Summary:
Instead of writing boolean values temporarily into 32-bit VGPRs
if they are involved in PHIs or are observed from outside a loop,
we use bitwise masking operations to combine lane masks in a way
that is consistent with wave control flow.
Move SIFixSGPRCopies to before this pass, since that pass
incorrectly attempts to move SGPR phis to VGPRs.
This should recover most of the code quality that was lost with
the bug fix in "AMDGPU: Remove PHI loop condition optimization".
There are still some relevant cases where code quality could be
improved, in particular:
- We often introduce redundant masks with EXEC. Ideally, we'd
have a generic computeKnownBits-like analysis to determine
whether masks are already masked by EXEC, so we can avoid this
masking both here and when lowering uniform control flow.
- The criterion we use to determine whether a def is observed
from outside a loop is conservative: it doesn't check whether
(loop) branch conditions are uniform.
Change-Id: Ibabdb373a7510e426b90deef00f5e16c5d56e64b
Reviewers: arsenm, rampitec, tpr
Subscribers: kzhuravl, jvesely, wdng, mgorny, yaxunl, dstuttard, t-tye, eraman, llvm-commits
Differential Revision: https://reviews.llvm.org/D53496
llvm-svn: 345719
Summary:
The optimization to early break out of loops if all threads are dead was
never fully implemented.
But the PHI node analyzing is actually causing a number of problems, so
remove all the extra code for it.
(This does actually regress code quality in a few places because it
ends up relying more heavily on phi's of i1, which we don't do a
great job with. However, since it fixes real bugs in the wild, we
should take this change. I have some prototype changes to improve
i1 lowering in general -- not just for control flow -- which should
help recover the code quality, I just need to make those changes
fit for general consumption. -- Nicolai)
Change-Id: I6fc6c6c8961857ac6009fcfb9f7e5e48dc23fbb1
Patch-by: Christian König <christian.koenig@amd.com>
Reviewers: arsenm, rampitec, tpr
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D53359
llvm-svn: 345718
Before this patch, class PredicateExpander only knew how to expand simple
predicates that performed checks on instruction operands.
In particular, the new scheduling predicate syntax was not rich enough to
express checks like this one:
Foo(MI->getOperand(0).getImm()) == ExpectedVal;
Here, the immediate operand value at index zero is passed in input to function
Foo, and ExpectedVal is compared against the value returned by function Foo.
While this predicate pattern doesn't show up in any X86 model, it shows up in
other upstream targets. So, being able to support those predicates is
fundamental if we want to be able to modernize all the scheduling models
upstream.
With this patch, we allow users to specify if a register/immediate operand value
needs to be passed in input to a function as part of the predicate check. Now,
register/immediate operand checks all derive from base class CheckOperandBase.
This patch also changes where TIIPredicate definitions are expanded by the
instructon info emitter. Before, definitions were expanded in class
XXXGenInstrInfo (where XXX is a target name).
With the introduction of this new syntax, we may want to have TIIPredicates
expanded directly in XXXInstrInfo. That is because functions used by the new
operand predicates may only exist in the derived class (i.e. XXXInstrInfo).
This patch is a non functional change for the existing scheduling models.
In future, we will be able to use this richer syntax to better describe complex
scheduling predicates, and expose them to llvm-mca.
Differential Revision: https://reviews.llvm.org/D53880
llvm-svn: 345714
Our a16 support was only enabled for sample/gather and buffer
load/store, but not for image load/store operations (which take an i16
as the pixel index rather than a half).
Fix our isel lowering and add test cases to prove it out.
Differential Revision: https://reviews.llvm.org/D53750
llvm-svn: 345710
For some unclear reason rewriteLoopExitValues considers recalculation
after the loop profitable if it has some "soft uses" outside the loop (i.e. any
use other than call and return), even if we have proved that it has a user inside
the loop which we think will not be optimized away.
There is no existing unit test that would explain this. This patch provides an
example when rematerialisation of exit value is not profitable but it passes
this check due to presence of a "soft use" outside the loop.
It makes no sense to recalculate value on exit if we are going to compute it
due to some irremovable within the loop. This patch disallows applying this
transform in the described situation.
Differential Revision: https://reviews.llvm.org/D51581
Reviewed By: etherzhhb
llvm-svn: 345708
optsize using masked wide loads
Under Opt for Size, the vectorizer does not vectorize interleave-groups that
have gaps at the end of the group (such as a loop that reads only the even
elements: a[2*i]) because that implies that we'll require a scalar epilogue
(which is not allowed under Opt for Size). This patch extends the support for
masked-interleave-groups (introduced by D53011 for conditional accesses) to
also cover the case of gaps in a group of loads; Targets that enable the
masked-interleave-group feature don't have to invalidate interleave-groups of
loads with gaps; they could now use masked wide-loads and shuffles (if that's
what the cost model selects).
Reviewers: Ayal, hsaito, dcaballe, fhahn
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D53668
llvm-svn: 345705
Turns out it's not always possible to figure out whether an asm()
statement argument points to a valid memory region.
