Summary: This avoid the need to duplicate all pattern and actually end up exposing some opportunity to optimize existing pattern that did not exists in both directions on an existing test case.
Reviewers: mkuper, spatel, bkramer, RKSimon, zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29541
llvm-svn: 294125
Recommiting after fixing X86 inc/dec chain bug.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 293893
Summary:
The affected transforms all implicitly use associativity of addition,
for which we usually require unsafe math to be enabled.
The "Aggressive" flag is only meant to convey information about the
performance of the fused ops relative to a fmul+fadd sequence.
Fixes Bug 31626.
Reviewers: spatel, hfinkel, mehdi_amini, arsenm, tstellarAMD
Subscribers: jholewinski, nemanjai, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D28675
llvm-svn: 293635
The type system already requires that the number of vector elements must fit in 32-bits so an index should as well. Even if the type of the index were larger all we care about is that the constant index can fit in 64-bits so that we can call getZExtValue.
llvm-svn: 293413
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 293184
The previous patch (https://reviews.llvm.org/rL289538) got reverted because of a bug. Chandler also requested some changes to the algorithm.
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161212/413479.html
This is an updated patch. The key difference is that collectBitProviders (renamed to calculateByteProvider) now collects the origin of one byte, not the whole value. It simplifies the implementation and allows to stop the traversal earlier if we know that the result won't be used.
From the original commit:
Match a pattern where a wide type scalar value is loaded by several narrow loads and combined by shifts and ors. Fold it into a single load or a load and a bswap if the targets supports it.
Assuming little endian target:
i8 *a = ...
i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24)
=>
i32 val = *((i32)a)
i8 *a = ...
i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3]
=>
i32 val = BSWAP(*((i32)a))
This optimization was discussed on llvm-dev some time ago in "Load combine pass" thread. We came to the conclusion that we want to do this transformation late in the pipeline because in presence of atomic loads load widening is irreversible transformation and it might hinder other optimizations.
Eventually we'd like to support folding patterns like this where the offset has a variable and a constant part:
i32 val = a[i] | (a[i + 1] << 8) | (a[i + 2] << 16) | (a[i + 3] << 24)
Matching the pattern above is easier at SelectionDAG level since address reassociation has already happened and the fact that the loads are adjacent is clear. Understanding that these loads are adjacent at IR level would have involved looking through geps/zexts/adds while looking at the addresses.
The general scheme is to match OR expressions by recursively calculating the origin of individual bytes which constitute the resulting OR value. If all the OR bytes come from memory verify that they are adjacent and match with little or big endian encoding of a wider value. If so and the load of the wider type (and bswap if needed) is allowed by the target generate a load and a bswap if needed.
Reviewed By: RKSimon, filcab, chandlerc
Differential Revision: https://reviews.llvm.org/D27861
llvm-svn: 293036
clang already emits this with -cl-no-signed-zeros, but codegen
doesn't do anything with it. Treat it like the other fast math
attributes, and change one place to use it.
llvm-svn: 293024
Even with aggressive fusion enabled, this requires duplicating
the fmul, or increases an fadd to another fma which is not an
improvement.
llvm-svn: 291642
Summary:
Instead of matching:
(a + i) + 1 -> (a + i, undef, 1)
Now it matches:
(a + i) + 1 -> (a, i, 1)
Reviewers: rengolin
Differential Revision: http://reviews.llvm.org/D26367
From: Evgeny Stupachenko <evstupac@gmail.com>
llvm-svn: 291012
1.Fix pessimized case in FIXME.
2.Add tests for it.
3.The canonicalisation on shifts results in different sequence for
tests of machine-licm.Correct some check lines.
Differential Revision: https://reviews.llvm.org/D27916
llvm-svn: 290410
This is for splitMergedValStore in DAG Combine to share the target query interface
with similar logic in CodeGenPrepare.
Differential Revision: https://reviews.llvm.org/D24707
llvm-svn: 290363
idiom.
r289538: Match load by bytes idiom and fold it into a single load
r289540: Fix a buildbot failure introduced by r289538
r289545: Use more detailed assertion messages in the code ...
r289646: Add a couple of assertions to the load combine code ...
This DAG combine has a bad crash in it that is quite hard to trigger
sadly -- it relies on sneaking code with UB through the SDAG build and
into this particular combine. I've responded to the original commit with
a test case that reproduces it.
However, the code also has other problems that will require substantial
changes to address and so I'm going ahead and reverting it for now. This
should unblock us and perhaps others that are hitting the crash in the
wild and will let a fresh patch with updated approach come in cleanly
afterward.
Sorry for any trouble or disruption!
llvm-svn: 289916
The original motivation for this patch comes from wanting to canonicalize
more IR to selects and also canonicalizing min/max.
If we're going to do that, we need more backend fixups to undo select codegen
when simpler ops will do. I chose AArch64 for the tests because that shows the
difference in the simplest way. This should fix:
https://llvm.org/bugs/show_bug.cgi?id=31175
Differential Revision: https://reviews.llvm.org/D27489
llvm-svn: 289738
Retrying after fixing after removing load-store factoring through
token factors in favor of improved token factor operand pruning
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 289659
Generalize sdiv/udiv/srem/urem combines using APInt::isPowerOf2, which only works for const/splat-const values, to call SelectionDAG::isKnownToBeAPowerOfTwo instead which recognises many more cases.
Added a DAGCombiner::BuildLogBase2 helper since PowerOf2 combines often involve taking the log2 of such a value.
Differential Revision: https://reviews.llvm.org/D27714
llvm-svn: 289654
Match a pattern where a wide type scalar value is loaded by several narrow loads and combined by shifts and ors. Fold it into a single load or a load and a bswap if the targets supports it.
Assuming little endian target:
i8 *a = ...
i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24)
=>
i32 val = *((i32)a)
i8 *a = ...
i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3]
=>
i32 val = BSWAP(*((i32)a))
This optimization was discussed on llvm-dev some time ago in "Load combine pass" thread. We came to the conclusion that we want to do this transformation late in the pipeline because in presence of atomic loads load widening is irreversible transformation and it might hinder other optimizations.
Eventually we'd like to support folding patterns like this where the offset has a variable and a constant part:
i32 val = a[i] | (a[i + 1] << 8) | (a[i + 2] << 16) | (a[i + 3] << 24)
Matching the pattern above is easier at SelectionDAG level since address reassociation has already happened and the fact that the loads are adjacent is clear. Understanding that these loads are adjacent at IR level would have involved looking through geps/zexts/adds while looking at the addresses.
The general scheme is to match OR expressions by recursively calculating the origin of individual bits which constitute the resulting OR value. If all the OR bits come from memory verify that they are adjacent and match with little or big endian encoding of a wider value. If so and the load of the wider type (and bswap if needed) is allowed by the target generate a load and a bswap if needed.
Reviewed By: hfinkel, RKSimon, filcab
Differential Revision: https://reviews.llvm.org/D26149
llvm-svn: 289538
Retrying after fixing overly aggressive load-store forwarding optimization.
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 289221
Handle the case where a sign extension has ended up being split into separate stages (typically to get around vector legal ops) and a zext + sext_in_reg gets inserted.
Differential Revision: https://reviews.llvm.org/D27461
llvm-svn: 288842
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
Forward store values to matching loads down through token
factors. Factored from D14834.
Reviewers: jyknight, hfinkel
Subscribers: hfinkel, nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D26080
llvm-svn: 287773
We have the following DAGCombiner transformations:
(mul (shl X, c1), c2) -> (mul X, c2 << c1)
(mul (shl X, C), Y) -> (shl (mul X, Y), C)
(shl (mul x, c1), c2) -> (mul x, c1 << c2)
Usually the constant shift is optimised by SelectionDAG::getNode when it is
constructed, by SelectionDAG::FoldConstantArithmetic, but when we're dealing
with vectors and one of those vector constants contains an undef element
FoldConstantArithmetic does not fold and we enter an infinite loop.
Fix this by making FoldConstantArithmetic use getNode to decide how to fold each
vector element, the same as FoldConstantVectorArithmetic does, and rather than
adding the constant shift to the work list instead only apply the transformation
if it's already been folded into a constant, as if it's not we're going to loop
endlessly. Additionally add missing NoOpaques to one of those transformations,
which I noticed when writing the tests for this.
Differential Revision: https://reviews.llvm.org/D26605
llvm-svn: 287766
Implemented widening (v2f32) and splitting (v16f64).
On splitting, I use "popcnt" to calculate memory increment.
More type legalization work will come in the next patches.
llvm-svn: 287761
The generic infrastructure to compute the Newton series for reciprocal and
reciprocal square root was conceived to allow a target to compute the series
itself. However, the original code did not properly consider this condition
if returned by a target. This patch addresses the issues to allow a target
to compute the series on its own.
Differential revision: https://reviews.llvm.org/D22975
llvm-svn: 286523
We were failing to extract a constant splat shift value if the shifted value was being masked.
The (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV) combine was unnecessarily preventing this.
llvm-svn: 286454
Summary:
Have MergeConsecutiveStores explicitly return information about the stores
that were merged, so that we can safely determine whether the starting
node has been freed.
Reviewers: chandlerc, bogner, niravd
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D25601
llvm-svn: 285916
2 new intrinsics covering AVX-512 compress/expand functionality.
This implementation includes syntax, DAG builder, operation lowering and tests.
Does not include: handling of illegal data types, codegen prepare pass and the cost model.
llvm-svn: 285876
Instead of asserting that the shift count is != 0 we just bail out
as it's not profitable trying to optimize a node which will be
removed anyway.
Differential Revision: https://reviews.llvm.org/D26098
llvm-svn: 285480
Summary:
Do *not* perform combines such as:
vector_shuffle<4,1,2,3>(build_vector(Ud, C0, C1 C2), scalar_to_vector(X))
->
build_vector(X, C0, C1, C2)
Keeping the shuffle allows lowering the constant build_vector to a materialized
constant vector (such as a vector-load from the constant-pool or some other idiom).
Reviewers: delena, igorb, spatel, mkuper, andreadb, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D25524
llvm-svn: 285063
0 - X --> 0, if the sub is NUW
0 - X --> 0, if X is 0 or the minimum signed value and the sub is NSW
0 - X --> X, if X is 0 or the minimum signed value
This is the DAG equivalent of:
https://reviews.llvm.org/rL284649
plus the fold for the NUW case which already existed in InstSimplify.
Note that we miss a vector fold because of a deficiency in the DAG version of
computeKnownBits().
llvm-svn: 284844
As discussed in D24815, let's start the process of killing off the broken fast-math global
state housed in TargetOptions and eliminate the need for function-level fast-math attributes.
Here we enable two similar folds that are possible when we don't care about signed-zero:
fadd nsz x, 0 --> x
fsub nsz 0, x --> -x
Note that although the test cases include a 'sin' function call, I'm side-stepping the
FMF-on-calls question (and lack of support in the DAG) for now. It's not needed for these
tests - isNegatibleForFree/GetNegatedExpression just look through a ISD::FSIN node.
Also, when we create an FNEG node and propagate the Flags of the FSUB to it, this doesn't
actually do anything today because Flags are silently dropped for any node that is not a
binary operator.
Differential Revision: https://reviews.llvm.org/D25297
llvm-svn: 284824
This is a retry of r284495 which was reverted at r284513 due to use-after-scope bugs
caused by faulty usage of StringRef.
This version also renames a pair of functions:
getRecipEstimateDivEnabled()
getRecipEstimateSqrtEnabled()
as suggested by Eric Christopher.
original commit msg:
[Target] remove TargetRecip class; move reciprocal estimate isel functionality to TargetLowering
This is a follow-up to https://reviews.llvm.org/D24816 - where we changed reciprocal estimates to be function attributes
rather than TargetOptions.
This patch is intended to be a structural, but not functional change. By moving all of the
TargetRecip functionality into TargetLowering, we can remove all of the reciprocal estimate
state, shield the callers from the string format implementation, and simplify/localize the
logic needed for a target to enable this.
If a function has a "reciprocal-estimates" attribute, those settings may override the target's
default reciprocal preferences for whatever operation and data type we're trying to optimize.
If there's no attribute string or specific setting for the op/type pair, just use the target
default settings.
As noted earlier, a better solution would be to move the reciprocal estimate settings to IR
instructions and SDNodes rather than function attributes, but that's a multi-step job that
requires infrastructure improvements. I intend to work on that, but it's not clear how long
it will take to get all the pieces in place.
Differential Revision: https://reviews.llvm.org/D25440
llvm-svn: 284746
This will get the same ConstantSDNode scalar or vector splat value as the current separate dyn_cast<ConstantSDNode> / isVector() approach.
llvm-svn: 284578
This is a follow-up to D24816 - where we changed reciprocal estimates to be function attributes
rather than TargetOptions.
This patch is intended to be a structural, but not functional change. By moving all of the
TargetRecip functionality into TargetLowering, we can remove all of the reciprocal estimate
state, shield the callers from the string format implementation, and simplify/localize the
logic needed for a target to enable this.
If a function has a "reciprocal-estimates" attribute, those settings may override the target's
default reciprocal preferences for whatever operation and data type we're trying to optimize.
If there's no attribute string or specific setting for the op/type pair, just use the target
default settings.
As noted earlier, a better solution would be to move the reciprocal estimate settings to IR
instructions and SDNodes rather than function attributes, but that's a multi-step job that
requires infrastructure improvements. I intend to work on that, but it's not clear how long
it will take to get all the pieces in place.
Differential Revision: https://reviews.llvm.org/D25440
llvm-svn: 284495
As noted in:
https://reviews.llvm.org/D25685
This is the next-to-smallest step needed to enable the ComputeNumSignBits fix in that patch.
In a minor attempt to keep some structure, we're pulling the FP helper over along with its
integer sibling, but clearly we can and should do more refactoring of the similar helper
functions in DAGCombiner and SelectionDAG to simplify and not duplicate functionality.
llvm-svn: 284421
This came up as part of:
https://reviews.llvm.org/D25485
Note that the vector case is missed because ComputeNumSignBits() is deficient for vectors.
llvm-svn: 284395
Eli noted this potential bug in the post-commit thread for:
https://reviews.llvm.org/rL284239
...but I'm not sure how to trigger it, so there's no test case yet.
llvm-svn: 284268
This will be needed by a future commit to support sign/zero extending from v8i8 to v8i64 which requires a sign/zero_extend_vector_inreg to be created which requires v8i8 to be concatenated upto v64i8 and goes through this code.
llvm-svn: 284204
Retrying after upstream changes.
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 284151
Add a number of helper functions to match scalar or vector equivalent constant/splat values to allow most of the combine patterns to be used by vectors.
Differential Revision: https://reviews.llvm.org/D25374
llvm-svn: 284015
This combiner breaks debug experience and should not be run when optimizations are disabled.
For example:
int main() {
int j = 0;
j += 2;
if (j == 2)
return 0;
return 5;
}
When debugging this code compiled in /O0, it should be valid to break at line "j+=2;" and edit the value of j. It should change the return value of the function.
Differential Revision: https://reviews.llvm.org/D19268
llvm-svn: 284014
Fixes a crash in the build_vector -> vector_shuffle combine
when the first vector input is twice as wide as the output,
and the second input vector is even wider.
llvm-svn: 283953
The non-obvious motivation for adding this fold (which already happens in InstCombine)
is that we want to canonicalize IR towards select instructions and canonicalize DAG
nodes towards boolean math. So we need to recreate some folds in the DAG to handle that
change in direction.
An interesting implementation difference for cases like this is that InstCombine
generally works top-down while the DAG goes bottom-up. That means we need to detect
different patterns. In this case, the SimplifyDemandedBits fold prevents us from
performing a zext to sext fold that would then be recognized as a negation of a sext.
llvm-svn: 283900
This generalizes the build_vector -> vector_shuffle combine to support any
number of inputs. The idea is to create a binary tree of shuffles, where
the first layer performs pairwise shuffles of the input vectors placing each
input element into the correct lane, and the rest of the tree blends these
shuffles together.
This doesn't try to be smart and create any sort of "optimal" shuffles.
The assumption is that even a "poor" shuffle sequence is better than extracting
and inserting the elements one by one.
Differential Revision: https://reviews.llvm.org/D24683
llvm-svn: 283480
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill
behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, nhaehnle, jyknight
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 282600
This check currently doesn't seem to do anything useful on any in-tree target:
On non-x86, it always evaluates to false, so we never hit the code path that
creates the shuffle with zero.
On x86, it just forwards to isShuffleMaskLegal(), which is a reasonable thing to
query in general, but doesn't make sense if only restricted to zero blends.
Differential Revision: https://reviews.llvm.org/D24625
llvm-svn: 282567
This allows us to, in some cases, create a vector_shuffle out of a build_vector, when
the inputs to the build are extract_elements from two different vectors, at least one
of which is wider than the output. (E.g. a <8 x i16> being constructed out of
elements from a <16 x i16> and a <8 x i16>).
Differential Revision: https://reviews.llvm.org/D24491
llvm-svn: 281402
To avoid assertion, we must ensure that the inner shift constant is within range before calling ConstantSDNode::getZExtValue(). We already know that the outer shift constant is in range.
Followup to D23007
llvm-svn: 281362
This should make it easier to add cases that we currently don't cover,
like supporting more kinds of type mismatches and more than 2 input vectors.
llvm-svn: 281283
Summary:
An IR load can be invariant, dereferenceable, neither, or both. But
currently, MI's notion of invariance is IR-invariant &&
IR-dereferenceable.
This patch splits up the notions of invariance and dereferenceability at
the MI level. It's NFC, so adds some probably-unnecessary
"is-dereferenceable" checks, which we can remove later if desired.
