Currently undef is used as a don’t-care vector when constructing a vector using a series of insertelement.
However, this is problematic because undef isn’t undefined enough.
Especially, a sequence of insertelement can be optimized to shufflevector, but using undef as its placeholder makes shufflevector a poison-blocking instruction because undef cannot be optimized to poison.
This makes a few straightforward optimizations incorrect, such as:
```
; https://bugs.llvm.org/show_bug.cgi?id=44185
define <4 x float> @insert_not_undef_shuffle_translate_commute(float %x, <4 x float> %y, <4 x float> %q) {
%xv = insertelement <4 x float> %q, float %x, i32 2
%r = shufflevector <4 x float> %y, <4 x float> %xv, <4 x i32> { 0, 6, 2, undef }
ret <4 x float> %r ; %r[3] is undef
}
=>
define <4 x float> @insert_not_undef_shuffle_translate_commute(float %x, <4 x float> %y, <4 x float> %q) {
%r = insertelement <4 x float> %y, float %x, i32 1
ret <4 x float> %r ; %r[3] = %y[3], incorrect if %y[3] = poison
}
Transformation doesn't verify!
ERROR: Target is more poisonous than source
```
I’d like to suggest
1. Using poison as insertelement’s placeholder value (IRBuilder::CreateVectorSplat should be patched too)
2. Updating shufflevector’s semantics to return poison element if mask is undef
Note that poison is currently lowered into UNDEF in SelDag, so codegen part is okay.
m_Undef() matches PoisonValue as well, so existing optimizations will still fire.
The only concern is hidden miscompilations that will go incorrect when poison constant is given.
A conservative way is copying all tests having `insertelement undef` & replacing it with `insertelement poison` & run Alive2 on it, but it will create many tests and people won’t like it. :(
Instead, I’ll simply locally maintain the tests and run Alive2.
If there is any bug found, I’ll report it.
Relevant links: https://bugs.llvm.org/show_bug.cgi?id=43958 , http://lists.llvm.org/pipermail/llvm-dev/2019-November/137242.html
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93586
I noticed an add example like the one from D91343, so here's a similar patch.
The logic is based on existing code for the single-use demanded bits fold.
But I only matched a constant instead of using compute known bits on the
operands because that was the motivating patterni that I noticed.
I think this will allow removing a special-case (but incomplete) dedicated
fold within visitAnd(), but I need to untangle the existing code to be sure.
https://rise4fun.com/Alive/V6fP
Name: add with low mask
Pre: (C1 & (-1 u>> countLeadingZeros(C2))) == 0
%a = add i8 %x, C1
%r = and i8 %a, C2
=>
%r = and i8 %x, C2
Differential Revision: https://reviews.llvm.org/D91415
We have a frequent pattern where we're merging two KnownBits to get the common/shared bits, and I just fell for the gotcha where I tried to use the & operator to merge them........
The 1st attempt (rG557b890) was reverted because it caused miscompiles.
That bug is avoided here by changing the order of folds and as verified
in the new tests.
Original commit message:
InstCombine currently has odd rules for folding insert-extract chains to shuffles,
so we miss collapsing seemingly simple cases as shown in the tests here.
But poison makes this not quite as easy as we might have guessed. Alive2 tests to
show the subtle difference (similar to the regression tests):
https://alive2.llvm.org/ce/z/hp4hv3 (this is ok)
https://alive2.llvm.org/ce/z/ehEWaN (poison leakage)
SLP tends to create these patterns (as shown in the SLP tests), and this could
help with solving PR16739.
Differential Revision: https://reviews.llvm.org/D86460
In getCastInstrCost when the instruction is a truncate we were relying
upon the implicit TypeSize -> uint64_t cast when asking if a given type
has the same size as a legal integer. I've changed the code to only
ask the question if the type is fixed length.
I have also changed InstCombinerImpl::SimplifyDemandedUseBits to bail
out for now if the type is a scalable vector.
I've added the following new tests:
Analysis/CostModel/AArch64/sve-trunc.ll
Transforms/InstCombine/AArch64/sve-trunc.ll
for both of these fixes.