One example would be per-CPU objects in the Linux kernel, for which the
addresses are calculated using the FS register and a small offset in the
.data..percpu section.
To avoid pulling all sorts of checks into the instrumentation, we replace
actual checking/unpoisoning code with calls to
msan_instrument_asm_load(ptr, size) and
msan_instrument_asm_store(ptr, size) functions in the runtime.
This patch doesn't implement the runtime hooks in compiler-rt, as there's
been no demand in assembly instrumentation for userspace apps so far.
llvm-svn: 345702
Emit pseudo instructions indicating unwind codes corresponding to each
instruction inside the prologue/epilogue. These are used by the MCLayer to
populate the .xdata section.
Differential Revision: https://reviews.llvm.org/D50288
llvm-svn: 345701
The debug-use flag must be set exactly for uses on DBG_VALUEs. This is
so obvious that it can be trivially inferred while parsing. This will
reduce noise when printing while omitting an information that has little
value to the user.
The parser will keep recognizing the flag for compatibility with old
`.mir` files.
Differential Revision: https://reviews.llvm.org/D53903
llvm-svn: 345671
InstCombine features an optimization that essentially replaces:
if (a)
free(a)
into:
free(a)
Right now, this optimization is gated by the minsize attribute and therefore
we only perform it if we can prove that we are going to be able to eliminate
the branch and the destination block.
However when casts are involved the optimization would fail to apply, because
the optimization was not smart enough to realize that it is possible to also
move the casts away from the destination block and that is harmless to the
performance since they are just noops.
E.g.,
foo(int *a)
if (a)
free((char*)a)
Wouldn't be optimized by instcombine, because
- We would refuse to hoist the `bitcast i32* %a to i8` in the source block
- We would fail to see that `bitcast i32* %a to i8` and %a are the same value.
This patch fixes both these problems:
- It teaches the pattern matching of the comparison how to look
through casts.
- It checks that whether the additional instruction in the destination block
can be hoisted and are harmless performance-wise.
- It hoists all the code of the destination block in the source block.
Differential Revision: D53356
llvm-svn: 345644
Summary:
Thunk functions in Windows are varag functions that call a musttail function
to pass the arguments after the fixup is done. We need to make sure that we
forward the arguments from the caller vararg to the callee vararg function.
This is the same mechanism that is used for Windows on X86.
Reviewers: ssijaric, eli.friedman, TomTan, mgrang, mstorsjo, rnk, compnerd, efriedma
Reviewed By: efriedma
Subscribers: efriedma, kristof.beyls, chrib, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D53843
llvm-svn: 345641
Summary:
Normalize the offset for endianess before checking
if the store cover the load in ForwardStoreValueToDirectLoad.
Without this we missed out on some optimizations for big
endian targets. If for example having a 4 bytes store followed
by a 1 byte load, loading the least significant byte from the
store, the STCoversLD check would fail (see @test4 in
test/CodeGen/AArch64/load-store-forwarding.ll).
This patch also fixes a problem seen in an out-of-tree target.
The target has i40 as a legal type, it is big endian,
and the StoreSize for i40 is 48 bits. So when normalizing
the offset for endianess we need to take the StoreSize into
account (assuming that padding added when storing into
a larger StoreSize always is added at the most significant
end).
Reviewers: niravd
Reviewed By: niravd
Subscribers: javed.absar, kristof.beyls, llvm-commits, uabelho
Differential Revision: https://reviews.llvm.org/D53776
llvm-svn: 345636
Prevents the post-RA scheduler from modifying the prologue sequences
emitting by frame lowering. This is roughly similar to what we do for
other targets: TargetInstrInfo::isSchedulingBoundary checks
isPosition(), which checks for CFI_INSTRUCTION.
isSEHInstruction is taken from D50288; it'll land with whatever patch
lands first.
Differential Revision: https://reviews.llvm.org/D53851
llvm-svn: 345634
Re-apply r345315 with testcase fixes.
Include all of the store's source vector operands when creating the
MachineMemOperand. Previously, we were missing the first operand,
making the store size seem smaller than it really is.
Differential Revision: https://reviews.llvm.org/D52816
llvm-svn: 345631
The CONCAT_VECTORS case was using the original mask element count to determine how to adjust the broadcast index. But if we looked through a bitcast the original mask size doesn't tell us anything about the concat_vectors.
This patch switchs to using the concat_vectors input element count directly instead.
Differential Revision: https://reviews.llvm.org/D53823
llvm-svn: 345626
The LRV and STRV nodes carry an extra operand to indicate the
type of the memory access. This is redundant, since the nodes
are actually of class MemIntrinsicNode and therefore hold that
same information already as MemoryVT.
NFC intended.
llvm-svn: 345618
Correct costings of SK_ExtractSubvector requires the SubTy argument to indicate the type/size of the extracted subvector.
Unlike the rest of the shuffle kinds this means that the main Ty argument represents the source vector type not the destination!