Reviewers: chandlerc, tstellarAMD
Subscribers: jholewinski, arsenm, nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D23371
llvm-svn: 281151
I might have called this "r246507, the sequel". It fixes the same issue, as the
issue has cropped up in a few more places. The underlying problem is that
isSetCCEquivalent can pick up select_cc nodes with a result type that is not
legal for a setcc node to have, and if we use that type to create new setcc
nodes, nothing fixes that (and so we've violated the contract that the
infrastructure has with the backend regarding setcc node types).
Fixes PR30276.
For convenience, here's the commit message from r246507, which explains the
problem is greater detail:
[DAGCombine] Fixup SETCC legality checking
SETCC is one of those special node types for which operation actions (legality,
etc.) is keyed off of an operand type, not the node's value type. This makes
sense because the value type of a legal SETCC node is determined by its
operands' value type (via the TLI function getSetCCResultType). When the
SDAGBuilder creates SETCC nodes, it either creates them with an MVT::i1 value
type, or directly with the value type provided by TLI.getSetCCResultType.
The first problem being fixed here is that DAGCombine had several places
querying TLI.isOperationLegal on SETCC, but providing the return of
getSetCCResultType, instead of the operand type directly. This does not mean
what the author thought, and "luckily", most in-tree targets have SETCC with
Custom lowering, instead of marking them Legal, so these checks return false
anyway.
The second problem being fixed here is that two of the DAGCombines could create
SETCC nodes with arbitrary (integer) value types; specifically, those that
would simplify:
(setcc a, b, op1) and|or (setcc a, b, op2) -> setcc a, b, op3
(which is possible for some combinations of (op1, op2))
If the operands of the and|or node are actual setcc nodes, then this is not an
issue (because the and|or must share the same type), but, the relevant code in
DAGCombiner::visitANDLike and DAGCombiner::visitORLike actually calls
DAGCombiner::isSetCCEquivalent on each operand, and that function will
recognise setcc-like select_cc nodes with other return types. And, thus, when
creating new SETCC nodes, we need to be careful to respect the value-type
constraint. This is even true before type legalization, because it is quite
possible for the SELECT_CC node to have a legal type that does not happen to
match the corresponding TLI.getSetCCResultType type.
To be explicit, there is nothing that later fixes the value types of SETCC
nodes (if the type is legal, but does not happen to match
TLI.getSetCCResultType). Creating SETCCs with an MVT::i1 value type seems to
work only because, either MVT::i1 is not legal, or it is what
TLI.getSetCCResultType returns if it is legal. Fixing that is a larger change,
however. For the time being, restrict the relevant transformations to produce
only SETCC nodes with a value type matching TLI.getSetCCResultType (or MVT::i1
prior to type legalization).
Fixes PR24636.
llvm-svn: 280767
For the store of a wide value merged from a pair of values, especially int-fp pair,
sometimes it is more efficent to split it into separate narrow stores, which can
remove the bitwise instructions or sink them to colder places.
Now the feature is only enabled on x86 target, and only store of int-fp pair is
splitted. It is possible that the application scope gets extended with perf evidence
support in the future.
Differential Revision: https://reviews.llvm.org/D22840
llvm-svn: 280505
This fixes a regression introduced by revision 268094.
Revision 268094 added the following dag combine rule:
// trunc (shl x, K) -> shl (trunc x), K => K < vt.size / 2
That rule converts a truncate of a shift-by-constant into a shift of a truncated
value. We do this only if the shift count is less than half the size in bits of
the truncated value (K < vt.size / 2).
The problem is that the constraint on the shift count is incorrect, so the rule
doesn't work well in some cases involving vector types. The combine rule should
have been written instead like this:
// trunc (shl x, K) -> shl (trunc x), K => K < vt.getScalarSizeInBits()
Basically, if K is smaller than the "scalar size in bits" of the truncated value
then we know that by "sinking" the truncate into the operand of the shift we
would never accidentally make the shift undefined.
This patch fixes the check on the shift count, and adds test cases to make sure
that we don't regress the behavior.
Differential Revision: https://reviews.llvm.org/D24154
llvm-svn: 280482
Legalization tends to create anyext(trunc) patterns. This should always be
combined - into either a single trunc, a single ext, or nothing if the
types match exactly. But if we happen to combine the trunc first, we may pull
the trunc away from the anyext or make it implicit (e.g. the truncate(extract)
-> extract(bitcast) fold).
To prevent this, we can avoid doing the fold, similarly to how we already handle
fpround(fpextend).
Differential Revision: https://reviews.llvm.org/D23893
llvm-svn: 280386
If the input vector to INSERT_SUBVECTOR is another INSERT_SUBVECTOR, and this inserted subvector replaces the last insertion, then insert into the common source vector.
i.e.
INSERT_SUBVECTOR( INSERT_SUBVECTOR( Vec, SubOld, Idx ), SubNew, Idx ) --> INSERT_SUBVECTOR( Vec, SubNew, Idx )
Differential Revision: https://reviews.llvm.org/D23330
llvm-svn: 278211
As detailed on D22726, much of the shift combining code assume constant values will fit into a uint64_t value and calls ConstantSDNode::getZExtValue where it probably shouldn't (leading to asserts). Using APInt directly avoids this problem but we encounter other assertions if we attempt to compare/operate on 2 APInt of different bitwidths.
This patch adds a helper function to ensure that 2 APInt values are zero extended as required so that they can be safely used together. I've only added an initial example use for this to the '(SHIFT (SHIFT x, c1), c2) --> (SHIFT x, (ADD c1, c2))' combines. Further cases can easily be added as required.
Differential Revision: https://reviews.llvm.org/D23007
llvm-svn: 278141
On modern Intel processors hardware SQRT in many cases is faster than RSQRT
followed by Newton-Raphson refinement. The patch introduces a simple heuristic
to choose between hardware SQRT instruction and Newton-Raphson software
estimation.
The patch treats scalars and vectors differently. The heuristic is that for
scalars the compiler should optimize for latency while for vectors it should
optimize for throughput. It is based on the assumption that throughput bound
code is likely to be vectorized.
Basically, the patch disables scalar NR for big cores and disables NR completely
for Skylake. Firstly, scalar SQRT has shorter latency than NR code in big cores.
Secondly, vector SQRT has been greatly improved in Skylake and has better
throughput compared to NR.
Differential Revision: https://reviews.llvm.org/D21379
llvm-svn: 277725
We used to combine "sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)"
Instead, we should combine to (select (setcc x, y, cc), T, 0) where the value
of T is 1 or -1, depending on the type of the setcc, and getBooleanContents()
for the type if it is not i1.
This fixes PR28504.
llvm-svn: 277371
Using getZExtValue() will assert if the value doesn't fit into uint64_t - SHL was already doing this, I've just updated ASHR/LSHR to match
As mentioned on D22726
llvm-svn: 276855
Summary:
Instead, we take a single flags arg (a bitset).
Also add a default 0 alignment, and change the order of arguments so the
alignment comes before the flags.
This greatly simplifies many callsites, and fixes a bug in
AMDGPUISelLowering, wherein the order of the args to getLoad was
inverted. It also greatly simplifies the process of adding another flag
to getLoad.
Reviewers: chandlerc, tstellarAMD
Subscribers: jholewinski, arsenm, jyknight, dsanders, nemanjai, llvm-commits
Differential Revision: http://reviews.llvm.org/D22249
llvm-svn: 275592
As a result, the urem instruction will not be expanded to a sequence of umull,
lsrs, muls and sub instructions, but just a call to __aeabi_uidivmod.
Differential Revision: http://reviews.llvm.org/D22131
llvm-svn: 274843
Summary:
findBetterNeighborChains may or may not find a better chain for each node it finds, which include the node ("St") that visitSTORE is currently processing. If no better chain is found for St, visitSTORE should continue instead of return SDValue(St, 0), as if it's CombinedTo'ed.
This fixes bug 28130. There might be other ways to make the test pass (see D21409). I think both of the patches are fixing actual bugs revealed by the same testcase.
Reviewers: echristo, wschmidt, hfinkel, kbarton, amehsan, arsenm, nemanjai, bogner
Subscribers: mehdi_amini, nemanjai, llvm-commits
Differential Revision: http://reviews.llvm.org/D21692
llvm-svn: 274644
This reverts commit r259387 because it inserts illegal code after legalization
in some backends where i64 OR type is illegal for example.
llvm-svn: 274573
For the most part this simplifies all callers. There were two places in X86 that needed an explicit makeArrayRef to shorten a statically sized array.
llvm-svn: 274337
Recommiting after correcting over-eager Debug Value transfer fixing PR28270.
[DAG] Previously debug values would transfer debuginfo for the selected
start node for a replacement which allows for debug to be dropped.
Push debug value transfer to occur with node/value replacement in
SelectionDAG, remove now extraneous transfers of debug values.
This refixes PR9817 which was being incompletely checked in the
testsuite.
Reviewers: jyknight
Subscribers: dblaikie, llvm-commits
Differential Revision: http://reviews.llvm.org/D21037
llvm-svn: 273585
Recommiting after fixing over-aggressive assertion
[DAG] Previously debug values would transfer debuginfo for the selected
start node for a replacement which allows for debug to be dropped.
Push debug value transfer to occur with node/value replacement in
SelectionDAG, remove now extraneous transfers of debug values.
This refixes PR9817 which was being incompletely checked in the
testsuite.
Reviewers: jyknight
Subscribers: dblaikie, llvm-commits
Differential Revision: http://reviews.llvm.org/D21037
llvm-svn: 273456
When calculating a square root using Newton-Raphson with two constants,
a naive implementation is to use five multiplications (four muls to calculate
reciprocal square root and another one to calculate the square root itself).
However, after some reassociation and CSE the same result can be obtained
with only four multiplications. Unfortunately, there's no reliable way to do
such a reassociation in the back-end. So, the patch modifies NR code itself
so that it directly builds optimal code for SQRT and doesn't rely on any
further reassociation.
Patch by Nikolai Bozhenov!
Differential Revision: http://reviews.llvm.org/D21127
llvm-svn: 272920
[DAG] Previously debug values would transfer debuginfo for the selected
start node for a replacement which allows for debug to be dropped.
Push debug value transfer to occur with node/value replacement in
SelectionDAG, remove now extraneous transfers of debug values.
This refixes PR9817 which was being incompletely checked in the
testsuite.
Reviewers: jyknight
Subscribers: dblaikie, llvm-commits
Differential Revision: http://reviews.llvm.org/D21037
llvm-svn: 272792
This used to be free, copying and moving DebugLocs became expensive
after the metadata rewrite. Passing by reference eliminates a ton of
track/untrack operations. No functionality change intended.
llvm-svn: 272512
As suggested by clang-tidy's performance-unnecessary-copy-initialization.
This can easily hit lifetime issues, so I audited every change and ran the
tests under asan, which came back clean.
llvm-svn: 272126
Although this was intended to be NFC, the test case wiggle shows a change in
code scheduling/RA caused by a difference in the SDLoc() generation.
Depending on how you look at it, this is the (dis)advantage of exact checking
in regression tests.
llvm-svn: 271526
This should have been converting the size to bytes, but wasn't really.
These should probably all be using getStoreSize instead.
I haven't been able to come up with a meaningful testcase for this.
I can trigger it using combinations of struct loads and stores,
but can't observe a difference in non-broken testcases.
isAlias is only really used during store merging, so I'm not sure how
to get into the vector splitting situation the comment describes
since store merging is only done before type legalization.
llvm-svn: 271356
visitAND, when folding and (load) forgets to check which output of
an indexed load is involved, happily folding the updated address
output on the following testcase:
target datalayout = "e-m:e-i64:64-n32:64"
target triple = "powerpc64le-unknown-linux-gnu"
%typ = type { i32, i32 }
define signext i32 @_Z8access_pP1Tc(%typ* %p, i8 zeroext %type) {
%b = getelementptr inbounds %typ, %typ* %p, i64 0, i32 1
%1 = load i32, i32* %b, align 4
%2 = ptrtoint i32* %b to i64
%3 = and i64 %2, -35184372088833
%4 = inttoptr i64 %3 to i32*
%_msld = load i32, i32* %4, align 4
%zzz = add i32 %1, %_msld
ret i32 %zzz
}
Fix this by checking ResNo.
I've found a few more places that currently neglect to check for
indexed load, and tightened them up as well, but I don't have test
cases for them. In fact, they might not be triggerable at all,
at least with current targets. Still, better safe than sorry.
Differential Revision: http://reviews.llvm.org/D19202
llvm-svn: 267420
The original patch caused crashes because it could derefence a null pointer
for SelectionDAGTargetInfo for targets that do not define it.
Evaluates fmul+fadd -> fmadd combines and similar code sequences in the
machine combiner. It adds support for float and double similar to the existing
integer implementation. The key features are:
- DAGCombiner checks whether it should combine greedily or let the machine
combiner do the evaluation. This is only supported on ARM64.
- It gives preference to throughput over latency: the heuristic used is
to combine always in loops. The targets decides whether the machine
combiner should optimize for throughput or latency.
- Supports for fmadd, f(n)msub, fmla, fmls patterns
- On by default at O3 ffast-math
llvm-svn: 267328
Evaluates fmul+fadd -> fmadd combines and similar code sequences in the
machine combiner. It adds support for float and double similar to the existing
integer implementation. The key features are:
- DAGCombiner checks whether it should combine greedily or let the machine
combiner do the evaluation. This is only supported on ARM64.
- It gives preference to throughput over latency: the heuristic used is
to combine always in loops. The targets decides whether the machine
combiner should optimize for throughput or latency.
- Supports for fmadd, f(n)msub, fmla, fmls patterns
- On by default at O3 ffast-math
llvm-svn: 267098
This patch fixes a bug (PR26827) when using anti-aliasing in store
merging. This sets the chain users of the component stores to point to
the new store instead of the component stores chain parent.
Reviewers: jyknight
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D18909
llvm-svn: 266217
xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) was only being combined at the AfterLegalizeTypes stage, this patch permits the combine to occur anytime before then as well.
The main aim with this to improve the ability to recognise bitmasks that can be converted to shuffles.
I had to modify a number of AVX512 mask tests as the basic bitcast to/from scalar pattern was being stripped out, preventing testing of the mmask bitops. By replacing the bitcasts with loads we can get almost the same result.
Differential Revision: http://reviews.llvm.org/D18944
llvm-svn: 265998
In Memcpy lowering we had missed a dependence from the load of the
operation to successor operations. This causes us to potentially
construct an in initial DAG with a memory dependence not fully
represented in the chain sub-DAG but rather require looking at the
entire DAG breaking alias analysis by allowing incorrect repositioning
of memory operations.
To work around this, r200033 changed DAGCombiner::GatherAllAliases to be
conservative if any possible issues to happen. Unfortunately this check
forbade many non-problematic situations as well. For example, it's
common for incoming argument lowering to add a non-aliasing load hanging
off of EntryNode. Then, if GatherAllAliases visited EntryNode, it would
find that other (unvisited) use of the EntryNode chain, and just give up
entirely. Furthermore, the check was incomplete: it would not actually
detect all such potentially problematic DAG constructions, because
GatherAllAliases did not guarantee to visit all chain nodes going up to
the root EntryNode. This is in general fine -- giving up early will just
miss a potential optimization, not generate incorrect results. But, for
this non-chain dependency detection code, it's possible that you could
have a load attached to a higher-up chain node than any which were
visited. If that load aliases your store, but the only dependency is
through the value operand of a non-aliasing store, it would've been
missed by this code, and potentially reordered.
With the dependence added, this check can be removed and Alias Analysis
can be much more aggressive. This fixes code quality regression in the
Consecutive Store Merge cleanup (D14834).
Test Change:
ppc64-align-long-double.ll now may see multiple serializations
of its stores
Differential Revision: http://reviews.llvm.org/D18062
llvm-svn: 265836
Change isConsecutiveLoads to check that loads are non-volatile as this
is a requirement for any load merges. Propagate change to two callers.
Reviewers: RKSimon
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D18546
llvm-svn: 265013
When merging stores in DAGCombiner, add check to ensure that no
dependenices exist that would cause the construction of a cycle in our
DAG. This may happen if one store has a data dependence on another
instruction (e.g. a load) which itself has a (chain) dependence on
another store being merged. These stores cannot be merged safely and
doing so results in a cycle that is discovered in LegalizeDAG.
This test is only done in cases where Antialias analysis is used (UseAA)
as non-AA store merge candidates will be merged logically after all
loads which have been checked to not alias.
Reviewers: ahatanak, spatel, niravd, arsenm, hfinkel, tstellarAMD, jyknight
Subscribers: llvm-commits, tberghammer, danalbert, srhines
Differential Revision: http://reviews.llvm.org/D18336
llvm-svn: 264461
Summary:
extract_vector_elt can cause an implicit any_ext if the types don't
match. When processing the following pattern:
(and (extract_vector_elt (load ([non_ext|any_ext|zero_ext] V))), c)
DAGCombine was ignoring the possible extend, and sometimes removing
the AND even though it was required to maintain some of the bits
in the result to 0, resulting in a miscompile.
This change fixes the issue by limiting the transformation only to
cases where the extract_vector_elt doesn't perform the implicit
extend.
Reviewers: t.p.northover, jmolloy
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D18247
llvm-svn: 263935
Generalise the existing SIGN_EXTEND to SIGN_EXTEND_VECTOR_INREG combine to support zero extension as well and get rid of a lot of unnecessary ANY_EXTEND + mask patterns.
Reapplied with a fix for PR26870 (avoid premature use of TargetConstant in ZERO_EXTEND_VECTOR_INREG expansion).