Differential revision: https://reviews.llvm.org/D86432
InstCombine currently has odd rules for folding insert-extract chains to shuffles,
so we miss collapsing seemingly simple cases as shown in the tests here.
But poison makes this not quite as easy as we might have guessed. Alive2 tests to
show the subtle difference (similar to the regression tests):
https://alive2.llvm.org/ce/z/hp4hv3 (this is ok)
https://alive2.llvm.org/ce/z/ehEWaN (poison leakage)
SLP tends to create these patterns (as shown in the SLP tests), and this could
help with solving PR16739.
Differential Revision: https://reviews.llvm.org/D86460
There's a potential motivating case to increase this limit in PR47191:
http://bugs.llvm.org/PR47191
But first we should make it less hacky. The limit in InstCombine is directly tied
to this value because an increase there can cause asserts in the underlying value
tracking calls if not changed together. The usage in VectorUtils is independent,
but the comment suggests that we should use the same value unless there's a known
reason to diverge. There are similar limits in codegen analysis, but I think we
should leave those independent in case we intentionally want the optimization
power/cost to be different there.
Differential Revision: https://reviews.llvm.org/D86113
This is a retry of rL300977 which was reverted because of infinite loops.
We have fixed all of the known places where that would happen, but there's
still a chance that this patch will cause infinite loops.
This matches the demanded bits behavior in the DAG and should fix:
https://bugs.llvm.org/show_bug.cgi?id=32706
Differential Revision: https://reviews.llvm.org/D32255
For a long time, the InstCombine pass handled target specific
intrinsics. Having target specific code in general passes was noted as
an area for improvement for a long time.
D81728 moves most target specific code out of the InstCombine pass.
Applying the target specific combinations in an extra pass would
probably result in inferior optimizations compared to the current
fixed-point iteration, therefore the InstCombine pass resorts to newly
introduced functions in the TargetTransformInfo when it encounters
unknown intrinsics.
The patch should not have any effect on generated code (under the
assumption that code never uses intrinsics from a foreign target).
This introduces three new functions:
TargetTransformInfo::instCombineIntrinsic
TargetTransformInfo::simplifyDemandedUseBitsIntrinsic
TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic
A few target specific parts are left in the InstCombine folder, where
it makes sense to share code. The largest left-over part in
InstCombineCalls.cpp is the code shared between arm and aarch64.
This allows to move about 3000 lines out from InstCombine to the targets.
Differential Revision: https://reviews.llvm.org/D81728
v3i16 and v3f16 currently cannot be legalized and lowered so they should
not be emitted by inst combining.
Moved the check down to still allow extracting 1 or 2 elements via the dmask.
Fixes image intrinsics being combined to return v3x16.
Differential Revision: https://reviews.llvm.org/D84223
- Now all SalvageDebugInfo() calls will mark undef if the salvage
attempt fails.
Reviewed by: vsk, Orlando
Differential Revision: https://reviews.llvm.org/D78369
Summary:
This patch fixes the following issues in visitInsertElementInst:
1. Bail out for scalable type when analysis requires fixed size number of vector elements.
2. Use cast<FixedVectorType> to get vector number of elements. This ensure assertion
on scalable vector type.
3. For scalable type, avoid folding a chain of insertelement into splat:
insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ...
->
shufflevector(insertelt(X, %k, 0), undef, zero)
The length of scalable vector is unknown at compile-time, therefore we don't know if
given insertelement sequence is valid for splat.
Reviewers: sdesmalen, efriedma, spatel, nikic
Reviewed By: sdesmalen, efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78895
Summary:
There are at least three clients for KnownBits calculations:
ValueTracking, SelectionDAG and GlobalISel. To reduce duplication the
common logic should be moved out of these clients and into KnownBits
itself.
This patch does this for AND, OR and XOR calculations by implementing
and using appropriate operator overloads KnownBits::operator& etc.
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74060
Summary:
Remove usages of asserting vector getters in Type in preparation for the
VectorType refactor. The existence of these functions complicates the
refactor while adding little value.
Reviewers: sdesmalen, rriddle, efriedma
Reviewed By: sdesmalen
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77263
Instead, represent the mask as out-of-line data in the instruction. This
should be more efficient in the places that currently use
getShuffleVector(), and paves the way for further changes to add new
shuffles for scalable vectors.
