU custom lowering
for FP_TO_FP16. I made a few minor fixups (using G_UNMERGE_VALUES
instead of creating shift/trunc to extract the two halves, and zexting
an inverted compare instead of select_cc).
This also does not include the fast math expansion the DAG which
converts to f32 and then to f16. I think that belongs in a
pre-legalize combine instead.
Also greatly improve i64 lowering. LegalizeIntegerTypes does the
correct narrowing if i64 isn't legal. Just workaround this for
SelectionDAG by making i64 legal and splitting in the patterns.
When SI_IF is inserted, it constrains the source register with a
register class, which was quite likely a G_ICMP. This was incorrectly
treating it as a scalar, and then applyMappingImpl would end up
producing invalid MIR since this was unexpected.
Also fix not using all VGPR sources for vcc outputs.
Load extra bits if suitably aligned. This allows using widened
3-vector loads on SI, and fixes legalization for <9 x s32> (which LSV
apparently forms frequently on lowered kernel argument lists).
Fix incorrectly treating these as legal on SI. This should emit a
64-bit store and a 32-bit store.
I think all of the load and store rules are just about complete, but
due for a rewrite.
The isNegatibleForFree/getNegatedExpression methods currently rely on a raw char value to indicate whether a negation is beneficial or not.
This patch replaces the char return value with an NegatibleCost enum to more clearly demonstrate what is implied.
It also renames isNegatibleForFree to getNegatibleCost to more accurately reflect whats going on.
Differential Revision: https://reviews.llvm.org/D74221
Summary:
Consider:
%r = call i32 @llvm.amdgcn.writelane(i32 0, i32 1, i32 2)
This produces a value that is 0 on lane 1, and 2 everywhere else; i.e.,
it is divergent.
Reported-by: Marek Olsak <Marek.Olsak@amd.com>
Reviewers: arsenm, foad, mareko
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74400
This was creating a select on true/false values, and then comparing
that later. This produced more work for later combines, which can be
avoided by just using the boolean values. This was copied from the
original DAG expansion, which also has the same problem. This doesn't
have a observable change using SelectionDAG, but since GlobalISel is
missing these optimizations, the final code was noticeably longer.
These have nicer expansions implemented in the DAG. Ideally we would
either directly implement all of these special expansions, or stop
expanding division in the IR.
This is apparently worse than 1-byte alignment. This does not attempt
to decompose 2-byte aligned wide stores, but will stop trying to
produce them.
Also fix bug in LoadStoreVectorizer which was decreasing the alignment
and vectorizing stack accesses. It was assuming a stack object was an
alloca that could have its base alignment changed, which is not true
if the pointer is derived from a function argument.
Since natural fdiv lowering is now more conservative even with
denormals disabled, we get a slower expansion from just a plain
1.0/fdiv. Directly emit the rcp intrinsic when using it to implement
integer division to avoid a pointlessly complex sequence.
Summary:
SIInstrInfo::expandPostRAPseudo converts ENTER_WWM in-place into an
S_OR_SAVEEXEC instruction that needs certain implicit operands. Without
this patch I get errors like this that make it harder to use -stop-after
to bisect the pass pipeline:
$ llc -march=amdgcn test/CodeGen/AMDGPU/wqm.ll -stop-after=postrapseudos -o - | sed -E 's/ (from|into) custom "TargetCustom[0-9]+"//' | llc -march=amdgcn -x=mir
error: <stdin>:1295:70: missing implicit register operand 'implicit-def $scc'
renamable $sgpr2_sgpr3 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec
^
Note that this error is currently only generated by MIParser but it
comes with a FIXME comment:
// FIXME: Move the implicit operand verification to the machine verifier.
Reviewers: critson, arsenm, rampitec, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74428
Summary:
Add a new method (tryParseRegister) that attempts to parse a register specification.
MASM allows the use of IFDEF <register>, as well as IFDEF <symbol>. To accommodate this, we make it possible to check whether a register specification can be parsed at the current location, without failing the entire parse if it can't.
Reviewers: thakis
Reviewed By: thakis
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73486
Summary:
As far as I know this did not affect code generation, but it did affect
the order of -debug-only=si-wqm output and the naming of autonamed
values in -print-after=si-wqm output.
Reviewers: arsenm, rampitec, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, mgrang, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74317
These are generated and do not need to have the same values.