Differential Revision: http://reviews.llvm.org/D17691
llvm-svn: 263159
The divrem combine assumed the type of the div/rem is simple, which isn't
necessarily true. This probably worked fine until r250825, since it only
saw legal types, but now breaks when it runs as a pre-type-legalization
combine.
This fixes PR26835.
Differential Revision: http://reviews.llvm.org/D17878
llvm-svn: 262746
When div+rem calls on the same arguments are found, the ARM back-end merges the
two calls into one __aeabi_divmod call for up to 32-bits values. However,
for 64-bit values, which also have a lib call (__aeabi_ldivmod), it wasn't
merging the calls, and thus calling ldivmod twice and spilling the temporary
results, which generated pretty bad code.
This patch legalises 64-bit lib calls for divmod, so that now all the spilling
and the second call are gone. It also relaxes the DivRem combiner a bit on the
legal type check, since it was already checking for isLegalOrCustom on every
value, so the extra check for isTypeLegal was redundant.
Second attempt, creating TLI.isOperationCustom like isOperationExpand, to make
sure we only emit valid types or the ones that were explicitly marked as custom.
Now, passing check-all and test-suite on x86, ARM and AArch64.
This patch fixes PR17193 (and a long time FIXME in the tests).
llvm-svn: 262738
Generalise the existing SIGN_EXTEND to SIGN_EXTEND_VECTOR_INREG combine to support zero extension as well and get rid of a lot of unnecessary ANY_EXTEND + mask patterns.
Differential Revision: http://reviews.llvm.org/D17691
llvm-svn: 262599
When div+rem calls on the same arguments are found, the ARM back-end merges the
two calls into one __aeabi_divmod call for up to 32-bits values. However,
for 64-bit values, which also have a lib call (__aeabi_ldivmod), it wasn't
merging the calls, and thus calling ldivmod twice and spilling the temporary
results, which generated pretty bad code.
This patch legalises 64-bit lib calls for divmod, so that now all the spilling
and the second call are gone. It also relaxes the DivRem combiner a bit on the
legal type check, since it was already checking for isLegalOrCustom on every
value, so the extra check for isTypeLegal was redundant.
This patch fixes PR17193 (and a long time FIXME in the tests).
llvm-svn: 262507
On AMDGPU where operations i64 operations are often bitcasted to v2i32
and back, this pattern shows up regularly where it breaks some
expected combines on i64, such as load width reducing.
This fixes some test failures in a future commit when i64 loads
are changed to promote.
llvm-svn: 262397
This reverts commit r262316.
It seems that my change breaks an out-of-tree chromium buildbot, so
I'm reverting this in order to investigate the situation further.
llvm-svn: 262387
Summary:
This patch modifies the existing comparison, branch, conditional-move
and select patterns, and adds new ones where needed. Also, the updated
SLT{u,i,iu} set of instructions generate a GPR width result.
The majority of the code changes in the Mips back-end fix the wrong
assumption that the result of SETCC nodes always produce an i32 value.
The changes in the common code path account for the fact that in 64-bit
MIPS targets, i1 is promoted to i32 instead of i64.
Reviewers: dsanders
Subscribers: dsanders, llvm-commits
Differential Revision: http://reviews.llvm.org/D10970
llvm-svn: 262316
In the case where op = add, y = base_ptr, and x = offset, this
transform:
(op y, (op x, c1)) -> (op (op x, y), c1)
breaks the canonical form of add by putting the base pointer in the
second operand and the offset in the first.
This fix is important for the R600 target, because for some address
spaces the base pointer and the offset are stored in separate register
classes. The old pattern caused the ISel code for matching addressing
modes to put the base pointer and offset in the wrong register classes,
which required no-trivial code transformations to fix.
llvm-svn: 262148
Summary:
This patch skips DAG combine of fp_round (fp_round x) if it results in
an fp_round from f80 to f16.
fp_round from f80 to f16 always generates an expensive (and as yet,
unimplemented) libcall to __truncxfhf2. This prevents selection of
native f16 conversion instructions from f32 or f64. Moreover, the first
(value-preserving) fp_round from f80 to either f32 or f64 may become a
NOP in platforms like x86.
Reviewers: ab
Subscribers: srhines, llvm-commits
Differential Revision: http://reviews.llvm.org/D17221
llvm-svn: 260769
This patch consists of two parts: a performance fix in DAGCombiner.cpp
and a correctness fix in SelectionDAG.cpp.
The test case tests the bug that's uncovered by the performance fix, and
fixed by the correctness fix.
The performance fix keeps the containers required by the
hasPredecessorHelper (which is a lazy DFS) and reuse them. Since
hasPredecessorHelper is called in a loop, the overall efficiency reduced
from O(n^2) to O(n), where n is the number of SDNodes.
The correctness fix keeps iterating the neighbor list even if it's time
to early return. It will return after finishing adding all neighbors to
Worklist, so that no neighbors are discarded due to the original early
return.
llvm-svn: 259691
Summary:
This is an extension to the existing implementation of r242436 which
restricts to only select inputs. This version fixes missed opportunities
in pr26084 by attempting to lower conditional compare sequences of
and/or trees with setcc leafs. This will additionaly handle the case
when a tree with select input is not a conjunction-disjunction tree
but some of the sub trees are conjunction-disjunction trees.
Reviewers: jmolloy, t.p.northover, mcrosier, MatzeB
Subscribers: mcrosier, llvm-commits, junbuml, haicheng, mssimpso, gberry
Differential Revision: http://reviews.llvm.org/D16291
llvm-svn: 259387
Summary:
findBetterNeighborChains does not handle volatile or indexed stores.
However, it did not check when adding stores to ChainedStores.
Reviewers: arsenm
Differential Revision: http://reviews.llvm.org/D16463
llvm-svn: 259024
This reapplies r258296 and r258366, and also fixes an existing bug in
SelectionDAG.cpp's isMemSrcFromString, neglecting to account for the
offset in a GlobalAddressSDNode, which is uncovered by those patches.
llvm-svn: 258482
This reverts r258296 and the follow up r258366. With this change, we
miscompiled the following program on Windows:
#include <string>
#include <iostream>
static const char kData[] = "asdf jkl;";
int main() {
std::string s(kData + 3, sizeof(kData) - 3);
std::cout << s << '\n';
}
llvm-svn: 258465
SelectionDAG previously missed opportunities to fold constants into
GlobalAddresses in several areas. For example, given `(add (add GA, c1), y)`, it
would often reassociate to `(add (add GA, y), c1)`, missing the opportunity to
create `(add GA+c, y)`. This isn't often visible on targets such as X86 which
effectively reassociate adds in their complex address-mode folding logic,
however it is currently visible on WebAssembly since it currently has very
simple address mode folding code that doesn't reassociate anything.
This patch fixes this by making SelectionDAG fold offsets into GlobalAddresses
at the same times that it folds constants together, so that it doesn't miss any
opportunities to perform such folding.
Differential Revision: http://reviews.llvm.org/D16090
llvm-svn: 258296
In an inbounds getelementptr, when an index produces a constant non-negative
offset to add to the base, the add can be assumed to not have unsigned overflow.
This relies on the assumption that addresses can't occupy more than half the
address space, which isn't possible in C because it wouldn't be possible to
represent the difference between the start of the object and one-past-the-end
in a ptrdiff_t.
Setting the NoUnsignedWrap flag is theoretically useful in general, and is
specifically useful to the WebAssembly backend, since it permits stronger
constant offset folding.
Differential Revision: http://reviews.llvm.org/D15544
llvm-svn: 256890
Almost all these changes are conditioned and only apply to the new
x86-64 f128 type configuration, which will be enabled in a follow up
patch. They are required together to make new f128 work. If there is
any error, we should fix or revert them as a whole.
These changes should have no impact to current configurations.
* Relax type legalization checks to accept new f128 type configuration,
whose TypeAction is TypeSoftenFloat, not TypeLegal, but also has
TLI.isTypeLegal true.
* Relax GetSoftenedFloat to return in some cases f128 type SDValue,
which is TLI.isTypeLegal but not "softened" to i128 node.
* Allow customized FABS, FNEG, FCOPYSIGN on new f128 type configuration,
to generate optimized bitwise operators for libm functions.
* Enhance related Lower* functions to handle f128 type.
* Enhance DAGTypeLegalizer::run, SoftenFloatResult, and related functions
to keep new f128 type in register, and convert f128 operators to library calls.
* Fix Combiner, Emitter, Legalizer routines that did not handle f128 type.
* Add ExpandConstant to handle i128 constants, ExpandNode
to handle ISD::Constant node.
* Add one more parameter to getCommonSubClass and firstCommonClass,
to guarantee that returned common sub class will contain the specified
simple value type.
This extra parameter is used by EmitCopyFromReg in InstrEmitter.cpp.
* Fix infinite loop in getTypeLegalizationCost when f128 is the value type.
* Fix printOperand to handle null operand.
* Enhance ISD::BITCAST node to handle f128 constant.
* Expand new f128 type for BR_CC, SELECT_CC, SELECT, SETCC nodes.
* Enhance X86AsmPrinter to emit f128 values in comments.
Differential Revision: http://reviews.llvm.org/D15134
llvm-svn: 254653
Summary:
Many target lowerings copy-paste the code to test SDValues for known constants.
This code can instead be shared in SelectionDAG.cpp, and reused in the targets.
Reviewers: MatzeB, andreadb, tstellarAMD
Subscribers: arsenm, jyknight, llvm-commits
Differential Revision: http://reviews.llvm.org/D14945
llvm-svn: 254085
When MergeConsecutiveStores() combines two loads and two stores into
wider loads and stores, the chain users of both of the original loads
must be transfered to the new load, because it may be that a chain
user only depends on one of the loads.
New test case: test/CodeGen/SystemZ/dag-combine-01.ll
Reviewed by James Y Knight.
Bugzilla: https://llvm.org/bugs/show_bug.cgi?id=25310#c6
llvm-svn: 253779
Summary:
Don't fold
(zext (and (load x), cst)) -> (and (zextload x), (zext cst))
if
(and (load x) cst)
will match as a zextload already and has additional users.
For example, the following IR:
%load = load i32, i32* %ptr, align 8
%load16 = and i32 %load, 65535
%load64 = zext i32 %load16 to i64
store i32 %load16, i32* %dst1, align 4
store i64 %load64, i64* %dst2, align 8
used to produce the following aarch64 code:
ldr w8, [x0]
and w9, w8, #0xffff
and x8, x8, #0xffff
str w9, [x1]
str x8, [x2]
but with this change produces the following aarch64 code:
ldrh w8, [x0]
str w8, [x1]
str x8, [x2]
Reviewers: resistor, mcrosier
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D14340
llvm-svn: 252789
1) PR25154. This is basically a repeat of PR18102, which was fixed in
r200201, and broken again by r234430. The latter changed which of the
store nodes was merged into from the first to the last. Thus, we now
also need to prefer merging a later store at a given address into the
target node, instead of an earlier one.
2) While investigating that, I also realized I'd introduced a bug in
r236850. There, I removed a check for alignment -- not realizing that
nothing except the alignment check was ensuring that none of the stores
were overlapping! This is a really bogus way to ensure there's no
aliased stores.
A better solution to both of these issues is likely to always use the
code added in the 'if (UseAA)' branches which rearrange the chain based
on a more principled analysis. I'll look into whether that can be used
always, but in the interest of getting things back to working, I think a
minimal change makes sense.
llvm-svn: 251816
This is a usage of the IR-level fast-math-flags now that they are propagated to SDNodes.
This was originally part of D8900.
Removing the global 'enable-unsafe-fp-math' checks will require auto-upgrade and
possibly other changes.
Differential Revision: http://reviews.llvm.org/D9708
llvm-svn: 251450
When taking the remainder of a value divided by a constant, visitREM()
attempts to convert the REM to a longer but faster sequence of instructions.
This conversion calls combine() on a speculative DIV instruction. Commit
rL250825 may cause this combine() to return a DIVREM, corrupting nearby nodes.
Flow eventually hits unreachable().
This patch adds a test case and a check to prevent visitREM() from trying
to convert the REM instruction in cases where a DIVREM is possible.
See http://reviews.llvm.org/D14035
llvm-svn: 251373
We don't need a mask of a rotation result to be a constant splat - any constant scalar/vector can be usefully folded.
Followup to D13851.
llvm-svn: 251197
This patch adds support for lowering to the XOP VPROT / VPROTI vector bit rotation instructions.
This has required changes to the DAGCombiner rotation pattern matching to support vector types - so far I've only changed it to support splat vectors, but generalising this further is feasible in the future.
Differential Revision: http://reviews.llvm.org/D13851
llvm-svn: 251188
When we fold "mul ((add x, c1), c1)" -> "add ((mul x, c2), c1*c2)", we bail if (add x, c1) has multiple
users which would result in an extra add instruction.
In such cases, this patch adds a check to see if we can eliminate a multiply instruction in exchange for the extra add.
I also added the capability of doing the existing optimization with non-splatted vectors (splatted also works).
Differential Revision: http://reviews.llvm.org/D13740
llvm-svn: 251028
Summary:
In addition to moving the code over, this patch amends the DIV,REM -> DIVREM
combining to run on all affected nodes at once: if the nodes are converted
to DIVREM one at a time, then the resulting DIVREM may get legalized by the
backend into something target-specific that we won't be able to recognize
and correlate with the remaining nodes.
The motivation is to "prepare terrain" for D13862: when we set DIV and REM
to be legalized to libcalls, instead of the DIVREM, we otherwise lose the
ability to combine them together. To prevent this, we need to take the
DIV,REM -> DIVREM combining out of the lowering stage.
Reviewers: RKSimon, eli.friedman, rengolin
Subscribers: john.brawn, rengolin, llvm-commits
Differential Revision: http://reviews.llvm.org/D13733
llvm-svn: 250825
Summary:
Caching SDLoc(N), instead of recreating it in every single
function call, keeps the code denser, and allows to unwrap long lines.
Reviewers: sunfish, atrick, sdmitrouk
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D13726
llvm-svn: 250305
Summary: The two implementations had more code in common than not.
Reviewers: sunfish, MatzeB, sdmitrouk
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D13724
llvm-svn: 250302
The comment says this was stopped because it was unlikely to be
profitable. This is not true if you want to combine vector loads
with multiple components.
For a simple case that looks like
t0 = load t0 ...
t1 = load t0 ...
t2 = load t0 ...
t3 = load t0 ...
t4 = store t0:1, t0:1
t5 = store t4, t1:0
t6 = store t5, t2:0
t7 = store t6, t3:0
We want to get all of these stores onto a chain
that is a TokenFactor of these N loads. This mostly
solves the AMDGPU merge-stores.ll regressions
with -combiner-alias-analysis for merging vector
stores of vector loads.
llvm-svn: 250138
This basic combine was surprisingly missing.
AMDGPU legalizes many operations in terms of 32-bit vector components,
so not doing this results in many extra copies and subregister extracts
that need to be cleaned up later.
InstCombine already does this for the hasOneUse case. The target hook
is to fix a handful of tests which break (e.g. ARM/vmov.ll) which turn
from a vector materialize repeated immediate instruction to a constant
vector load with more scalar copies from it.
llvm-svn: 250129
We have a number of functions that implement constant folding of vectors (unary and binary ops) in near identical manners (and the differences don't appear to be critical).
This patch introduces a common implementation (SelectionDAG::FoldConstantVectorArithmetic) and calls this in both the unary and binary op cases.
After this initial patch I intend to begin enabling vector constant folding for a wider number of opcodes in SelectionDAG::getNode().
Differential Revision: http://reviews.llvm.org/D13665
llvm-svn: 250118
visitSIGN_EXTEND_INREG calls SelectionDAG::getNode to constant fold scalar constants but handles vector constants itself, despite getNode being capable of dealing with them.
This required a minor change to the getNode implementation to actually deal with cases where the scalars of a BUILD_VECTOR were wider integers than the vector type - which was the only extra ability of the visitSIGN_EXTEND_INREG implementation.
No codegen intended and all existing tests remain the same.
llvm-svn: 249236
When AA is being used, non-aliasing stores are canonicalized to use the same
chain, and DAGCombiner::getStoreMergeAndAliasCandidates can take advantage of
this by looking only as users of a store's chain operand. However, user
iteration is not result-number specific, we need to check that the use is as a
chain operand, and not via some other operand. It is certainly possible to have
another potentially-aliasing store, which shares the first's base pointer, and
uses the first's chain's node via some other operand.
Failure to catch this situation caused, at least in the included test case, an
assert later because the relative sequence-number ordering caused later
replacement to create a cycle in the DAG.
llvm-svn: 248698
This is a redo of D7208 ( r227242 - http://llvm.org/viewvc/llvm-project?view=revision&revision=227242 ).
The patch was reverted because an AArch64 target could infinite loop after the change in DAGCombiner
to merge vector stores. That happened because AArch64's allowsMisalignedMemoryAccesses() wasn't telling
the truth. It reported all unaligned memory accesses as fast, but then split some 128-bit unaligned
accesses up in performSTORECombine() because they are slow.
This patch attempts to fix the problem in AArch's allowsMisalignedMemoryAccesses() while preserving
existing (perhaps questionable) lowering behavior.
The x86 test shows that store merging is working as intended for a target with fast 32-byte unaligned
stores.
Differential Revision: http://reviews.llvm.org/D12635
llvm-svn: 248622
If the stores are storing values from loads which partially
alias the stores, we could end up placing the merged loads
and stores on the same chain which has the potential to break.
Each store may have a different chain dependency on only some
of the original loads. Create a new TokenFactor to capture all
of the required dependencies of the stores rather than assuming
all stores can use the same chain.