This doesn't change the syntax in textual IR. And I don't currently plan
to change the bitcode encoding in this patch, although we'll probably
need to do something once we extend shufflevector for scalable types.
I expect that once this is finished, we can then replace the raw "mask"
with something more appropriate for scalable vectors. Not sure exactly
what this looks like at the moment, but there are a few different ways
we could handle it. Maybe we could try to describe specific shuffles.
Or maybe we could define it in terms of a function to convert a fixed-length
array into an appropriate scalable vector, using a "step", or something
like that.
Differential Revision: https://reviews.llvm.org/D72467
Optimize the common case of splat vector constant. For large vector
going through all elements is expensive. For splatr/broadcast cases we
can skip going through all elements.
Differential Revision: https://reviews.llvm.org/D76664
Ideally SimplifyDemanded should compute the same known bits as
computeKnownBits(). This patch addresses one discrepancy, where
ValueTracking is more powerful: If we have a shl nsw shift, we
know that the sign bit of the input and output must be the same.
If this results in a conflict, the result is poison.
This is implemented in
2c4ca6832f/lib/Analysis/ValueTracking.cpp (L1175-L1179)
and
2c4ca6832f/lib/Analysis/ValueTracking.cpp (L904-L908).
This implements the same basic logic in SimplifyDemanded. It's
slightly stronger, because I return undef instead of zero for the
poison case (which is not an option inside ValueTracking).
As mentioned in https://reviews.llvm.org/D75801#inline-698484,
we could detect poison in more cases, this just establishes parity
with the existing logic.
Differential Revision: https://reviews.llvm.org/D76489
Fixes a regression from D75801. SimplifyDemandedUseBits() is also
supposed to compute the known bits (of the demanded subset) of the
instruction. For unknown instructions it does so by directly calling
computeKnownBits(). For known instructions it will compute known
bits itself. However, for instructions where only some cases are
handled directly (e.g. a constant shift amount) the known bits
invocation for the unhandled case is sometimes missing. This patch
adds the missing calls and thus removes the main discrepancy with
ExpensiveCombines mode.
Differential Revision: https://reviews.llvm.org/D75804
This fixes a small mistake from D72944: The worklist add should
happen before assigning the new operand, not after.
In case an actual replacement happens, the old operand needs to
be added for DCE. If no actual replacement happens, then old/new
are the same, so it doesn't matter.
This drops one iteration from the annotated test case.
When simplifying demanded bits, we currently only report the
instruction on which SimplifyDemandedBits was called as changed.
However, this is a recursive call, and the actually modified
instruction will usually be further up the chain. Additionally,
all the intermediate instructions should also be revisited,
as additional combines may be possible after the demanded bits
simplification. We fix this by explicitly adding them back to the
worklist.
Differential Revision: https://reviews.llvm.org/D72944
Summary:
This was a very odd API, where you had to pass a flag into a zext
function to say whether the extended bits really were zero or not. All
callers passed in a literal true or false.
I think it's much clearer to make the function name reflect the
operation being performed on the value we're tracking (rather than on
the KnownBits Zero and One fields), so zext means the value is being
zero extended and new function anyext means the value is being extended
with unknown bits.
NFC.
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74482
This renames Worklist.AddDeferred() to Worklist.add() and
Worklist.Add() to Worklist.push(). The intention here is that
Worklist.add() should be the go-to method for explicit worklist
management, while the raw Worklist.push() is mostly for
InstCombine internals. I will then migrate uses of Worklist.push()
to Worklist.add() in followup changes.
As suggested by spatel on D73411 I'm also changing the remaining
method names to lowercase first character, in line with current
coding standards.
Differential Revision: https://reviews.llvm.org/D73745
Summary:
Add trimming of unused components of s_buffer_load.
For s_buffer_load and unformatted buffer_load also trim unused
components at the beginning of vector and update offset accordingly.