I've done my best to fix a number of vectorizer uses:
SLP - the reduction epilogue costs should be using a SK_PermuteSingleSrc shuffle as these all occur at the hardware vector width - we're not extracting (illegal) subvector types. This is causing the cost model diffs as SK_ExtractSubvector costs are poorly handled and tend to just return 1 at the moment.
LV - I'm not clear on what the SK_ExtractSubvector should represents for recurrences - I've used a <1 x ?> subvector extraction as that seems to match the VF delta.
Differential Revision: https://reviews.llvm.org/D53573
llvm-svn: 345617
shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC'
The motivating case is at least a couple of steps away: I noticed that
SLPVectorizer does not analyze shuffles as well as sequences of
insert/extract in PR34724:
https://bugs.llvm.org/show_bug.cgi?id=34724
...so SLP may fail to vectorize when source code has shuffles to start
with or instcombine has converted insert/extract to shuffles.
Independent of that, an insertelement is always a simpler op for IR
analysis vs. a shuffle, so we should transform to insert when possible.
I don't think there's any codegen concern here - if a target can't insert
a scalar directly to some fixed element in a vector (x86?), then this
should get expanded to the insert+shuffle that we started with.
Differential Revision: https://reviews.llvm.org/D53507
llvm-svn: 345607
The SchedModel allows the addition of ReadAdvances to express that certain
operands of the instructions are needed at a later point than the others.
RegAlloc may add pseudo operands that are not part of the instruction
descriptor, and therefore cannot have any read advance entries. This meant
that in some cases the desired read advance was nullified by such a pseudo
operand, which still had the original latency.
This patch fixes this by making sure that such pseudo operands get a zero
latency during DAG construction.
Review: Matthias Braun, Ulrich Weigand.
https://reviews.llvm.org/D49671
llvm-svn: 345606
This commit is a combination of two patches:
* "Fix in getScalarizationOverhead()"
If target returns false in TTI.prefersVectorizedAddressing(), it means the
address registers will not need to be extracted. Therefore, there should
be no operands scalarization overhead for a load instruction.
* "Don't pass the instruction pointer from getMemInstScalarizationCost."
Since VF is always > 1, this is a cost query for an instruction in the
vectorized loop and it should not be evaluated within the scalar
context of the instruction.
Review: Ulrich Weigand, Hal Finkel
https://reviews.llvm.org/D52351https://reviews.llvm.org/D52417
llvm-svn: 345603
Narrowing vector binops came up in the demanded bits discussion in D52912.
I don't think we're going to be able to do this transform in IR as a canonicalization
because of the risk of creating unsupported widths for vector ops, but we already have
a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is
currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression
test diffs).
This is artificially limited to not look through bitcasts because there are so many
test diffs already, but that's marked with a TODO and is a small follow-up.
Differential Revision: https://reviews.llvm.org/D53784
llvm-svn: 345602
Sub, SDiv and UDiv are not commutative, so only the RHS operand can fold a
load. This patch adds a check for this.
Review: Ulrich Weigand
https://reviews.llvm.org/D53791
llvm-svn: 345596
This fixes an assertion when constant folding a GEP when the part of the offset
was in i32 (IndexSize, as per DataLayout) and part in the i64 (PointerSize) in
the newly created test case.
Differential Revision: https://reviews.llvm.org/D52609
llvm-svn: 345585
Summary:
The final pattern.
There is no test changes:
* We are looking for the pattern with one-use of it's mask,
* If the mask is one-use, D48768 will unfold it into pattern d.
* Thus, the tests have extra-use on the mask.
* Thus, only the BMI2 BZHI can be tested, and it already worked.
* So there is no BMI1 test coverage, we just assume it works since it uses the same codepath.
Reviewers: craig.topper, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53575
llvm-svn: 345584
Similar to FoldCONCAT_VECTORS, this patch adds FoldBUILD_VECTOR to simplify cases that can avoid the creation of the BUILD_VECTOR - if all the operands are UNDEF or if the BUILD_VECTOR simplifies to a copy.
This exposed an assumption in some AMDGPU code that getBuildVector was guaranteed to be a BUILD_VECTOR node that I've tried to handle.
Differential Revision: https://reviews.llvm.org/D53760
llvm-svn: 345578
Summary: Previously if we had a bitcast vector output type that needs promotion and a vector input type that needs widening we would just do a stack store and load to handle the conversion. We can do a little better if we can widen the bitcast to a legal vector type the same size as the widened input type. Then we can do the bitcast between this widened type and the widened input type. Afterwards we can extract_subvector back to the original output and any_extend that. Type legalization will then circle back and handle promotion of the extract_subvector and the any_extend will just be removed. This will avoid going through the stack and allows us to remove a custom version of this legalization from X86.
Reviewers: efriedma, RKSimon
Reviewed By: efriedma
Subscribers: javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D53229
llvm-svn: 345567
Use SelectionDAG::EVTToAPFloatSemantics. Make the LogicVT calculation in LowerFABSorFNEG similar to LowerFCOPYSIGN. Use APInt::getSignedMaxValue instead of ~APInt::getSignMask.
llvm-svn: 345565
David Bolvansky