The testcase is a situation where this happens, although
it does not have an observable change from this. The DAG nodes
just happened to not be reordered before despite this missing
chain dependency.
This is based on an off-list report for an out of tree target
which regressed due to r246307 and I haven't managed to find a case
where the nodes do end up reordered with an in tree target.
llvm-svn: 248468
This patch adds support for combining patterns such as (FMUL(FADD(1.0, x), y)) and (FMUL(FSUB(x, 1.0), y)) to their FMA equivalents.
This is useful in particular for linear interpolation cases such as (FADD(FMUL(x, t), FMUL(y, FSUB(1.0, t))))
Differential Revision: http://reviews.llvm.org/D13003
llvm-svn: 248210
If storing multiple FP constants, some subset of the stores
would be replaced with integers due to visit order, so
MergeConsecutiveStores would only partially merge
these.
llvm-svn: 248169
After D10403, we had FMF in the DAG but disabled by default. Nick reported no crashing errors after some stress testing,
so I enabled them at r243687. However, Escha soon notified us of a bug not covered by any in-tree regression tests:
if we don't propagate the flags, we may fail to CSE DAG nodes because differing FMF causes them to not match. There is
one test case in this patch to prove that point.
This patch hopes to fix or leave a 'TODO' for all of the in-tree places where we create nodes that are FMF-capable. I
did this by putting an assert in SelectionDAG.getNode() to find any FMF-capable node that was being created without FMF
( D11807 ). I then ran all regression tests and test-suite and confirmed that everything passes.
This patch exposes remaining work to get DAG FMF to be fully functional: (1) add the flags to non-binary nodes such as
FCMP, FMA and FNEG; (2) add the flags to intrinsics; (3) use the flags as conditions for transforms rather than the
current global settings.
Differential Revision: http://reviews.llvm.org/D12095
llvm-svn: 247815
Summary:
The BUILD_VECTOR node will truncate its operators to match the
type. We need to take this into account when constant folding -
we need to perform a truncation before constant folding the elements.
This is because the upper bits can change the result, depending on
the operation type (for example this is the case for min/max).
This change also adds a regression test.
Reviewers: jmolloy
Subscribers: jmolloy, llvm-commits
Differential Revision: http://reviews.llvm.org/D12697
llvm-svn: 247265
Use and check the 'IsFast' optional parameter to TLI.allowsMemoryAccess() any time
we have a merged access candidate. Without this patch, we were generating unaligned
16-byte (SSE) memops for x86 targets where those accesses are slow.
This change was mentioned in:
http://reviews.llvm.org/D10662 and
http://reviews.llvm.org/D10905
and will help solve PR21711.
Differential Revision: http://reviews.llvm.org/D12573
llvm-svn: 246771
SETCC is one of those special node types for which operation actions (legality,
etc.) is keyed off of an operand type, not the node's value type. This makes
sense because the value type of a legal SETCC node is determined by its
operands' value type (via the TLI function getSetCCResultType). When the
SDAGBuilder creates SETCC nodes, it either creates them with an MVT::i1 value
type, or directly with the value type provided by TLI.getSetCCResultType.
The first problem being fixed here is that DAGCombine had several places
querying TLI.isOperationLegal on SETCC, but providing the return of
getSetCCResultType, instead of the operand type directly. This does not mean
what the author thought, and "luckily", most in-tree targets have SETCC with
Custom lowering, instead of marking them Legal, so these checks return false
anyway.
The second problem being fixed here is that two of the DAGCombines could create
SETCC nodes with arbitrary (integer) value types; specifically, those that
would simplify:
(setcc a, b, op1) and|or (setcc a, b, op2) -> setcc a, b, op3
(which is possible for some combinations of (op1, op2))
If the operands of the and|or node are actual setcc nodes, then this is not an
issue (because the and|or must share the same type), but, the relevant code in
DAGCombiner::visitANDLike and DAGCombiner::visitORLike actually calls
DAGCombiner::isSetCCEquivalent on each operand, and that function will
recognise setcc-like select_cc nodes with other return types. And, thus, when
creating new SETCC nodes, we need to be careful to respect the value-type
constraint. This is even true before type legalization, because it is quite
possible for the SELECT_CC node to have a legal type that does not happen to
match the corresponding TLI.getSetCCResultType type.
To be explicit, there is nothing that later fixes the value types of SETCC
nodes (if the type is legal, but does not happen to match
TLI.getSetCCResultType). Creating SETCCs with an MVT::i1 value type seems to
work only because, either MVT::i1 is not legal, or it is what
TLI.getSetCCResultType returns if it is legal. Fixing that is a larger change,
however. For the time being, restrict the relevant transformations to produce
only SETCC nodes with a value type matching TLI.getSetCCResultType (or MVT::i1
prior to type legalization).
Fixes PR24636.
llvm-svn: 246507
This was part of D7208 (r227242), but that commit was reverted because it exposed
a bug in AArch64 lowering. I should have that fixed and the rest of the commit
reinstated soon.
llvm-svn: 246493
DAGCombine has a utility wrapper around TLI's getSetCCResultType; use it in the
one place in DAGCombine still directly calling the TLI function. NFC.
llvm-svn: 246482
This code was dead when it was committed in r23665 (Oct 7, 2005), and before it
reaches its 10th anniversary, it really should go. We can always bring it back
if we'd like, but it forms more SETCC nodes, and the way we do legality
checking on SETCC nodes is wrong in a number of places, and removing this means
fewer places to fix. NFC.
llvm-svn: 246466
When combiner AA is enabled, look at stores on the same chain.
Non-aliasing stores are moved to the same chain so the existing
code fails because it expects to find an adajcent store on a consecutive
chain.
Because of how DAGCombiner tries these store combines,
MergeConsecutiveStores doesn't see the correct set of stores on the chain
when it visits the other stores. Each store individually has its chain
fixed before trying to merge consecutive stores, and then tries to merge
stores from that point before the other stores have been processed to
have their chains fixed. To fix this, attempt to use FindBetterChain
on any possibly neighboring stores in visitSTORE.
Suppose you have 4 32-bit stores that should be merged into 1 vector
store. One store would be visited first, fixing the chain. What happens is
because not all of the store chains have yet been fixed, 2 of the stores
are merged. The other 2 stores later have their chains fixed,
but because the other stores were already merged, they have different
memory types and merging the two different sized stores is not
supported and would be more difficult to handle.
llvm-svn: 246307
Fixes PR24602: r245689 introduced an unguarded use of
SelectionDAG::FoldConstantArithmetic, which returns 0 when it fails
because of opaque (hoisted) constants.
llvm-svn: 246217
The FP16_TO_FP node only uses the bottom 16 bits of its input, so the
following pattern can be optimised by removing the AND:
(FP16_TO_FP (AND op, 0xffff)) -> (FP16_TO_FP op)
This is a common pattern for ARM targets when functions have __fp16
arguments, as they are passed as floats (so that they get passed in the
correct registers), but then bitcast and truncated to ignore the top 16
bits.
llvm-svn: 245832
This is intended to improve code generation for GEPs, as the index value is
shifted by the element size and in GEPs of multi-dimensional arrays the index
of higher dimensions is multiplied by the lower dimension size.
Differential Revision: http://reviews.llvm.org/D12197
llvm-svn: 245689
We still need to add constant folding of vector comparisons to fold the tests for targets that don't support the respective min/max nodes
I needed to update 2011-12-06-AVXVectorExtractCombine to load a vector instead of using a constant vector to prevent it folding
Differential Revision: http://reviews.llvm.org/D12118
llvm-svn: 245503
Check to see if this is a CONCAT_VECTORS of a bunch of EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector inputs come from at most two distinct vectors the same size as the result, attempt to turn this into a legal shuffle.
Differential Revision: http://reviews.llvm.org/D12125
llvm-svn: 245490
This removes the isPow2SDivCheap() query, as it is not currently used in
any meaningful way. isIntDivCheap() no longer relies on a state variable
(as all in-tree target set it to false), but the interface allows querying
based on the type optimization level.
NFC.
Differential Revision: http://reviews.llvm.org/D12082
llvm-svn: 245430
The current code normalizes select(C0, x, select(C1, x, y)) towards
select(C0|C1, x, y) if the targets prefers that form. This patch adds an
additional rule that if the select(C1, x, y) part already exists in the
function then we want to normalize into the other direction because the
effects of reusing the existing value are bigger than transforming into
the target preferred form.
This addresses regressions following r238793, see also:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20150727/290272.html
Differential Revision: http://reviews.llvm.org/D11616
llvm-svn: 245350
For cases where we TRUNCATE and then ZERO_EXTEND to a larger size (often from vector legalization), see if we can mask the source data and then ZERO_EXTEND (instead of after a ANY_EXTEND). This can help avoid having to generate a larger mask, and possibly applying it to several sub-vectors.
(zext (truncate x)) -> (zext (and(x, m))
Includes a minor patch to SystemZ to better recognise 8/16-bit zero extension patterns from RISBG bit-extraction code.
This is the first of a number of minor patches to help improve the conversion of byte masks to clear mask shuffles.
Differential Revision: http://reviews.llvm.org/D11764
llvm-svn: 245160
This commit removes the global manager variable which is responsible for
storing and allocating pseudo source values and instead it introduces a new
manager class named 'PseudoSourceValueManager'. Machine functions now own an
instance of the pseudo source value manager class.
This commit also modifies the 'get...' methods in the 'MachinePointerInfo'
class to construct pseudo source values using the instance of the pseudo
source value manager object from the machine function.
This commit updates calls to the 'get...' methods from the 'MachinePointerInfo'
class in a lot of different files because those calls now need to pass in a
reference to a machine function to those methods.
This change will make it easier to serialize pseudo source values as it will
enable me to transform the mips specific MipsCallEntry PseudoSourceValue
subclass into two target independent subclasses.
Reviewers: Akira Hatanaka
llvm-svn: 244693
Summary:
For example:
s6 = s0*s5;
s2 = s6*s6 + s6;
...
s4 = s6*s3;
We notice that it is possible for s2 is folded to fma (s0, s5, fmul (s6 s6)).
This only happens when Aggressive is true, otherwise hasOneUse() check
already prevents from folding the multiplication with more uses.
Test Plan: test/CodeGen/NVPTX/fma-assoc.ll
Patch by Xuetian Weng
Reviewers: hfinkel, apazos, jingyue, ohsallen, arsenm
Subscribers: arsenm, jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D11855
llvm-svn: 244649
Create wrapper methods in the Function class for the OptimizeForSize and MinSize
attributes. We want to hide the logic of "or'ing" them together when optimizing
just for size (-Os).
Currently, we are not consistent about this and rely on a front-end to always set
OptimizeForSize (-Os) if MinSize (-Oz) is on. Thus, there are 18 FIXME changes here
that should be added as follow-on patches with regression tests.
This patch is NFC-intended: it just replaces existing direct accesses of the attributes
by the equivalent wrapper call.
Differential Revision: http://reviews.llvm.org/D11734
llvm-svn: 243994
The XformToShuffleWithZero method currently checks AND masks at the per-lane level for all-one and all-zero constants and attempts to convert them to legal shuffle clear masks.
This patch generalises XformToShuffleWithZero, splitting and checking the sub-lanes of the constants down to the byte level to see if any legal shuffle clear masks are possible. This allows a lot of masks (often from legalization or truncation) to be folded into existing shuffle patterns and removes a lot of constant mask loading.
There are a few examples of poor shuffle lowering that are exposed by this patch that will be cleaned up in future patches (e.g. merging shuffles that are separated by bitcasts, x86 legalized v8i8 zero extension uses PMOVZX+AND+AND instead of AND+PMOVZX, etc.)
Differential Revision: http://reviews.llvm.org/D11518
llvm-svn: 243831
Making allowableAlignment() more accessible was suggested as a predecessor patch
for D10662, so I've pulled it into TargetLowering. This let's us remove 4 instances
of duplicate logic in LegalizeDAG.
There's a subtle functional change in the implementation: the existing
allowableAlignment() code was using getPrefTypeAlignment() when checking
alignment with the DataLayout and assumed that was fast. In this implementation,
we use getABITypeAlignment() and assume that is fast. See the TODO comment or the
discussion in the Phab review for future improvements in this implementation
(don't use the data layout at all).
There are no regression test changes from this difference, and I'm not sure how to
expose it via a test. I think we actually do want to provide the 'Fast' param when
checking this from DAGCombiner::MergeConsecutiveStores(). Ie, we shouldn't merge
stores if the new stores are not going to be fast. But that change will require
fixing allowsMisalignedMemoryAccess() overrides as noted in D10662.
Differential Revision: http://reviews.llvm.org/D10905
llvm-svn: 243549
PR24141: https://llvm.org/bugs/show_bug.cgi?id=24141
contains a test case where we have duplicate entries in a node's uses() list.
After r241826, we use CombineTo() to delete dead nodes when combining the uses into
reciprocal multiplies, but this fails if we encounter the just-deleted node again in
the list.
The solution in this patch is to not add duplicate entries to the list of users that
we will subsequently iterate over. For the test case, this avoids triggering the
combine divisors logic entirely because there really is only one user of the divisor.
Differential Revision: http://reviews.llvm.org/D11345
llvm-svn: 243500
This fix was suggested as part of D11345 and is part of fixing PR24141.
With this change, we can avoid walking the uses of a divisor node if the target
doesn't want the combineRepeatedFPDivisors transform in the first place.
There is no NFC-intended other than that.
Differential Revision: http://reviews.llvm.org/D11531
llvm-svn: 243498
We don't bitcast the UNDEFs - that is done in visitVECTOR_SHUFFLE, and the getValueType should come from the operand's SDValue not the SDNode.
llvm-svn: 242640
The motivation is to allow GatherAllAliases / FindBetterChain
to not give up on dependent loads of a pointer from constant memory.
This is important for AMDGPU, because most loads are pointers
derived from a load of a kernel argument from constant memory.
llvm-svn: 241948
This patch fixes bugs that were exposed by the addition of fast-math-flags in the DAG:
r237046 ( http://reviews.llvm.org/rL237046 ):
1. When replacing a division node, it's not enough to RAUW.
We should call CombineTo() to delete dead nodes and combine again.
2. Because we are changing the DAG, we can't return an empty SDValue
after the transform. As the code comments say:
Visitation implementation - Implement dag node combining for different node types.
The semantics are as follows: Return Value:
SDValue.getNode() == 0 - No change was made
SDValue.getNode() == N - N was replaced, is dead and has been handled.
otherwise - N should be replaced by the returned Operand.
The new test case shows no difference with or without this patch, but it will crash if
we re-apply r237046 or enable FMF via the current -enable-fmf-dag cl::opt.
Differential Revision: http://reviews.llvm.org/D9893
llvm-svn: 241826
Summary:
This change is part of a series of commits dedicated to have a single
DataLayout during compilation by using always the one owned by the
module.
Reviewers: echristo
Subscribers: jholewinski, llvm-commits, rafael, yaron.keren
Differential Revision: http://reviews.llvm.org/D11040
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 241778
Summary:
This change is part of a series of commits dedicated to have a single
DataLayout during compilation by using always the one owned by the
module.
Reviewers: echristo
Subscribers: jholewinski, llvm-commits, rafael, yaron.keren
Differential Revision: http://reviews.llvm.org/D11037
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 241776
Summary:
This change is part of a series of commits dedicated to have a single
DataLayout during compilation by using always the one owned by the
module.
Reviewers: echristo
Subscribers: jholewinski, ted, yaron.keren, rafael, llvm-commits
Differential Revision: http://reviews.llvm.org/D11028
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 241775
Summary:
This change is part of a series of commits dedicated to have a single
DataLayout during compilation by using always the one owned by the
module.
Reviewers: echristo
Subscribers: llvm-commits, rafael, yaron.keren
Differential Revision: http://reviews.llvm.org/D11017
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 241655
Summary:
SelectionDAG itself is not invoking directly the DataLayout in the
TargetMachine, but the "TargetLowering" class is still using it. I'll
address it in a following commit.
This change is part of a series of commits dedicated to have a single
DataLayout during compilation by using always the one owned by the
module.
Reviewers: echristo
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11000
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 241618
The code responsible for shl folding in the DAGCombiner was assuming incorrectly that all constants are less than 64 bits. This patch simply changes the way values are compared.
It has been reverted previously because of some problems with comparing APInt with raw uint64_t. That has been fixed/changed with r241204.
llvm-svn: 241254
Summary: This patch fixes the cases of sext/zext constant folding in DAG combiner where constans do not fit 64 bits. The fix simply removes un$
Test Plan: New regression test included.
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: RKSimon, llvm-commits
Differential Revision: http://reviews.llvm.org/D10607
llvm-svn: 240991
We had a hack in SDAGBuilder in place to work around this but now we
can avoid that. Call BuildExactSDIV from BuildSDIV so DAGCombiner can
perform this trick automatically.
The added check in DAGCombiner is necessary to prevent exact sdiv by pow2
from regressing as the target-specific pow2 lowering is not aware of
exact bits yet.
This is mostly covered by existing tests. One side effect is that we
get the better lowering for exact vector sdivs now too :)
llvm-svn: 240891
Summary: The code responsible for shl folding in the DAGCombiner was assuming incorrectly that all constants are less than 64 bits. This patch simply changes the way values are compared.
Test Plan: A regression test included.
Reviewers: andreadb
Reviewed By: andreadb
Subscribers: andreadb, test, llvm-commits
Differential Revision: http://reviews.llvm.org/D10602
llvm-svn: 240291
This will allow classes to implement the AA interface without deriving
from the class or referencing an internal enum of some other class as
their return types.