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71785
In certain situations after inlining and simplification we end up with
code that is _almost_ a min/max pattern, but contains constants that
have been demand-bit optimised to the wrong values, ending up with code
like:
%1 = icmp slt i32 %shr, -128
%2 = select i1 %1, i32 128, i32 %shr
%.inv = icmp sgt i32 %shr, 127
%spec.select.i = select i1 %.inv, i32 127, i32 %2
%conv7 = trunc i32 %spec.select.i to i8
This should be turned into a min/max pattern, but the -128 in the first
select was instead transformed into 128, as only the bottom byte was
ever demanded.
To fix this, I've put in further canonicalisation for the immediates of
selects, preferring to use the same value as the icmp if available.
Differential Revision: https://reviews.llvm.org/D71516
Summary:
Add trimming of unused components of s_buffer_load.
Extend trimming of *buffer_load to also include
unused components at the beginning of vectors and update offset.
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70315
This has two main effects:
- Optimizes debug info size by saving 221.86 MB of obj file size in a
Windows optimized+debug build of 'all'. This is 3.03% of 7,332.7MB of
object file size.
- Incremental step towards decoupling target intrinsics.
The enums are still compact, so adding and removing a single
target-specific intrinsic will trigger a rebuild of all of LLVM.
Assigning distinct target id spaces is potential future work.
Part of PR34259
Reviewers: efriedma, echristo, MaskRay
Reviewed By: echristo, MaskRay
Differential Revision: https://reviews.llvm.org/D71320
And simultaneously enhance SimplifyDemandedVectorElts() to rcognize that
pattern. That preserves some of the old optimizations in IR.
Given a shuffle that includes undef elements in an otherwise identity mask like:
define <4 x float> @shuffle(<4 x float> %arg) {
%shuf = shufflevector <4 x float> %arg, <4 x float> undef, <4 x i32> <i32 undef, i32 1, i32 2, i32 3>
ret <4 x float> %shuf
}
We were simplifying that to the input operand.
But as discussed in PR43958:
https://bugs.llvm.org/show_bug.cgi?id=43958
...that means that per-vector-element poison that would be stopped by the shuffle can now
leak to the result.
Also note that we still have (and there are tests for) the same transform with no undef
elements in the mask (a fully-defined identity mask). I don't think there's any
controversy about that case - it's a valid transform under any interpretation of
shufflevector/undef/poison.
Looking at a few of the diffs into codegen, I don't see any difference in final asm. So
depending on your perspective, that's good (no real loss of optimization power) or bad
(poison exists in the DAG, so we only partially fixed the bug).
Differential Revision: https://reviews.llvm.org/D70246
Summary:
Allow for ignoring the check for a single use in SimplifyDemandedVectorElts
to be able to simplify operands if DemandedElts is known to contain
the union of elements used by all users.
It is a responsibility of a caller of SimplifyDemandedVectorElts to
supply correct DemandedElts.
Simplify a series of extractelement instructions if only a subset of
elements is used.
Reviewers: reames, arsenm, majnemer, nhaehnle
Reviewed By: nhaehnle
Subscribers: wdng, jvesely, nhaehnle, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67345
llvm-svn: 375395
Summary:
This is something of a workaround to avoid a crash later on in type
legalizer (WidenVectorResult()).
Also added some f16 tests, including a non-working v3f16 case with
a FIXME.
Reviewers: arsenm, tpr, nhaehnle
Reviewed By: arsenm
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68865
llvm-svn: 374993
Extend the mechanism to overload intrinsic arguments by using either
backward or forward references to the overloadable arguments.
In for example:
def int_something : Intrinsic<[LLVMPointerToElt<0>],
[llvm_anyvector_ty], []>;
LLVMPointerToElt<0> is a forward reference to the overloadable operand
of type 'llvm_anyvector_ty' and would allow intrinsics such as:
declare i32* @llvm.something.v4i32(<4 x i32>);
declare i64* @llvm.something.v2i64(<2 x i64>);
where the result pointer type is deduced from the element type of the
first argument.
If the returned pointer is not a pointer to the element type, LLVM will
give an error:
Intrinsic has incorrect return type!
i64* (<4 x i32>)* @llvm.something.v4i32
Reviewers: RKSimon, arsenm, rnk, greened
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D62995
llvm-svn: 363233
The demanded elts rules introduced for GEPs in https://reviews.llvm.org/rL356293 replaced vector constants with undefs (by design). It turns out that the LangRef disallows such cases when indexing structs. The right fix is probably to relax the langref requirement, and update other passes to expect the result, but for the moment, limit the transform to avoid compiler crashes.