We are defining separate subregs for R600 and GCN but then
using AMDGPU subregs on R600.
Differential Revision: https://reviews.llvm.org/D74248
Based on D72931
This adds a new feature called A16 which is enabled for gfx10.
gfx9 keeps the R128A16 feature so it can share all the instruction encodings
with gfx7/8.
Differential Revision: https://reviews.llvm.org/D73956
Vector indexing with a constant index should be folded out in the
legalizer, but this was accidentally falling through. This would
produce the indexing operation with $noreg. Handle this case as a
dynamic index just in case a bug like this happens again in the
future.
We were failing to find constants that were casted. I feel like the
artifact combiner should have folded the constant in the trunc before
the custom lowering, but that doesn't happen.
Reverts part of 6524a7a2b9. Since that
commit, the expansion was ignoring the actual save exec register
produced by the instruction, and looking at other instructions. I do
not understand why it was looking at other instructions, but relying
on this scan was wrong.
Fixes verifier errors after SI_IF is tail duplicated, which should be
correct to do. The results were fed into a phi, which was lowered to
the S_MOV_B64_term instructions.
Summary:
The accuracy limit to use rcp is adjusted to 1.0 ulp from 2.5 ulp.
Also, afn instead of arcp is used to allow inaccurate rcp to be used.
Reviewers:
arsenm
Differential Revision: https://reviews.llvm.org/D73588
Summary:
Current implementation of matchSwap in SIShrinkInstructions searches the entire
use_nodbg_operands set to find the possible pattern to generate v_swap instruction.
This approach will lead to a O(N^3) in compile time for SIShrinkInstructions.
But in reality, the matching pattern only exists within nearby instructions in the
same basic block. This work limits the search to a maximum of 16 instructions, and has
a linear compile time comsumption.
Reviewers:
rampitec, arsenm
Differential Revision: https://reviews.llvm.org/D74180
Allows more flexible use of buildMerge in places where
use operands are available as SrcOp since it does not
require explicit conversion to Register.
Simplify code with new buildMerge.
Differential Revision: https://reviews.llvm.org/D74223
We were executing this in a waterfall loop as a placeholder, but this
should really be converted to a MUBUF load. Also execute in a
waterfall loop if the resource isn't an SGPR. This is a case where the
DAG handling was wrong because doing the right thing was too hard.
Currently, this will mishandle 96-bit loads. There's currently no way
to track the original memory size with an MMO, so these loads will be
widened andd the resulting memory size will be 128-bits.
The type passed to lower was invalid, so I'm not sure how this was
even working before. The source and destination type also do not have
to match, so make sure to use the right ones.
Summary: This patch introduces an API for MemOp in order to simplify and tighten the client code.
Reviewers: courbet
Subscribers: arsenm, nemanjai, jvesely, nhaehnle, hiraditya, kbarton, jsji, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73964
We are using countPopulation on a LaneBitmask to determine
a number of registers it covers. This is the assumption which
does not necessarily need to be true. It is not changed but
factored into a single call SIRegisterInfo::getNumCoveredRegs().
Some other places are cleaned up with respect to assumptions
about subreg indexes values and tablegen behavior.
Differential Revision: https://reviews.llvm.org/D74177
Summary:
Current implementation of matchSwap in SIShrinkInstructions searches the entire
use_nodbg_operands set to find the possible pattern to generate v_swap instruction.
This approach will lead to a O(N^3) in compile time for SIShrinkInstructions.
But in reality, the matching pattern only exists within nearby instructions in the
same basic block. This work limits the search to a maximum of 16 instructions, and has
a linear compile time comsumption.
Reviewers:
rampitec, arsenm
Differential Revision: https://reviews.llvm.org/D74180
When multiple instructions are moved into a waterfall loop, it's
possible some of them re-use the same operands. Avoid creating
multiple sequences of readfirstlanes for them. None of the current
uses will hit this, but will be used in a future patch.
Really the intrinsic definition is wrong, but work around this
here. The DAG lowering introduces an MMO. We have to introduce a new
operation to avoid the verifier complaining about the missing mayLoad.
Use cmp ord instead of cmp_class compared to the DAG version for the
nan check, but mostly try to match the existsing pattern.
I think the sign doesn't matter for fract, so we could do a little
better with the source modifier matching.
I think this is also still broken as in D22898, but I'm leaving it
as-is for now while I don't have an SI system to test on.