Also, to a pretty fundamental extent, concepts such as 'NoAlias',
'MayAlias', and 'MustAlias' are first class concepts in LLVM and we
aren't saving anything by scoping them heavily.
My mild preference would have been to use a scoped enum, but that
feature is essentially completely broken AFAICT. I'm extremely
disappointed. For example, we cannot through any reasonable[1] means
construct an enum class (or analog) which has scoped names but converts
to a boolean in order to test for the possibility of aliasing.
[1]: Richard Smith came up with a "solution", but it requires class
templates, and lots of boilerplate setting up the enumeration multiple
times. Something like Boost.PP could potentially bundle this up, but
even that would be quite painful and it doesn't seem realistically worth
it. The enum class solution would probably work without the need for
a bool conversion.
Differential Revision: http://reviews.llvm.org/D10495
llvm-svn: 240255
The patch is generated using this command:
tools/clang/tools/extra/clang-tidy/tool/run-clang-tidy.py -fix \
-checks=-*,llvm-namespace-comment -header-filter='llvm/.*|clang/.*' \
llvm/lib/
Thanks to Eugene Kosov for the original patch!
llvm-svn: 240137
that it is its own entity in the form of MemoryLocation, and update all
the callers.
This is an entirely mechanical change. References to "Location" within
AA subclases become "MemoryLocation", and elsewhere
"AliasAnalysis::Location" becomes "MemoryLocation". Hope that helps
out-of-tree folks update.
llvm-svn: 239885
This is an updated version of the patch that was checked in at:
http://reviews.llvm.org/rL237046
but subsequently reverted because it exposed a bug in the DAG Combiner:
http://reviews.llvm.org/D9893
This time, there's an enablement flag ("EnableFMFInDAG") around the code in
SelectionDAGBuilder where we copy the set of FP optimization flags from IR
instructions to DAG nodes. So, in theory, there should be no functional change
from this patch as-is, but it will allow testing with the added functionality
to proceed via "-enable-fmf-dag" passed to llc.
This patch adds the minimum plumbing necessary to use IR-level
fast-math-flags (FMF) in the backend without actually using
them for anything yet. This is a follow-on to:
http://reviews.llvm.org/rL235997
Differential Revision: http://reviews.llvm.org/D10403
llvm-svn: 239828
Reapply r239539. Don't assume the collected number of
stores is the same vector size. Just take the first N
stores to fill the vector.
llvm-svn: 239825
Now actually stores the non-zero constant instead of 0.
I somehow forgot to include this part of r238108.
The test change was just an independent instruction order swap,
so just add another check line to satisfy CHECK-NEXT.
llvm-svn: 239539
For targets with a free fneg, this fold is always a net loss if it
ends up duplicating the multiply, so definitely avoid it.
This might be true for some targets without a free fneg too, but
I'll leave that for future investigation.
llvm-svn: 239167
Also, moved test cases from CodeGen/X86/fold-buildvector-bug.ll into
CodeGen/X86/buildvec-insertvec.ll and regenerated CHECK lines using
update_llc_test_checks.py.
llvm-svn: 239142
Method 'visitBUILD_VECTOR' in the DAGCombiner knows how to combine a
build_vector of a bunch of extract_vector_elt nodes and constant zero nodes
into a shuffle blend with a zero vector.
However, method 'visitBUILD_VECTOR' forgot that a floating point
build_vector may contain negative zero as well as positive zero.
Example:
define <2 x double> @example(<2 x double> %A) {
entry:
%0 = extractelement <2 x double> %A, i32 0
%1 = insertelement <2 x double> undef, double %0, i32 0
%2 = insertelement <2 x double> %1, double -0.0, i32 1
ret <2 x double> %2
}
Before this patch, llc (with -mattr=+sse4.1) wrongly generated
movq %xmm0, %xmm0 # xmm0 = xmm0[0],zero
So, the sign bit of the negative zero was effectively lost.
This patch fixes the problem by adding explicit checks for positive zero.
With this patch, llc produces the following code for the example above:
movhpd .LCPI0_0(%rip), %xmm0
where .LCPI0_0 referes to a 'double -0'.
llvm-svn: 239070
On GPU targets, materializing constants is cheap and stores are
expensive, so only doing this for zero vectors was silly.
Most of the new testcases aren't optimally merged, and are for
later improvements.
llvm-svn: 238108
This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization).
It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner.
Differential Revision: http://reviews.llvm.org/D9848
llvm-svn: 237885
DAG.FoldConstantArithmetic() can fail even though both operands are
Constants if OpaqueConstants are involved. Continue trying other combine
possibilities in tis case.
Differential Revision: http://reviews.llvm.org/D6946
Somewhat related to PR21801 / rdar://19211454
llvm-svn: 237822
In CombineToPreIndexedLoadStore, when the offset is a constant, we have code
that looks for other uses of the pointer which are constant offset computations
so that they can be rewritten in terms of the updated pointer so that we don't
need to keep a copy of the base pointer to compute these constant offsets.
Unfortunately, when it iterated over the uses, it did so by SDNodes, and so we
could confuse ourselves if the base pointer was produced by a node that had
multiple results (because we would not immediately exclude uses of the other
node results). This was reported as PR22755. Unfortunately, we don't have a
test case (and I've also been unable to produce one thus far), but at least the
mistake is clear. The right way to fix this problem is to make use of the information
contained in the use iterators to filter out any uses of other results of the
node producing the base pointer.
This should be mostly NFC, but should also fix PR22755 (for which,
unfortunately, we have no in-tree test case).
llvm-svn: 237576
This is a less ambitious version of:
http://reviews.llvm.org/rL236546
because that was reverted in:
http://reviews.llvm.org/rL236600
because it caused memory corruption that wasn't related to FMF
but was actually due to making nodes with 2 operands derive from a
plain SDNode rather than a BinarySDNode.
This patch adds the minimum plumbing necessary to use IR-level
fast-math-flags (FMF) in the backend without actually using
them for anything yet. This is a follow-on to:
http://reviews.llvm.org/rL235997
...which split the existing nsw / nuw / exact flags and FMF
into their own struct.
llvm-svn: 237046
The bug showed up as a compile-time assertion failure:
Assertion `NumBits >= MIN_INT_BITS && "bitwidth too small"' failed
when building msan tests on x86-64.
Prior to r236850, this bug was masked due to a bogus alignment check,
which also accidentally rejected non-byte-sized accesses. Afterwards,
an invalid ElementSizeBytes == 0 got further into the function, and
triggered the assertion failure.
It would probably be a good idea to allow it to handle merging stores
of unusual widths as well, but for now, to un-break it, I'm just
making the minimal fix.
Differential Revision: http://reviews.llvm.org/D9626
llvm-svn: 236927
1) check whether the alignment of the memory is sufficient for the
*merged* store or load to be efficient.
Not doing so can result in some ridiculously poor code generation, if
merging creates a vector operation which must be aligned but isn't.
2) DON'T check that the alignment of each load/store is equal. If
you're merging 2 4-byte stores, the first *might* have 8-byte
alignment, but the second certainly will have 4-byte alignment. We do
want to allow those to be merged.
llvm-svn: 236850
This patch adds the minimum plumbing necessary to use IR-level
fast-math-flags (FMF) in the backend without actually using
them for anything yet. This is a follow-on to:
http://reviews.llvm.org/rL235997
...which split the existing nsw / nuw / exact flags and FMF
into their own struct.
There are 2 structural changes here:
1. The main diff is that we're preparing to extend the optimization
flags to affect more than just binary SDNodes. Eg, IR intrinsics
( https://llvm.org/bugs/show_bug.cgi?id=21290 ) or non-binop nodes
that don't even exist in IR such as FMA, FNEG, etc.
2. The other change is that we're actually copying the FP fast-math-flags
from the IR instructions to SDNodes.
Differential Revision: http://reviews.llvm.org/D8900
llvm-svn: 236546
This patch makes ReplaceExtractVectorEltOfLoadWithNarrowedLoad convert
the element number from getVectorIdxTy() to PtrTy before doing pointer
arithmetic on it. This is needed on z, where element numbers are i32
but pointers are i64.
Original patch by Richard Sandiford.
llvm-svn: 236530
For little-endian, the function would convert (extract_vector_elt (load X), Y)
to X + Y*sizeof(elt). For big-endian it would instead use
X + sizeof(vec) - Y*sizeof(elt). The big-endian case wasn't right since
vector index order always follows memory/array order, even for big-endian.
(Note that the current handling has to be wrong for Y==0 since it would
access beyond the end of the vector.)
Original patch by Richard Sandiford.
llvm-svn: 236529
At the least it should be guarded by some kind of target hook.
It also introduced catastrophic compile time and code quality
regressions on some out of tree targets (test case still being
reduced/sanitized).
Sanjay agreed with reverting this patch until these issues can be
resolved.
llvm-svn: 236199
This is a compromise: with this simple patch, we should always handle a chain of exactly 3
operations optimally, but we're not generating the optimal balanced binary tree for a longer
sequence.
In general, this transform will reduce the dependency chain for a sequence of instructions
using N operands from a worst case N-1 dependent operations to N/2 dependent operations.
The optimal balanced binary tree would reduce the chain to log2(N).
The trade-off for not dealing with longer sequences is: (1) we have less complexity in the
compiler, (2) we avoid unknown compile-time blowup calculating a balanced tree, and (3) we
don't need to worry about the increased register pressure required to parallelize longer
sequences. It also seems unlikely that we would ever encounter really long strings of
dependent ops like that in the wild, but I'm not sure how to verify that speculation.
FWIW, I see no perf difference for test-suite running on btver2 (x86-64) with -ffast-math
and this patch.
We can extend this patch to cover other associative operations such as fmul, fmax, fmin,
integer add, integer mul.
This is a partial fix for:
https://llvm.org/bugs/show_bug.cgi?id=17305
and if extended:
https://llvm.org/bugs/show_bug.cgi?id=21768https://llvm.org/bugs/show_bug.cgi?id=23116
The issue also came up in:
http://reviews.llvm.org/D8941
Differential Revision: http://reviews.llvm.org/D9232
llvm-svn: 236031
This is a preliminary step to using the IR-level floating-point fast-math-flags in the SDAG (D8900).
In this patch, we introduce the optimization flags as their own struct. As noted in the TODO comment,
we should eventually share this data between the IR passes and the backend.
We also switch the existing nsw / nuw / exact bit functionality of the BinaryWithFlagsSDNode class to
use the new struct.
The tradeoff is that instead of using the free but limited space of SDNode's SubclassData, we add a
data member to the subclass. This means we don't have to repeat all of the get/set methods per flag,
but we're potentially adding size to all nodes of this subclassi type.
In practice on 64-bit systems (measured on Linux and MacOS X), there is no size difference between an
SDNode and BinaryWithFlagsSDNode after this change: they're both 80 bytes. This means that we had at
least one free byte to play with due to struct alignment.
Differential Revision: http://reviews.llvm.org/D9325
llvm-svn: 235997
[DebugInfo] Add debug locations to constant SD nodes
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
Differential Revision: http://reviews.llvm.org/D9084
llvm-svn: 235989
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
Differential Revision: http://reviews.llvm.org/D9084
llvm-svn: 235977
right scaling.
In the function canFoldInAddressingMode, VT is computed as the type of the
destination/source of a LOAD/STORE operations, instead of the memory type of the
operation.
On targets with a scaling factor on the offset of the LOAD/STORE operations, the
function may return false for actually valid cases. This may then prevent the
selection of profitable pre or post indexed load/store operations, and instead
select pre or post indexed load/store for unprofitable cases.
Patch by Francois de Ferriere <francois.de-ferriere@st.com>!
Differential Revision: http://reviews.llvm.org/D9146
llvm-svn: 235780
Patch to remove extra bitcasts from shuffles, this is often a legacy of XformToShuffleWithZero being used to combine bitmaskings (of float vectors bitcast to integer vectors) into shuffles: bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1)
Differential Revision: http://reviews.llvm.org/D9097
llvm-svn: 235578
This turned up after r235333, but was a pre-existing bug. The optimization
which transforms select(c, load, load) into a load of a select of the addresses
does not handle indexed loads (pre/post inc/dec). However, it did not check for
them either, leading to a crash if it tried to transform one of them.
llvm-svn: 235497
Summary:
This patch adds legalization support to operate on FP16 as a load/store type
and do operations on it as floats.
Tests for ARM are added to test/CodeGen/ARM/fp16-promote.ll
Reviewers: srhines, t.p.northover
Differential Revision: http://reviews.llvm.org/D8755
llvm-svn: 235215
The only type that isn't an integer, isn't floating point, and isn't
a vector; ladies and gentlemen, the gift that keeps on giving: x86_mmx!
Fixes PR23246.
Original message (reverted in r235062):
[CodeGen] Combine concat_vectors of scalars into build_vector.
Combine something like:
(v8i8 concat_vectors (v2i8 bitcast (i16)) x4)
into:
(v8i8 (bitcast (v4i16 BUILD_VECTOR (i16) x4)))
If any of the scalars are floating point, use that throughout.
Differential Revision: http://reviews.llvm.org/D8948
llvm-svn: 235072
Combine something like:
(v8i8 concat_vectors (v2i8 bitcast (i16)) x4)
into:
(v8i8 (bitcast (v4i16 BUILD_VECTOR (i16) x4)))
If any of the scalars are floating point, use that throughout.
Differential Revision: http://reviews.llvm.org/D8948
llvm-svn: 234809
In case of different types used for the condition of the selects the
select(select) -> select(and) normalisation cannot be performed.
See also: http://reviews.llvm.org/D7622
llvm-svn: 234763
We already do:
concat_vectors(scalar, undef) -> scalar_to_vector(scalar)
When the scalar is legal.
When it's not, but is a truncated legal scalar, we can also do:
concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
Which is equivalent, since the upper lanes are undef anyway.
While there, teach the combine to look at more than 2 operands.
Differential Revision: http://reviews.llvm.org/D8883
llvm-svn: 234530
The bug manifests when there are two loads and two stores chained as follows in
a DAG,
(ld v3f32) -> (st f32) -> (ld v3f32) -> (st f32)
and the stores' values are extracted from the preceding vector loads.
MergeConsecutiveStores would replace the first store in the chain with the
merged vector store, which would create a cycle between the merged store node
and the last load node that appears in the chain.
This commits fixes the bug by replacing the last store in the chain instead.
rdar://problem/20275084
Differential Revision: http://reviews.llvm.org/D8849
llvm-svn: 234430
This patch attempts to fold the shuffling of 'scalar source' inputs - BUILD_VECTOR and SCALAR_TO_VECTOR nodes - if the shuffle node is the only user. This folds away a lot of unnecessary shuffle nodes, and allows quite a bit of constant folding that was being missed.
Differential Revision: http://reviews.llvm.org/D8516
llvm-svn: 234004
DAGCombiner::ReassociateOps was correctly testing for an constant integer scalar but failed to correctly test for constant integer vectors (it was testing for any constant vector).
llvm-svn: 233482
This patch adds supports for the vector constant folding of TRUNCATE and FP_EXTEND instructions and tidies up the SINT_TO_FP and UINT_TO_FP instructions to match.
It also moves the vector constant folding for the FNEG and FABS instructions to use the DAG.getNode() functionality like the other unary instructions.
Differential Revision: http://reviews.llvm.org/D8593
llvm-svn: 233224
Reverts the code change from r221168 and the relevant test.
It was a mistake to disable the combiner, and based on the ultimate
definition of 'optnone' we shouldn't have considered the test case
as failing in the first place.
llvm-svn: 233153
This is very related to the bug fixed in r174431. The problem is that
SelectionDAG does not include alignment in the uniquing of loads and
stores. When an otherwise no-op DAGCombine would increase the alignment
of a load or store, the original node would be returned (with the
alignment increased), which would cause the node not to be processed by
any further DAGCombines.
I don't have a direct testcase for this that manifests on an in-tree
target, but I did see some noise in the tests for other targets and have
updated them for it.
llvm-svn: 232780
Targets which provide a rotate make it possible to replace a sequence of
(XOR (SHL 1, x), -1) with (ROTL ~1, x). This saves an instruction on
architectures like X86 and POWER(64).
Differential Revision: http://reviews.llvm.org/D8350
llvm-svn: 232572
We have an increasing number of cases where we are creating commuted shuffle masks - all implementing nearly the same code.
This patch adds a static helper function - ShuffleVectorSDNode::commuteMask() and replaces a number of cases to use it.
Differential Revision: http://reviews.llvm.org/D8139
llvm-svn: 231581
This patch fixes the logic in the DAGCombiner that folds an AND node according
to rule: (and (X (load V)), C) -> (X (load V))
An AND between a vector load 'X' and a constant build_vector 'C' can be folded
into the load itself only if we can prove that the AND operation is redundant.
The algorithm implemented by 'visitAND' firstly computes the splat value 'S'
from C, and then checks if S has the lower 'B' bits set (where B is the size in
bits of the vector element type). The algorithm takes into account also the
'undef' bits in the splat mask.
Unfortunately, the algorithm only worked under the assumption that the size of S
is a multiple of the vector element type. With this patch, we conservatively
avoid folding the AND if the splat bits are not compatible with the vector
element type.
Added X86 test and-load-fold.ll
Differential Revision: http://reviews.llvm.org/D8085
llvm-svn: 231563
This patch attempts to convert a SCALAR_TO_VECTOR using an operand from an EXTRACT_VECTOR_ELT into a VECTOR_SHUFFLE.
This prevents many cases of spilling scalar data between the gpr + simd registers.
At present the optimization only accepts cases where there is no TRUNC of the scalar type (i.e. all types must match).