This should fix https://bugs.llvm.org/show_bug.cgi?id=41624.
llvm-svn: 359633
This fixes a miscompile which was introduced in r356510 (https://reviews.llvm.org/D57372).
The problem is that the original patch removed pointer operands where the load results we're demanded, but without considering the legality of the load itself. If the masked.gather had active, but undemanded, lanes, then we could end up creating a load which loaded from an undef address. The result could be a segfault, or, in theory, an arbitrary read from a random memory location into an used register.
llvm-svn: 358299
This helps to avoid the situation where RA spots that only 3 of the
v4f32 result of a load are used, and immediately reallocates the 4th
register for something else, requiring a stall waiting for the load.
Differential Revision: https://reviews.llvm.org/D58906
Change-Id: I947661edfd5715f62361a02b100f14aeeada29aa
llvm-svn: 356768
Teach instcombine to propagate demanded elements through a masked load or masked gather instruction. This is in the broader context of improving vector pointer instcombine under https://reviews.llvm.org/D57140.
Differential Revision: https://reviews.llvm.org/D57372
llvm-svn: 356510
A change of two parts:
1) A generic enhancement for all callers of SDVE to exploit the fact that if all lanes are undef, the result is undef.
2) A GEP specific piece to strengthen/fix the vector index undef element handling, and call into the generic infrastructure when visiting the GEP.
The result is that we replace a vector gep with at least one undef in each lane with a undef. We can also do the same for vector intrinsics. Once the masked.load patch (D57372) has landed, I'll update to include call tests as well.
Differential Revision: https://reviews.llvm.org/D57468
llvm-svn: 356293
Before r355981, this was under LLVM_DEBUG. I don't think the assert is
quite right, but this really should be a verifier check. Instcombine
should not be asserting on this sort of thing.
llvm-svn: 356219
This indicates an intrinsic parameter is required to be a constant,
and should not be replaced with a non-constant value.
Add the attribute to all AMDGPU and generic intrinsics that comments
indicate it should apply to. I scanned other target intrinsics, but I
don't see any obvious comments indicating which arguments are intended
to be only immediates.
This breaks one questionable testcase for the autoupgrade. I'm unclear
on whether the autoupgrade is supposed to really handle declarations
which were never valid. The verifier fails because the attributes now
refer to a parameter past the end of the argument list.
llvm-svn: 355981
Summary:
The description of KnownBits::zext() and
KnownBits::zextOrTrunc() has confusingly been telling
that the operation is equivalent to zero extending the
value we're tracking. That has not been true, instead
the user has been forced to explicitly set the extended
bits as known zero afterwards.
This patch adds a second argument to KnownBits::zext()
and KnownBits::zextOrTrunc() to control if the extended
bits should be considered as known zero or as unknown.
Reviewers: craig.topper, RKSimon
Reviewed By: RKSimon
Subscribers: javed.absar, hiraditya, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D58650
llvm-svn: 355099
Summary:
The fix added in r352904 is not quite correct, or rather misleading:
1. When the texfailctrl (TFC) argument was non-constant, the fix assumed
non-TFE/LWE, which is incorrect.
2. Regardless, this code path cannot even be hit for correct
TFE/LWE-enabled calls, because those return a struct. Added
a test case for those for completeness.
Change-Id: I92d314dbc67a2670f6d7adaab765ef45f56a49cf
Reviewers: hliao, dstuttard, arsenm
Subscribers: kzhuravl, jvesely, wdng, yaxunl, tpr, t-tye, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D57681
llvm-svn: 353097
GEPs can produce either scalar or vector results. If we're extracting only a subset of the vector lanes, simplifying the operands is helpful in eliminating redundant computation, and (eventually) allowing further optimizations
Differential Revision: https://reviews.llvm.org/D57177
llvm-svn: 352440
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
Also revert fix r347876
One of the buildbots was reporting a failure in some relevant tests that I can't
repro or explain at present, so reverting until I can isolate.
llvm-svn: 347911
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
llvm-svn: 347871
Support funnel shifts in InstCombine demanded bits simplification.