Rewrite the result register pair into the expected sinigle register
format in the legalizer.
I'm also operating under the assumption that TFE doesn't apply to
stores or atomics, but don't know if this is true or not.
The mask results of these should be uniform. The trickier part is the
dummy booleans used as IR glue need to be treated as divergent. This
should make the divergence analysis results correct for the IR the DAG
is constructed from.
This should allow us to eliminate requiresUniformRegister, which has
an expensive, recursive scan over all users looking for control flow
intrinsics. This should avoid recent compile time regressions.
The 96-bit results need to be widened.
I find the interaction between LegalizerHelper and MIRBuilder somewhat
awkward. The custom legalization is called by the LegalizerHelper, but
then does not have access to the helper. You have to construct a new
helper, which then does not own the MachineIRBuilder, but does modify
it. Maybe custom legalization should be passed the helper?
The adjusted iterator range included the last we just inserted, and
don't want to process. Figure out the new iterator range before
inserting phis. This was a harmless problem, but added an unnecessary
complication for a future patch.
If we have s_pack_* instructions, legalize this to
G_BUILD_VECTOR_TRUNC from s32 elements. This is closer to how how the
s_pack_* instructions really behave.
If we don't have s_pack_ instructions, expand this by creating a merge
to s32 and bitcasting. This expands to the expected bit operations. I
think this eventually should go in a new bitcast legalize action type
in LegalizerHelper.
We already directly emit the shift operations in RegBankSelect for the
vector case. This could possibly be cleaned up, but I also may want to
defer doing this expansion to selection anyway. I'll see about that
when I try to actually match VOP3P instructions.
This breaks the selection of the build_vector since tablegen doesn't
know how to match G_BUILD_VECTOR_TRUNC yet, so just xfail it for now.
scalar_to_vector takes only one argument, not two.
The a16 tests now also check the packing of coordinates into registers
Differential Revision: https://reviews.llvm.org/D73482
This should lower the amount of used registers for gfx9.
I updated some of the changed tests with the update script because
changing them by hand is tedious.
Differential Revision: https://reviews.llvm.org/D73884
The old version might be faster on EG (RECIP_IEEE is Trans only),
but it'd need extra corner case checks.
This gives correct corner case behaviour and saves a register.
Fixes OCL CTS sqrt test (1-thread, scalar) on Turks.
Reviewer: arsenm
Differential Revision: https://reviews.llvm.org/D74017
This was incorrectly rounding up to the next power of 2. v4f32 was
rounding up to v8f32, which was just wrong. There are also v3i16/v3f16
available in MVT, so we don't even need to round the f16 cases
anymore. Additionally, this field is really an EVT so we don't even
need to consider this.
Also switch some asserts to return invalid. We should have an IR
verifier for these intrinsic return types, but for now it's better to
not assert on IR that passes the verifier.
This should also probably be fixed to consider that dmask is really
eliminating some of the loaded components.
https://reviews.llvm.org/D72312 introduced an infinite loop which involves
DAGCombiner::visitFMA and AMDGPUTargetLowering::performFNegCombine.
fma( a, fneg(b), fneg(c) ) => fneg( fma (a, b, c) ) => fma( a, fneg(b), fneg(c) ) ...
This only breaks with types where 'isFNegFree' returns flase, e.g. v4f32.
Reproducing the issue also needs the attribute 'no-signed-zeros-fp-math',
and no source mods allowed on one of the users of the Op.
This fix makes changes to indicate that it is not free to negate a fma if it
has users with source mods.
Differential Revision: https://reviews.llvm.org/D73939
Prepare to accurately track the future denormal-fp-math attribute
changes. The way to actually set these separately is not wired in yet.
This is just a mechanical change, and mostly still assumes the input
and output mode match. This should be refined for some cases. For
example, fcanonicalize lowering should use the flushing variant if
either input or output flushing is enabled
The usage of the Imm out argument from SelectSMRDOffset is pretty
confusing. Stop trying to reject CI immediates in the case where the
offset field can be used. It's not an illegal way to encode the
immediate, so just prefer the better encoding pattern with
AddedComplexity.
We probably don't even really need the different opcodes for the
different offset types anymore, but that will be more work to cleanup.
The SMRD non-buffer load patterns could also use a cleanup to be done
separately.
Matt Arsenault