Differential Revision: http://reviews.llvm.org/D8132
llvm-svn: 231554
This is based on the following equivalences:
select(C0 & C1, X, Y) <=> select(C0, select(C1, X, Y), Y)
select(C0 | C1, X, Y) <=> select(C0, X, select(C1, X, Y))
Many target cannot perform and/or on the CPU flags and therefore the
right side should be choosen to avoid materializign the i1 flags in an
integer register. If the target can perform this operation efficiently
we normalize to the left form.
Differential Revision: http://reviews.llvm.org/D7622
llvm-svn: 231507
This is in preparation for changing visitSELECT to normalize towards
select(Cond0, select(Cond1, X, Y), Y);
select(Cond0, X, select(Cond1, X, Y)) which perfom an implicit and/or of
the conditions.
The factored function contains all DAGCombine rules which reduce two values
combined by an And/Or operation to a single value. This does not include rules
involving constants as visitSELECT already handles that case.
Differential Revision: http://reviews.llvm.org/D8026
llvm-svn: 231506
Currently shuffles may only be combined if they are of the same type, despite the fact that bitcasts are often introduced in between shuffle nodes (e.g. x86 shuffle type widening).
This patch allows a single input shuffle to peek through bitcasts and if the input is another shuffle will merge them, shuffling using the smallest sized type, and re-applying the bitcasts at the inputs and output instead.
Dropped old ShuffleToZext test - this patch removes the use of the zext and vector-zext.ll covers these anyhow.
Differential Revision: http://reviews.llvm.org/D7939
llvm-svn: 231380
When trying to convert a BUILD_VECTOR into a shuffle, we try to split a single source vector that is twice as wide as the destination vector.
We can not do this when we also need the zero vector to create a blend.
This fixes PR22774.
Differential Revision: http://reviews.llvm.org/D8040
llvm-svn: 231219
Accidentally committed a few more of these cleanup changes than
intended. Still breaking these out & tidying them up.
This reverts commit r231135.
llvm-svn: 231136
There doesn't seem to be any need to assert that iterator assignment is
between iterators over the same node - if you want to reuse an iterator
variable to iterate another node, that's perfectly acceptable. Just
don't mix comparisons between iterators into disjoint sequences, as
usual.
llvm-svn: 231135
We were missing a check for the following fold in DAGCombiner:
// fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
If 'x' is also a constant, then we shouldn't do anything. Otherwise, we could end up swapping the operands back and forth forever.
This should fix:
http://llvm.org/bugs/show_bug.cgi?id=22698
Differential Revision: http://reviews.llvm.org/D7917
llvm-svn: 230884
have the debugger step through each one individually. Turn off the
combine for adjacent stores at -O0 so we get this behavior.
Possibly, DAGCombine shouldn't run at all at -O0, but that's for
another day; see PR22346.
Differential Revision: http://reviews.llvm.org/D7181
llvm-svn: 230659
Author: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Mon Feb 23 23:04:28 2015 +0000
Fix based on post-commit comment on D7816 & rL230177 - BUILD_VECTOR operand truncation was using the the BV's output scalar type instead of the input type.
and
Author: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sun Feb 22 18:17:28 2015 +0000
[DagCombiner] Generalized BuildVector Vector Concatenation
The CONCAT_VECTORS combiner pass can transform the concat of two BUILD_VECTOR nodes into a single BUILD_VECTOR node.
This patch generalises this to support any number of BUILD_VECTOR nodes, and also permits UNDEF nodes to be included as well.
This was noticed as AVX vec128 -> vec256 canonicalization sometimes creates a CONCAT_VECTOR with a real vec128 lower and an vec128 UNDEF upper.
Differential Revision: http://reviews.llvm.org/D7816
as the root cause of PR22678 which is causing an assertion inside the DAG combiner.
I'll follow up to the main thread as well.
llvm-svn: 230358
The CONCAT_VECTORS combiner pass can transform the concat of two BUILD_VECTOR nodes into a single BUILD_VECTOR node.
This patch generalises this to support any number of BUILD_VECTOR nodes, and also permits UNDEF nodes to be included as well.
This was noticed as AVX vec128 -> vec256 canonicalization sometimes creates a CONCAT_VECTOR with a real vec128 lower and an vec128 UNDEF upper.
Differential Revision: http://reviews.llvm.org/D7816
llvm-svn: 230177
DAGCombine will rewrite an BUILD_VECTOR where all non-undef inputs some from
[US]INT_TO_FP, as a BUILD_VECTOR of integers with the conversion applied as a
vector operation. We check operation legality of the conversion, but fail to
check legality of the integer vector type itself. Because targets don't
normally override operation legality defaults for illegal types, we need to
check this also.
This came up in the context of the QPX vector entensions for PowerPC (which can
have legal floating-point vector types without corresponding legal integer
vector types). No in-tree test case for this yes, but one can be added once
the QPX support has been committed.
llvm-svn: 230176
This allows sharing of FMA forming combines to work
with instructions that have the same semantics as a separate
multiply and add.
This is expand by default, and only formed post legalization
so it shouldn't have much impact on targets that do not want it.
llvm-svn: 230070
First, don't combine bit masking into vector shuffles (even ones the
target can handle) once operation legalization has taken place. Custom
legalization of vector shuffles may exist for these patterns (making the
predicate return true) but that custom legalization may in some cases
produce the exact bit math this matches. We only really want to handle
this prior to operation legalization.
However, the x86 backend, in a fit of awesome, relied on this. What it
would do is mark VSELECTs as expand, which would turn them into
arithmetic, which this would then match back into vector shuffles, which
we would then lower properly. Amazing.
Instead, the second change is to teach the x86 backend to directly form
vector shuffles from VSELECT nodes with constant conditions, and to mark
all of the vector types we support lowering blends as shuffles as custom
VSELECT lowering. We still mark the forms which actually support
variable blends as *legal* so that the custom lowering is bypassed, and
the legal lowering can even be used by the vector shuffle legalization
(yes, i know, this is confusing. but that's how the patterns are
written).
This makes the VSELECT lowering much more sensible, and in fact should
fix a bunch of bugs with it. However, as you'll see in the test cases,
right now what it does is point out the *hilarious* deficiency of the
new vector shuffle lowering when it comes to blends. Fortunately, my
very next patch fixes that. I can't submit it yet, because that patch,
somewhat obviously, forms the exact and/or pattern that the DAG combine
is matching here! Without this patch, teaching the vector shuffle
lowering to produce the right code infloops in the DAG combiner. With
this patch alone, we produce terrible code but at least lower through
the right paths. With both patches, all the regressions here should be
fixed, and a bunch of the improvements (like using 2 shufps with no
memory loads instead of 2 andps with memory loads and an orps) will
stay. Win!
There is one other change worth noting here. We had hilariously wrong
vectorization cost estimates for vselect because we fell through to the
code path that assumed all "expand" vector operations are scalarized.
However, the "expand" lowering of VSELECT is vector bit math, most
definitely not scalarized. So now we go back to the correct if horribly
naive cost of "1" for "not scalarized". If anyone wants to add actual
modeling of shuffle costs, that would be cool, but this seems an
improvement on its own. Note the removal of 16 and 32 "costs" for doing
a blend. Even in SSE2 we can blend in fewer than 16 instructions. ;] Of
course, we don't right now because of OMG bad code, but I'm going to fix
that. Next patch. I promise.
llvm-svn: 229835
This is a follow-on patch to:
http://reviews.llvm.org/D7093
That patch canonicalized constant splats as build_vectors,
and this patch removes the constant check so we can canonicalize
all splats as build_vectors.
This fixes the 2nd test case in PR22283:
http://llvm.org/bugs/show_bug.cgi?id=22283
The unfortunate code duplication between SelectionDAG and DAGCombiner
is discussed in the earlier patch review. At least this patch is just
removing code...
This improves an existing x86 AVX test and changes codegen in an ARM test.
Differential Revision: http://reviews.llvm.org/D7389
llvm-svn: 229511
test.
This was just a matter of the DAG combine for vector shuffles being too
aggressive. This is a bit of a grey area, but I think generally if we
can re-use intermediate shuffles, we should. Certainly, given the test
cases I have available, this seems like the right call.
llvm-svn: 229285
Canonicalize access to function attributes to use the simpler API.
getAttributes().getAttribute(AttributeSet::FunctionIndex, Kind)
=> getFnAttribute(Kind)
getAttributes().hasAttribute(AttributeSet::FunctionIndex, Kind)
=> hasFnAttribute(Kind)
Also, add `Function::getFnStackAlignment()`, and canonicalize:
getAttributes().getStackAlignment(AttributeSet::FunctionIndex)
=> getFnStackAlignment()
llvm-svn: 229208
We used to do this DAG combine, but it's not always correct:
If the first fp_round isn't a value preserving truncation, it might
introduce a tie in the second fp_round, that wouldn't occur in the
single-step fp_round we want to fold to.
In other words, double rounding isn't the same as rounding.
Differential Revision: http://reviews.llvm.org/D7571
llvm-svn: 228911
Add new token factor node and its users to worklist if alias analysis is
turned on, in DAGCombiner::visitTokenFactor(). Alias analysis may cause
a lot of new token factors to be inserted into the DAG, and they need to
be optimized to avoid significant slow-downs.
Reviewed by Hal Finkel.
llvm-svn: 228841
nodes when folding bitcasts of constants.
We can't fold things and then check after-the-fact whether it was legal.
Once we have formed the DAG node, arbitrary other nodes may have been
collapsed to it. There is no easy way to go back. Instead, we need to
test for the specific folding cases we're interested in and ensure those
are legal first.
This could in theory make this less powerful for bitcasting from an
integer to some vector type, but AFAICT, that can't actually happen in
the SDAG so its fine. Now, we *only* whitelist specific int->fp and
fp->int bitcasts for post-legalization folding. I've added the test case
from the PR.
(Also as a note, this does not appear to be in 3.6, no backport needed)
llvm-svn: 228656
The combine that forms extloads used to be disabled on vector types,
because "None of the supported targets knows how to perform load and
sign extend on vectors in one instruction."
That's not entirely true, since at least SSE4.1 X86 knows how to do
those sextloads/zextloads (with PMOVS/ZX).
But there are several aspects to getting this right.
First, vector extloads are controlled by a profitability callback.
For instance, on ARM, several instructions have folded extload forms,
so it's not always beneficial to create an extload node (and trying to
match extloads is a whole 'nother can of worms).
The interesting optimization enables folding of s/zextloads to illegal
(splittable) vector types, expanding them into smaller legal extloads.
It's not ideal (it introduces some legalization-like behavior in the
combine) but it's better than the obvious alternative: form illegal
extloads, and later try to split them up. If you do that, you might
generate extloads that can't be split up, but have a valid ext+load
expansion. At vector-op legalization time, it's too late to generate
this kind of code, so you end up forced to scalarize. It's better to
just avoid creating egregiously illegal nodes.
This optimization is enabled unconditionally on X86.
Note that the splitting combine is happy with "custom" extloads. As
is, this bypasses the actual custom lowering, and just unrolls the
extload. But from what I've seen, this is still much better than the
current custom lowering, which does some kind of unrolling at the end
anyway (see for instance load_sext_4i8_to_4i64 on SSE2, and the added
FIXME).
Also note that the existing combine that forms extloads is now also
enabled on legal vectors. This doesn't have a big effect on X86
(because sext+load is usually combined to sext_inreg+aextload).
On ARM it fires on some rare occasions; that's for a separate commit.
Differential Revision: http://reviews.llvm.org/D6904
llvm-svn: 228325
This commit creates infinite loop in DAG combine for in the LLVM test-suite
for aarch64 with mcpu=cylcone (just having neon may be enough to expose this).
llvm-svn: 227272
This patch resolves part of PR21711 ( http://llvm.org/bugs/show_bug.cgi?id=21711 ).
The 'f3' test case in that report presents a situation where we have two 128-bit
stores extracted from a 256-bit source vector.
Instead of producing this:
vmovaps %xmm0, (%rdi)
vextractf128 $1, %ymm0, 16(%rdi)
This patch merges the 128-bit stores into a single 256-bit store:
vmovups %ymm0, (%rdi)
Differential Revision: http://reviews.llvm.org/D7208
llvm-svn: 227242
This is a 2nd try at the same optimization as http://reviews.llvm.org/D6698.
That patch was checked in at r224611, but reverted at r225031 because it
caused a failure outside of the regression tests.
The cause of the crash was not recognizing consecutive stores that have mixed
source values (loads and vector element extracts), so this patch adds a check
to bail out if any store value is not coming from a vector element extract.
This patch also refactors the shared logic of the constant source and vector
extracted elements source cases into a helper function.
Differential Revision: http://reviews.llvm.org/D6850
llvm-svn: 226845
This solves PR22276.
Splats of constants would sometimes produce redundant shuffles, sometimes ridiculously so (see the PR for details). Fold these shuffles into BUILD_VECTORs early on instead.
Differential Revision: http://reviews.llvm.org/D7093
Fixed recommit of r226811.
llvm-svn: 226816
This solves PR22276.
Splats of constants would sometimes produce redundant shuffles, sometimes ridiculously so (see the PR for details). Fold these shuffles into BUILD_VECTORs early on instead.
Differential Revision: http://reviews.llvm.org/D7093
llvm-svn: 226811
The problem occurs when after vectorization we have type
<2 x i32>. This type is promoted to <2 x i64> and then requires
additional efforts for expanding loads and truncating stores.
I added EXPAND / TRUNCATE attributes to the masked load/store
SDNodes. The code now contains additional shuffles.
I've prepared changes in the cost estimation for masked memory
operations, it will be submitted separately.
llvm-svn: 226808
Type MVT::i1 became legal in KNL, but store operation can't be narrowed to this type,
since the size of VT (1 bit) is not equal to its actual store size(8 bits).
Added a test provided by David (dag@cray.com)
llvm-svn: 226805
Loading 2 2x32-bit float vectors into the bottom half of a 256-bit vector
produced suboptimal code in AVX2 mode with certain IR combinations.
In particular, the IR optimizer folded 2f32 + 2f32 -> 4f32, 4f32 + 4f32
(undef) -> 8f32 into a 2f32 + 2f32 -> 8f32, which seems more canonical,
but then mysteriously generated rather bad code; the movq/movhpd combination
didn't match.
The problem lay in the BUILD_VECTOR optimization path. The 2f32 inputs
would get promoted to 4f32 by the type legalizer, eventually resulting
in a BUILD_VECTOR on two 4f32 into an 8f32. The BUILD_VECTOR then, recognizing
these were both half the output size, concatted them and then produced
a shuffle. However, the resulting concat + shuffle was more complex than
it should be; in the case where the upper half of the output is undef, we
probably want to generate shuffle + concat instead.
This enhancement causes the vector_shuffle combine step to recognize this
suboptimal pattern and correct it. I included it there instead of in BUILD_VECTOR
in case the same suboptimal pattern occurs for other reasons.
This results in the optimizer correctly producing the optimal movq + movhpd
sequence for all three variations on this IR, even with AVX2.
I've included a test case.
Radar link: rdar://problem/19287012
Fix for PR 21943.
From: Fiona Glaser <fglaser@apple.com>
llvm-svn: 226360
utils/sort_includes.py.
I clearly haven't done this in a while, so more changed than usual. This
even uncovered a missing include from the InstrProf library that I've
added. No functionality changed here, just mechanical cleanup of the
include order.
llvm-svn: 225974
In case folding a node end up with a NaN as operand for the select,
the folding of the condition of the selectcc node returns "UNDEF".
Differential Revision: http://reviews.llvm.org/D6889
llvm-svn: 225952
This requires a new hook to prevent expanding sqrt in terms
of rsqrt and reciprocal. v_rcp_f32, v_rsq_f32, and v_sqrt_f32 are
all the same rate, so this expansion would just double the number
of instructions and cycles.
llvm-svn: 225828
As pointed out by Aditya (and Owen), when we elide an FP extend to form an FMA,
we need to extend the incoming operands so that the resulting node will really
be legal. This is currently enabled only for PowerPC, and it happens to work
there regardless, but this should fix the functionality for everyone else
should anyone else wish to use it.
llvm-svn: 225492
As pointed out by Aditya (and Owen), there are two things wrong with this code.
First, it adds patterns which elide FP extends when forming FMAs, and that might
not be profitable on all targets (it belongs behind the pre-existing
aggressive-FMA-formation flag). This is fixed by this change.
Second, the resulting nodes might have operands of different types (the
extensions need to be re-added). That will be fixed in the follow-up commit.
llvm-svn: 225485
type (in addition to the memory type).
The *LoadExt* legalization handling used to only have one type, the
memory type. This forced users to assume that as long as the extload
for the memory type was declared legal, and the result type was legal,
the whole extload was legal.
However, this isn't always the case. For instance, on X86, with AVX,
this is legal:
v4i32 load, zext from v4i8
but this isn't:
v4i64 load, zext from v4i8
Whereas v4i64 is (arguably) legal, even without AVX2.
Note that the same thing was done a while ago for truncstores (r46140),
but I assume no one needed it yet for extloads, so here we go.
Calls to getLoadExtAction were changed to add the value type, found
manually in the surrounding code.
Calls to setLoadExtAction were mechanically changed, by wrapping the
call in a loop, to match previous behavior. The loop iterates over
the MVT subrange corresponding to the memory type (FP vectors, etc...).
I also pulled neighboring setTruncStoreActions into some of the loops;
those shouldn't make a difference, as the additional types are illegal.
(e.g., i128->i1 truncstores on PPC.)
No functional change intended.