If the shift amount is constant, we can determine both the demanded
bits of the operands, as well as the known bits of the result.
If one of the operands has no demanded bits, it will be replaced
by undef and the funnel shift will be simplified into a simple shift
due to the simplifications added in D54778.
Differential Revision: https://reviews.llvm.org/D54869
llvm-svn: 347515
This is the umin alternative to the umax code from rL344237. We use
DeMorgans law on the umax case to bring us to the same thing on umin,
but using countLeadingOnes, not countLeadingZeros.
Differential Revision: https://reviews.llvm.org/D53036
llvm-svn: 344239
Use the demanded bits of umax(A,C) to prove we can just use A so long as the
lowest non-zero bit of DemandMask is higher than the highest non-zero bit of C
Differential Revision: https://reviews.llvm.org/D53033
llvm-svn: 344237
We established the (unfortunately complicated) rules for UB/poison
propagation with vector ops in:
D48893
D48987
D49047
It's clear from the affected tests that we are potentially creating
poison where none existed before the transforms. For add/sub/mul,
the answer is simple: just drop the flags because the extra undef
vector lanes are generally more valuable for analysis and codegen.
llvm-svn: 343819
We're a long way from D50992 and D51553, but this is where we have to start.
We weren't back-propagating undefs into binop constant values for anything but
add/sub/mul/and/or/xor.
This is likely because we have to be careful about not introducing UB/poison
with div/rem/shift. But I suspect we already are getting the poison part wrong
for add/sub/mul (although it may not be possible to expose the bug currently
because we use SimplifyDemandedVectorElts from a limited set of opcodes).
See the discussion/implementation from D48987 and D49047.
This patch just enables functionality for FP ops because those do not have
UB/poison potential.
llvm-svn: 343727
I noticed that we were not back-propagating undef lanes to shuffle masks when we have a
shuffle that reduces the vector width. This is part of investigating/solving PR38691:
https://bugs.llvm.org/show_bug.cgi?id=38691
The DAG equivalent was proposed with:
D51696
Differential Revision: https://reviews.llvm.org/D51433
llvm-svn: 341981
I'm preparing to add the same functionality both here and to the DAG
version of this code in D51696 / D51433, so try to make those cases
as similar as possible to avoid bugs.
llvm-svn: 341545
This converts them to what clang is now using for codegen. Unfortunately, there seem to be a few kinks to work out still. I'll try to address with follow up patches.
llvm-svn: 336871
Summary:
This also removes the need for atomic pseudo instructions, since
we select the correct encoding directly in SITargetLowering::lowerImage
for dimension-aware image intrinsics.
Mesa uses dimension-aware image intrinsics since
commit a9a7993441.
Change-Id: I7473d20009476a4ed6d919cae4e6dca9ff42e77a
Reviewers: arsenm, rampitec, mareko, tpr, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48167
llvm-svn: 335231
Summary:
Use the expanded features of the TableGen generic tables to avoid manually
adding the combinatorially exploded set of intrinsics. The
getAMDGPUImageDimIntrinsic lookup function is early-out,
i.e. non-AMDGPU intrinsics will never look at the underlying table.
Use a generic approach for getting the new intrinsic overload to keep the
code simple, and make the image dmask handling more generic:
- handle non-sampler image loads
- handle the case where the set of demanded elements is not a prefix
There is some overlap between this code and an optimization that happens
in the backend during code generation. They currently complement each other:
- only the codegen optimization can generate vec3 loads
- only the InstCombine optimization can handle D16
The InstCombine optimization also likely covers more cases since the
codegen optimization is fairly ad-hoc. Ideally, we'll remove the optimization
in codegen once the infrastructure for vec3 is in place (which will probably
take a long time).
Modify the test cases to use dimension-aware intrinsics. This makes it
easier to see that the test coverage for the new intrinsics is equivalent,
and the old style intrinsics will be removed in a follow-up commit anyway.
Change-Id: I4b91ea661413d13004956fe4ef7d13d41b8ce3ad
Reviewers: arsenm, rampitec, majnemer
Subscribers: kzhuravl, wdng, mgorny, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48165
llvm-svn: 335230
Put in a conservatively correct estimate for now. Avoids miscompiling
clang in FDO mode. This is really tricky to trigger in reality as
basically all interesting cases will be folded away by computeKnownBits
earlier, I was unable to find a reasonably small test case.
llvm-svn: 331975
This is safe as long as the udiv is not exact. The pattern is not common in
C++ code, but comes up all the time in code generated by XLA's GPU backend.
Differential Revision: https://reviews.llvm.org/D46647
llvm-svn: 331933
This completes the work started in r329604 and r329605 when we changed clang to no longer use the intrinsics.
We lost some InstCombine SimplifyDemandedBit optimizations through this change as we aren't able to fold 'and', bitcast, shuffle very well.
llvm-svn: 329990
I was comparing the demanded-bits implementations between InstCombine
and TargetLowering as part of investigating questions in D42088 and
noticed that this was wrong in IR. We were losing all of the prior
known bits when we got back to the 'zext'.
llvm-svn: 322662
Recurse instead of returning on the first found optimization. Also, return early in the caller
instead of continuing because that allows another round of simplification before we might
potentially lose undef information from a shuffle mask by eliminating the shuffle.
As noted in the review, we could probably do better and be more efficient by moving all of
demanded elements into a separate pass, but this is yet another quick fix to instcombine.
Differential Revision: https://reviews.llvm.org/D37236
llvm-svn: 312248
Just create an all 1s demanded mask and continue recursing like normal. The recursive calls should be able to handle an all 1s mask and do the right thing.
The only time we should care about knowing whether the upper bit was demanded is when we need to know if we should clear the NSW/NUW flags.
Now that we have a consistent path through the code for all cases, use KnownBits::computeForAddSub to compute the known bits at the end since we already have the LHS and RHS.
My larger goal here is to move the code that turns add into xor if only 1 bit is demanded and no bits below it are non-zero from InstCombiner::OptAndOp to here. This will allow it to be more general instead of just looking for 'add' and 'and' with constant RHS.
Differential Revision: https://reviews.llvm.org/D36486
llvm-svn: 311789
There are cases where AShr have better chance to be optimized than LShr, especially when the demanded bits are not known to be Zero, and also known to be similar to the sign bit.
Differential Revision: https://reviews.llvm.org/D36936
llvm-svn: 311773
Summary:
As far as I can tell the earlier call getLimitedValue will guaranteed ShiftAmt is saturated to BitWidth-1 preventing it from ever being equal or greater than BitWidth.
At one point in the past the getLimitedValue call was only passed BitWidth not BitWidth - 1. This would have allowed the equality case to get here. And in fact this check was initially added as just BitWidth == ShiftAmt, but was changed shortly after to include > which should have never been possible.
Reviewers: spatel, majnemer, davide
Reviewed By: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D36123
llvm-svn: 309690
Previously the InstCombiner class contained a pointer to an IR builder that had been passed to the constructor. Sometimes this would be passed to helper functions as either a pointer or the pointer would be dereferenced to be passed by reference.
This patch makes it a reference everywhere including the InstCombiner class itself so there is more inconsistency. This a large, but mechanical patch. I've done very minimal formatting changes on it despite what clang-format wanted to do.
llvm-svn: 307451
Going through the Constant methods requires redetermining that the Constant is a ConstantInt and then calling isZero/isOne/isMinusOne.
llvm-svn: 307292
While there avoid resizing the DemandedMask twice. Make a copy into a separate variable instead. This potentially removes an allocation on large bit widths.
With the use of the zextOrTrunc methods on APInt and KnownBits these can be made almost source identical. The only difference is the zero of the upper bits for ZExt. This is similar to how its done in computeKnownBits in ValueTracking.
llvm-svn: 303791
The current code created a NewBits mask and used it as a mask several times. One of them just before a call to trunc making it unnecessary. A call to getActiveBits can get us the same information for the case. We also ORed with this mask later when we should have just sign extended the known bits.
We also called trunc on the guaranteed to be zero KnownZeros/Ones masks entering this code. Creating appropriately sized temporary APInts is probably better.
Differential Revision: https://reviews.llvm.org/D32098
llvm-svn: 303779
Sanjay Patel