Differential Revision: http://reviews.llvm.org/D6532
llvm-svn: 225421
Right now in DAG Combine check the validity of the returned type
only when -debug is given on the command line. However usually
the test cases in the validation does not use -debug.
An Assert build should always check this.
llvm-svn: 224779
This handles the case of a BUILD_VECTOR being constructed out of elements extracted from a vector twice the size of the result vector. Previously this was always scalarized. Now, we try to construct a shuffle node that feeds on extract_subvectors.
This fixes PR15872 and provides a partial fix for PR21711.
Differential Revision: http://reviews.llvm.org/D6678
llvm-svn: 224429
Add an option to disable optimization to shrink truncated larger type
loads to smaller type loads. On SI this prevents using scalar load
instructions in some cases, since there are no scalar extloads.
llvm-svn: 224084
Added instcombine optimizations for BSWAP with AND/OR/XOR ops:
OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
Since its just a one liner, I've also added BSWAP to the DAGCombiner equivalent as well:
fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
Refactored bswap-fold tests to use FileCheck instead of just checking that the bswaps had gone.
Differential Revision: http://reviews.llvm.org/D6407
llvm-svn: 223349
I'm recommiting the codegen part of the patch.
The vectorizer part will be send to review again.
Masked Vector Load and Store Intrinsics.
Introduced new target-independent intrinsics in order to support masked vector loads and stores. The loop vectorizer optimizes loops containing conditional memory accesses by generating these intrinsics for existing targets AVX2 and AVX-512. The vectorizer asks the target about availability of masked vector loads and stores.
Added SDNodes for masked operations and lowering patterns for X86 code generator.
Examples:
<16 x i32> @llvm.masked.load.v16i32(i8* %addr, <16 x i32> %passthru, i32 4 /* align */, <16 x i1> %mask)
declare void @llvm.masked.store.v8f64(i8* %addr, <8 x double> %value, i32 4, <8 x i1> %mask)
Scalarizer for other targets (not AVX2/AVX-512) will be done in a separate patch.
http://reviews.llvm.org/D6191
llvm-svn: 223348
This reverts commit r222632 (and follow-up r222636), which caused a host
of LNT failures on an internal bot. I'll respond to the commit on the
list with a reproduction of one of the failures.
Conflicts:
lib/Target/X86/X86TargetTransformInfo.cpp
llvm-svn: 222936
Introduced new target-independent intrinsics in order to support masked vector loads and stores. The loop vectorizer optimizes loops containing conditional memory accesses by generating these intrinsics for existing targets AVX2 and AVX-512. The vectorizer asks the target about availability of masked vector loads and stores.
Added SDNodes for masked operations and lowering patterns for X86 code generator.
Examples:
<16 x i32> @llvm.masked.load.v16i32(i8* %addr, <16 x i32> %passthru, i32 4 /* align */, <16 x i1> %mask)
declare void @llvm.masked.store.v8f64(i8* %addr, <8 x double> %value, i32 4, <8 x i1> %mask)
Scalarizer for other targets (not AVX2/AVX-512) will be done in a separate patch.
http://reviews.llvm.org/D6191
llvm-svn: 222632
Before this patch, the DAGCombiner only tried to convert build_vector dag nodes
into shuffles if all operands were either extract_vector_elt or undef.
This patch improves that logic and teaches the DAGCombiner how to deal with
build_vector dag nodes where one or more operands are zero. A build_vector
dag node with some zero operands is turned into a shuffle only if the resulting
shuffle mask is legal for the target.
llvm-svn: 222536
This patch simplifies the logic that combines a pair of shuffle nodes into
a single shuffle if there is a legal mask. Also added comments to better
describe the algorithm. No functional change intended.
llvm-svn: 222522
E.g., ( a / D; b / D ) -> ( recip = 1.0 / D; a * recip; b * recip)
A hook is added to allow the target to control whether it needs to do such combine.
Reviewed in http://reviews.llvm.org/D6334
llvm-svn: 222510
This is to be consistent with StringSet and ultimately with the standard
library's associative container insert function.
This lead to updating SmallSet::insert to return pair<iterator, bool>,
and then to update SmallPtrSet::insert to return pair<iterator, bool>,
and then to update all the existing users of those functions...
llvm-svn: 222334
Some optimisations in DAGCombiner cause miscompilations for targets that use
TargetLowering::UndefinedBooleanContent, because they assume that the results
of a SELECT_CC node are boolean values, and can be safely ANDed, ORed and
XORed. These optimisations are only valid for targets that use
ZeroOrOneBooleanContent or ZeroOrNegativeOneBooleanContent.
This is a follow-up to D6210/r221693.
llvm-svn: 222123
This patch teaches the DAGCombiner how to combine shuffles according to rules:
shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(B, A, M2)
shuffle(shuffle(A, B, M0), B, M1) -> shuffle(B, A, M2)
shuffle(shuffle(A, B, M0), A, M1) -> shuffle(B, A, M2)
llvm-svn: 222090
LLVM replaces the SelectionDAG pattern (xor (set_cc cc x y) 1) with
(set_cc !cc x y), which is only correct when the xor has type i1.
Instead, we should check that the constant operand to the xor is all
ones.
llvm-svn: 221693
This patch improves the folding of vector AND nodes into blend operations for
targets that feature SSE4.1. A vector AND node where one of the operands is
a constant build_vector with elements that are either zero or all-ones can be
converted into a blend.
This allows for example to simplify the following code:
define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
%1 = and <4 x i32> %A, <i32 0, i32 0, i32 0, i32 -1>
%2 = and <4 x i32> %B, <i32 -1, i32 -1, i32 -1, i32 0>
%3 = or <4 x i32> %1, %2
ret <4 x i32> %3
}
Before this patch llc (-mcpu=corei7) generated:
andps LCPI1_0(%rip), %xmm0, %xmm0
andps LCPI1_1(%rip), %xmm1, %xmm1
orps %xmm1, %xmm0, %xmm0
retq
With this patch we generate a single 'vpblendw'.
llvm-svn: 221343
call DAGCombiner. But we ran into a case (on Windows) where the
calling convention causes argument lowering to bail out of fast-isel,
and we end up in CodeGenAndEmitDAG() which does run DAGCombiner.
So, we need to make DAGCombiner check for 'optnone' after all.
Commit includes the test that found this, plus another one that got
missed in the original optnone work.
llvm-svn: 221168
Earlier this summer I fixed an issue where we were incorrectly combining
multiple loads that had different constraints such alignment, invariance,
temporality, etc. Apparently in one case I made copt paste error and swapped
alignment and invariance.
Tests included.
rdar://18816719
llvm-svn: 220933
This is a first step for generating SSE rsqrt instructions for
reciprocal square root calcs when fast-math is allowed.
For now, be conservative and only enable this for AMD btver2
where performance improves significantly - for example, 29%
on llvm/projects/test-suite/SingleSource/Benchmarks/BenchmarkGame/n-body.c
(if we convert the data type to single-precision float).
This patch adds a two constant version of the Newton-Raphson
refinement algorithm to DAGCombiner that can be selected by any target
via a parameter returned by getRsqrtEstimate()..
See PR20900 for more details:
http://llvm.org/bugs/show_bug.cgi?id=20900
Differential Revision: http://reviews.llvm.org/D5658
llvm-svn: 220570
v2: use dyn_cast
fixup comments
v3: use cast
Reviewed-by: Matt Arsenault <arsenm2@gmail.com>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 220044
This patch changes the fast-math implementation for calculating sqrt(x) from:
y = 1 / (1 / sqrt(x))
to:
y = x * (1 / sqrt(x))
This has 2 benefits: less code / faster code and one less estimate instruction
that may lose precision.
The only target that will be affected (until http://reviews.llvm.org/D5658 is approved)
is PPC. The difference in codegen for PPC is 2 less flops for a single-precision sqrtf
or vector sqrtf and 4 less flops for a double-precision sqrt.
We also eliminate a constant load and extra register usage.
Differential Revision: http://reviews.llvm.org/D5682
llvm-svn: 219445
The patch's author points out that, despite the function's documentation,
getSetCCResultType is only used to get the SETCC result type (with one
here-removed problematic exception). In one case, getSetCCResultType was being
used to get the predicate type to use for a SELECT node, and then
SIGN_EXTENDing (or truncating) to get the input predicate to match that type.
Unfortunately, this was happening inside visitSIGN_EXTEND, and creating new
SIGN_EXTEND nodes was causing an infinite loop. In addition, this behavior was
wrong if a target was not using ZeroOrNegativeOneBooleanContent. Lastly, the
extension/truncation seems unnecessary here: SELECT is defined as:
Select(COND, TRUEVAL, FALSEVAL). If the type of the boolean COND is not i1
then the high bits must conform to getBooleanContents.
So here we remove this use of getSetCCResultType and update
getSetCCResultType's documentation to reflect its actual uses.
Patch by deadal nix!
llvm-svn: 219141
that are unused.
This allows the combiner to delete math feeding shuffles where the math
isn't actually necessary. This improves some of the vperm2x128 tests
that regressed when the vector shuffle lowering started actually
generating vperm instructions rather than forcibly decomposing them.
Sadly, this isn't enough to get this *really* right because we still
form a completely unnecessary permutation. To fix that, we also need to
fold shuffles which just rearrange concatenated or inserted subvectors.
llvm-svn: 219086
It was hacky to use an opcode as a switch because it won't always match
(rsqrte != sqrte), and it looks like we'll need to add more special casing
per arch than I had hoped for. Eg, x86 will prefer a different NR estimate
implementation. ARM will want to use it's 'step' instructions. There also
don't appear to be any new estimate instructions in any arch in a long,
long time. Altivec vloge and vexpte may have been the first and last in
that field...
llvm-svn: 218698
Currently, the DAG Combiner only tries to convert type-legal build_vector nodes
into shuffles. This patch simply moves the logic that checks if a
build_vector has a legal value type up before we even start analyzing the
operands. This allows to early exit immediately from method
'visitBUILD_VECTOR' if the node type is known to be illegal.
No functional change intended.
llvm-svn: 218677
If there is a store followed by a store with the same value to the same location, then the store is dead/noop. It can be removed.
This problem is found in spec2006-197.parser.
For example,
stur w10, [x11, #-4]
stur w10, [x11, #-4]
Then one of the two stur instructions can be removed.
Patch by David Xu!
llvm-svn: 218569
This is purely refactoring. No functional changes intended. PowerPC is the only target
that is currently using this interface.
The ultimate goal is to allow targets other than PowerPC (certainly X86 and Aarch64) to turn this:
z = y / sqrt(x)
into:
z = y * rsqrte(x)
And:
z = y / x
into:
z = y * rcpe(x)
using whatever HW magic they can use. See http://llvm.org/bugs/show_bug.cgi?id=20900 .
There is one hook in TargetLowering to get the target-specific opcode for an estimate instruction
along with the number of refinement steps needed to make the estimate usable.
Differential Revision: http://reviews.llvm.org/D5484
llvm-svn: 218553
This is purely a plumbing patch. No functional changes intended.
The ultimate goal is to allow targets other than PowerPC (certainly X86 and Aarch64) to turn this:
z = y / sqrt(x)
into:
z = y * rsqrte(x)
using whatever HW magic they can use. See http://llvm.org/bugs/show_bug.cgi?id=20900 .
The first step is to add a target hook for RSQRTE, take the already target-independent code selfishly hoarded by PPC, and put it into DAGCombiner.
Next steps:
The code in DAGCombiner::BuildRSQRTE() should be refactored further; tests that exercise that logic need to be added.
Logic in PPCTargetLowering::BuildRSQRTE() should be hoisted into DAGCombiner.
X86 and AArch64 overrides for TargetLowering.BuildRSQRTE() should be added.
Differential Revision: http://reviews.llvm.org/D5425
llvm-svn: 218219
The heuristic used by DAGCombine to form FMAs checks that the FMUL has only one
use, but this is overly-conservative on some systems. Specifically, if the FMA
and the FADD have the same latency (and the FMA does not compete for resources
with the FMUL any more than the FADD does), there is no need for the
restriction, and furthermore, forming the FMA leaving the FMUL can still allow
for higher overall throughput and decreased critical-path length.
Here we add a new TLI callback, enableAggressiveFMAFusion, false by default, to
elide the hasOneUse check. This is enabled for PowerPC by default, as most
PowerPC systems will benefit.
Patch by Olivier Sallenave, thanks!
llvm-svn: 218120
Do
(shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
This is already done for multiplies, but since multiplies
by powers of two are turned into shifts, we also need
to handle it here.
This might want checks for isLegalAddImmediate to avoid
transforming an add of a legal immediate with one that isn't.
llvm-svn: 217610
This is an extension of the change made with r215820:
http://llvm.org/viewvc/llvm-project?view=revision&revision=215820
That patch allowed combining of splatted vector FP constants that are multiplied.
This patch allows combining non-uniform vector FP constants too by relaxing the
check on the type of vector. Also, canonicalize a vector fmul in the
same way that we already do for scalars - if only one operand of the fmul is a
constant, make it operand 1. Otherwise, we miss potential folds.
This fold is also done by -instcombine, but it's possible that extra
fmuls may have been generated during lowering.
Differential Revision: http://reviews.llvm.org/D5254
llvm-svn: 217599
This problem is bigger than just fsub, but this is the minimum fix to solve
fneg for PR20556 ( http://llvm.org/bugs/show_bug.cgi?id=20556 ), and we solve
zero subtraction with the same change.
llvm-svn: 217286
If an fmul was introduced by lowering, it wouldn't be folded
into a multiply by a constant since the earlier combine would
have replaced the fmul with the fadd.
llvm-svn: 216932
When I recommitted r208640 (in r216898) I added an exclusion for TargetConstant
offsets, as there is no guarantee that a backend can handle them on generic
ADDs (even if it generates them during address-mode matching) -- and,
specifically, applying this transformation directly with TargetConstants caused
a self-hosting failure on PPC64. Ignoring all TargetConstants, however, is less
than ideal. Instead, for non-opaque constants, we can convert them into regular
constants for use with the generated ADD (or SUB).
llvm-svn: 216908
I reverted r208640 in r209747 because r208640 broke self-hosting on PPC64. The
underlying cause of the failure is that pre-inc loads with increments
represented by ISD::TargetConstants were being transformed into ISD:::ADDs with
ISD::TargetConstant operands. PPC doesn't have a pattern for those, and so they
were selected as invalid r+r adds.
This recommits r208640, rebased and with an exclusion for ISD::TargetConstant
increments. This behavior seems correct, although in the future we might want
to ask the target to split out the indexing that uses ISD::TargetConstants.
Unfortunately, I don't yet have small test case where the relevant invalid
'add' instruction is not itself dead (and thus eliminated by
DeadMachineInstructionElim -- sometimes bugpoint is too good at removing things)
Original commit message (by Adam Nemet):
Right now the load may not get DCE'd because of the side-effect of updating
the base pointer.
This can happen if we lower a read-modify-write of an illegal larger type
(e.g. i48) such that the modification only affects one of the subparts (the
lower i32 part but not the higher i16 part). See the testcase.
In order to spot the dead load we need to revisit it when SimplifyDemandedBits
decided that the value of the load is masked off. This is the
CommitTargetLoweringOpt piece.
I checked compile time with ARM64 by sending SPEC bitcode files through llc.
No measurable change.
Fixes <rdar://problem/16031651>
llvm-svn: 216898
was marked custom. The target independent DAG combine has no way to know if
the shuffles it is introducing are ones that the target could support or not.
llvm-svn: 216678
isPow2DivCheap
That name doesn't specify signed or unsigned.
Lazy as I am, I eventually read the function and variable comments. It turns out that this is strictly about signed div. But I discovered that the comments are wrong:
srl/add/sra
is not the general sequence for signed integer division by power-of-2. We need one more 'sra':
sra/srl/add/sra
That's the sequence produced in DAGCombiner. The first 'sra' may be removed when dividing by exactly '2', but that's a special case.
This patch corrects the comments, changes the name of the flag bit, and changes the name of the accessor methods.
No functional change intended.
Differential Revision: http://reviews.llvm.org/D5010
llvm-svn: 216237
When combining a pair of shuffle nodes, check if the combined shuffle mask is
trivially Undef. In case, immediately fold that pair of shuffles to Undef.
The lack of checks for undef masks was the root-cause of a poor-codegen bug
in the dag combiner.
Example:
%1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 1, i32 1, i32 6>
%2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 4, i32 1, i32 6>
%3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 1, i32 5, i32 3, i32 3>
Before this patch, on x86 (with -mcpu=corei7) we failed to fold the entire
sequence to Undef value and therefore we generated:
shufps $-123, %xmm1, $xmm0
pshufd $-46, %xmm0, %xmm0
With this patch, the entire shuffle sequence is folded to Undef and no
shuffles are generated in the output assembly.
Added new test cases to test 'combine-vec-shuffle-5.ll'.
llvm-svn: 215797
This patch allows a vector fneg of a bitcasted integer value to be optimized in the same way that we already optimize a scalar fneg. If the integer variable is a constant, we can precompute the result and not require any logic ops.
This patch is very similar to a fabs patch committed at r214892.
Differential Revision: http://reviews.llvm.org/D4852
llvm-svn: 215646
input node after manually adding it to the worklist and using CombineTo.
Once we use CombineTo the input node may have been deleted. Despite this
being *completely confusing* and somewhat broken, the only way to
"correctly" return from a DAG combine after potentially deleting the
input node is to return *that exact node*....
But really, this code should just never have used CombineTo. It won't do
what it wants (returning the node as mentioned above just causes the
combine to infloop). The correct way to combine away a casted load to
a load of the correct type is to RAUW the chain directly and then return
the loaded value to replace the actual value node.
I managed to find this with the vector shuffle fuzzer even though it
clearly has nothing at all to do with vector shuffles and rather those
happen to trigger a load of a constant pool that hits this combine *just
right*. I've included the test as it is small and a nice stress test
that the infrastructure isn't asserting.
llvm-svn: 215622
This patch improves the existing algorithm in DAGCombiner that
attempts to fold shuffles according to rule:
shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)
Before this change, there were cases where the DAGCombiner conservatively
avoided folding shuffles even if the resulting mask would have been legal.
That is because the algorithm wrongly assumed that commuting
an illegal shuffle mask would always produce an illegal mask.
With this change, we now correctly compute the commuted shuffle mask before
calling method 'isShuffleMaskLegal' on it.
On X86, this improves for example the codegen for the following function:
define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
%1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 2, i32 6, i32 7>
%2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 2, i32 3>
ret <4 x i32> %2
}
Before this change the X86 backend (-mcpu=corei7) generated
the following assembly code for function @test:
shufps $-23, %xmm0, %xmm1 # xmm1 = xmm1[1,2],xmm0[2,3]
movhlps %xmm1, %xmm1 # xmm1 = xmm1[1,1]
movaps %xmm1, %xmm0
Now we produce:
movhlps %xmm0, %xmm0 # xmm0 = xmm0[1,1]
Added extra test cases in combine-vec-shuffle-2.ll to verify that we correctly
fold according to the above-mentioned rule.
llvm-svn: 215555
Allow vector fabs operations on bitcasted constant integer values to be optimized
in the same way that we already optimize scalar fabs.
So for code like this:
%bitcast = bitcast i64 18446744069414584320 to <2 x float> ; 0xFFFF_FFFF_0000_0000
%fabs = call <2 x float> @llvm.fabs.v2f32(<2 x float> %bitcast)
%ret = bitcast <2 x float> %fabs to i64
Instead of generating something like this:
movabsq (constant pool loadi of mask for sign bits)
vmovq (move from integer register to vector/fp register)
vandps (mask off sign bits)
vmovq (move vector/fp register back to integer return register)
We should generate:
mov (put constant value in return register)
I have also removed a redundant clause in the first 'if' statement:
N0.getOperand(0).getValueType().isInteger()
is the same thing as:
IntVT.isInteger()
Testcases for x86 and ARM added to existing files that deal with vector fabs.
One existing testcase for x86 removed because it is no longer ideal.
For more background, please see:
http://reviews.llvm.org/D4770
And:
http://llvm.org/bugs/show_bug.cgi?id=20354
Differential Revision: http://reviews.llvm.org/D4785
llvm-svn: 214892
This code is completely wrong. It is also dead, as if it were to *ever*
run, it would crash. Fortunately, after my work to the combiner, it is
at least *possible* to reach the code, and llvm-stress has found a test
case. Thanks to Patrick for reporting.
It would be really good if anyone who remembers how this code works and
what it was intended to do could add some more obvious test coverage
instead of my completely contrived and reduced test case. My test case
was so brittle I left a bread crumb comment in it to help the next
person to stumble on it and not know what it was actually testing for.
llvm-svn: 214785
combines) until they are legal.
Doing it the old way could, when the stars align *just* right, cause
a node to get into the combine set prior to being legalized. Then, when
the same node showed up as an operand to another node later on (but not
so much later on that it had been deleted as dead) we would fail to add
it back to the worklist thinking it had already been combined. This
would in turn cause it to not be legalized. Fortunately, we can also
walk the operands looking for uncombined (and thus potentially
un-legalized) nodes late. It will still ensure that we walk all operands
of all nodes and send all of them through both the legalizer without
changes and the combiner at least once. (Which was the original goal of
this).
I have a test case for this bug, but it is terribly brittle. For
example, it will stop finding the bug the moment I enable the new
shuffle lowering. I don't yet have any test case that reliably exercises
this bug, and it isn't clear that it will be possible to craft one. It
is entirely possible that with the new shuffle lowering the two forms of
doing this are precisely equivalent. That doesn't mean we shouldn't take
the more conservative approach of insisting on things in the combined
set having survived the legalizer.
llvm-svn: 214673
This is intended to be the minimal change needed to fix PR20354 ( http://llvm.org/bugs/show_bug.cgi?id=20354 ). The check for a vector operation was wrong; we need to check that the fabs itself is not a vector operation.
This patch will not generate the optimal code. A constant pool load and 'and' op will be generated instead of just returning a value that we can calculate in advance (as we do for the scalar case). I've put a 'TODO' comment for that here and expect to have that patch ready soon.
There is a very similar optimization that we can do in visitFNEG, so I've put another 'TODO' there and expect to have another patch for that too.
llvm-svn: 214670
so using a single helper which adds operands back onto the worklist.
Several places didn't rigorously do this but a couple already did.
Factoring them together and doing it rigorously is important to delete
things recursively early on in the combiner and get a chance to see
accurate hasOneUse values. While no existing test cases change, an
upcoming patch to add DAG combining logic for PSHUFB requires this to
work correctly.
llvm-svn: 214623
during DAGCombine in certain circumstances. Unfortunately, the circumstances required
to trigger the issue seem to require a pretty specific interaction of DAGCombines,
and I haven't been able to find a testcase that reproduces on X86, ARM, or AArch64.
The functionality added here is replicated in essentially every other DAG combine,
so it seems pretty obviously correct.
llvm-svn: 214622
Currently when DAGCombine converts loads feeding a switch into a switch of
addresses feeding a load the new load inherits the isInvariant flag of the left
side. This is incorrect since invariant loads can be reordered in cases where it
is illegal to reoarder normal loads.
This patch adds an isInvariant parameter to getExtLoad() and updates all call
sites to pass in the data if they have it or false if they don't. It also
changes the DAGCombine to use that data to make the right decision when
creating the new load.
llvm-svn: 214449
DAGCombine may choose to rewrite graphs where two loads feed a select into
graphs where a select of two addresses feed a load. While it sanity checks the
loads to make sure they are broadly equivalent it currently just uses the
alignment restriction of the left node. In cases where the right node has
stronger alignment requiresment this may lead to bad codegen, such as generating
an aligned load where an unaligned load is required. This patch makes the
combine generate a load with an alignment that is the same as whichever is more
restrictive of the two alignments.
Tests included.
rdar://17762530
llvm-svn: 214322
inspection in the proccess, and shuffle the logging in the DAG combiner
around a bit.
With this it is much easier to follow what the legalizer is doing. It
should even accurately present most of the strange legalization
operations where a single node is replaced by multiple nodes, etc. There
is still some information lost (we log SDNodes not SDValues so we don't
log which result is used for which thing), but I think this is much
closer to a usable system. Notably, this will make it *much* more
apparant when legalization is actually happening inside the combiner, or
when there is a cycle caused by interactions of the legalizer and the
combiner.
The "bug" I fixed here I'm not sure is remotely possible to trigger. We
were only adding one of the nodes in a replacement to the updated set
rather than all of the nodes in the replacement. Realistically, the
worst result of this are nodes not getting back onto the worklist in the
DAG combiner. I doubt it is possible to trigger this today, and
I certainly don't have any ideas about how, but this at least brings the
code into alignment with the principled operation of the routine.
llvm-svn: 214105
over each node in the worklist prior to combining.
This allows the combiner to produce new nodes which need to go back
through legalization. This is particularly useful when generating
operands to target specific nodes in a post-legalize DAG combine where
the operands are significantly easier to express as pre-legalized
operations. My immediate use case will be PSHUFB formation where we need
to build a constant shuffle mask with a build_vector node.
This also refactors the relevant functionality in the legalizer to
support this, and updates relevant tests. I've spoken to the R600 folks
and these changes look like improvements to them. The avx512 change
needs to be investigated, I suspect there is a disagreement between the
legalizer and the DAG combiner there, but it seems a minor issue so
leaving it to be re-evaluated after this patch.
Differential Revision: http://reviews.llvm.org/D4564
llvm-svn: 214020
with a result number outside the range of results for the node.
I don't know how we managed to not really check this very basic
invariant for so long, but the code is *very* broken at this point.
I have over 270 test failures with the assert enabled. I'm committing it
disabled so that others can join in the cleanup effort and reproduce the
issues. I've also included one of the obvious fixes that I already
found. More fixes to come.
llvm-svn: 213926
which have successfully round-tripped through the combine phase, and use
this to ensure all operands to DAG nodes are visited by the combiner,
even if they are only added during the combine phase.
This is critical to have the combiner reach nodes that are *introduced*
during combining. Previously these would sometimes be visited and
sometimes not be visited based on whether they happened to end up on the
worklist or not. Now we always run them through the combiner.
This fixes quite a few bad codegen test cases lurking in the suite while
also being more principled. Among these, the TLS codegeneration is
particularly exciting for programs that have this in the critical path
like TSan-instrumented binaries (although I think they engineer to use
a different TLS that is faster anyways).
I've tried to check for compile-time regressions here by running llc
over a merged (but not LTO-ed) clang bitcode file and observed at most
a 3% slowdown in llc. Given that this is essentially a worst case (none
of opt or clang are running at this phase) I think this is tolerable.
The actual LTO case should be even less costly, and the cost in normal
compilation should be negligible.
With this combining logic, it is possible to re-legalize as we combine
which is necessary to implement PSHUFB formation on x86 as
a post-legalize DAG combine (my ultimate goal).
Differential Revision: http://reviews.llvm.org/D4638
llvm-svn: 213898
In order to enable the preservation of noalias function parameter information
after inlining, and the representation of block-level __restrict__ pointer
information (etc.), additional kinds of aliasing metadata will be introduced.
This metadata needs to be carried around in AliasAnalysis::Location objects
(and MMOs at the SDAG level), and so we need to generalize the current scheme
(which is hard-coded to just one TBAA MDNode*).
This commit introduces only the necessary refactoring to allow for the
introduction of other aliasing metadata types, but does not actually introduce
any (that will come in a follow-up commit). What it does introduce is a new
AAMDNodes structure to hold all of the aliasing metadata nodes associated with
a particular memory-accessing instruction, and uses that structure instead of
the raw MDNode* in AliasAnalysis::Location, etc.
No functionality change intended.
llvm-svn: 213859
The target-independent DAGcombiner will generate:
asr w1, X, #31 w1 = splat sign bit.
add X, X, w1, lsr #28 X = X + 0 or pow2-1
asr w0, X, asr #4 w0 = X/pow2
However, the add + shifts is expensive, so generate:
add w0, X, 15 w0 = X + pow2-1
cmp X, wzr X - 0
csel X, w0, X, lt X = (X < 0) ? X + pow2-1 : X;
asr w0, X, asr 4 w0 = X/pow2
llvm-svn: 213758
insertions.
The old behavior could cause arbitrarily bad memory usage in the DAG
combiner if there was heavy traffic of adding nodes already on the
worklist to it. This commit switches the DAG combine worklist to work
the same way as the instcombine worklist where we null-out removed
entries and only add new entries to the worklist. My measurements of
codegen time shows slight improvement. The memory utilization is
unsurprisingly dominated by other factors (the IR and DAG itself
I suspect).
This change results in subtle, frustrating churn in the particular order
in which DAG combines are applied which causes a number of minor
regressions where we fail to match a pattern previously matched by
accident. AFAICT, all of these should be using AddToWorklist to directly
or should be written in a less brittle way. None of the changes seem
drastically bad, and a few of the changes seem distinctly better.
A major change required to make this work is to significantly harden the
way in which the DAG combiner handle nodes which become dead
(zero-uses). Previously, we relied on the ability to "priority-bump"
them on the combine worklist to achieve recursive deletion of these
nodes and ensure that the frontier of remaining live nodes all were
added to the worklist. Instead, I've introduced a routine to just
implement that precise logic with no indirection. It is a significantly
simpler operation than that of the combiner worklist proper. I suspect
this will also fix some other problems with the combiner.
I think the x86 changes are really minor and uninteresting, but the
avx512 change at least is hiding a "regression" (despite the test case
being just noise, not testing some performance invariant) that might be
looked into. Not sure if any of the others impact specific "important"
code paths, but they didn't look terribly interesting to me, or the
changes were really minor. The consensus in review is to fix any
regressions that show up after the fact here.
Thanks to the other reviewers for checking the output on other
architectures. There is a specific regression on ARM that Tim already
has a fix prepped to commit.
Differential Revision: http://reviews.llvm.org/D4616
llvm-svn: 213727
'Worklist' consistently rather than a deeply confusing mixture of
'WorkList' and 'Worklist'.
Notably, the very 'WorkList' of the DAG combiner was exposed to target
specific DAG combines under an interface 'AddToWorklist' which was
implemented by in turn calling 'AddToWorkList' in the combiner. This has
sent me circling with the wrong case in grep one too many times.
I chose to normalize on 'Worklist' because that one won the grep-vote
for llvm/lib/... by a hundered hits or so, and it is used in places
relatively "canonical" such as InstCombine's Worklist. Let's all jsut
pick this casing, whether "correct", "good", or "bad" and be
consistent...
llvm-svn: 213506
stack, filter all handle nodes from the DAG combiner worklist.
This will also handle cases where other handle nodes might be
(erroneously) added to the worklist and then cause bugs and explosions
when deleted. For example, when running the legalizer within the DAG
combiner, there are times when other handle nodes are used and can end
up here.
llvm-svn: 213505
Canonicalize shuffles according to rules:
* shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
* shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
* shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
This patch helps identifying more shuffle pairs that could be combined reusing
the already existing rules in the DAGCombiner.
Added new test 'combine-vec-shuffle-5.ll' to verify that the canonicalized
shuffles are now folded into a single shuffle node by the DAGCombiner.
Added more test cases to 'combine-vec-shuffle-4.ll'.
llvm-svn: 213504
It turns out that in most cases (the main exception being i1-related
types) once these operations are formed we cannot separate them and
the targets end up having to deal with them whether they want to or
not.
This is not a good situation, and a more reasonable default can be
formed by ackowledging this and having targets leave them as Legal.
Only x86 seems to be affected (other targets don't even try marking
the operation Expand).
Mostly there's no visible change here yet, but it will be useful to
have truly expanded EXTLOADS for MVT::f16 softening support.
llvm-svn: 213162
This patch adds two new rules to the DAGCombiner:
1. shuffle (shuffle A, Undef, M0), B, M1 -> shuffle A, B, M2
2. shuffle (shuffle A, Undef, M0), A, M1 -> shuffle A, Undef, M2
We only do this if the combined shuffle is legal for the target.
Example:
;;
define <4 x float> @test(<4 x float> %a, <4 x float> %b) {
%1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32><i32 6, i32 0, i32 1, i32 7>
%2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32><i32 1, i32 2, i32 4, i32 5>
ret <4 x i32> %2
}
;;
(using llc -mcpu=corei7 -march=x86-64)
Before, the x86 backend generated:
pshufd $120, %xmm0, %xmm0
shufps $-108, %xmm0, %xmm1
movaps %xmm1, %xmm0
Now the x86 backend generates:
movsd %xmm1, %xmm0
llvm-svn: 213069
This patch fixes a crasher in method 'DAGCombiner::visitOR' due to an invalid
call to method 'isShuffleMaskLegal'. On x86, method 'isShuffleMaskLegal'
always expects a legal vector value type in input.
With this patch, we immediately check if the input OR dag node has a legal
vector type; we only try to fold a OR dag node into a single shufflevector
if we know that the resulting shuffle will have a legal type.
This is to avoid calling method 'isShuffleMaskLegal' on a potentially
illegal vector value type.
Added a new test-case to file 'CodeGen/X86/combine-or.ll' to verify that
DAGCombiner doesn't crash in the attempt to check/combine an OR between shuffles
with illegal types.
llvm-svn: 213020
This patch teaches the DAGCombiner how to fold a pair of shuffles
according to rules:
1. shuffle(shuffle A, B, M0), B, M1) -> shuffle(A, B, M2)
2. shuffle(shuffle A, B, M0), A, M1) -> shuffle(A, B, M3)
The new rules would only trigger if the resulting shuffle has legal type and
legal mask.
Added test 'combine-vec-shuffle-3.ll' to verify that DAGCombiner correctly
folds shuffles on x86 when the resulting mask is legal. Also added some negative
cases to verify that we avoid introducing illegal shuffles.
llvm-svn: 213001
Verify that DAGCombiner does not crash when trying to fold a pair of shuffles
according to rule (added at r212539):
(shuffle (shuffle A, Undef, M0), Undef, M1) -> (shuffle A, Undef, M2)
The DAGCombiner avoids folding shuffles if the resulting shuffle dag node
is not legal for the target. That means, the resulting shuffle must have
legal type and legal mask.
Before, the DAGCombiner only called method
'TargetLowering::isShuffleMaskLegal' to check if it was "safe" to fold according
to the above-mentioned rule. However, this caused a crash in the x86 backend
since method 'isShuffleMaskLegal' always expects to be called on a
legal vector type.
llvm-svn: 212915
This patch teaches the DAGCombiner how to fold shuffles according to the
following new rules:
1. shuffle(shuffle(x, y), undef) -> x
2. shuffle(shuffle(x, y), undef) -> y
3. shuffle(shuffle(x, y), undef) -> shuffle(x, undef)
4. shuffle(shuffle(x, y), undef) -> shuffle(y, undef)
The backend avoids to combine shuffles according to rules 3. and 4. if
the resulting shuffle does not have a legal mask. This is to avoid introducing
illegal shuffles that are potentially expanded into a sub-optimal sequence of
target specific dag nodes during vector legalization.
Added test case combine-vec-shuffle-2.ll to verify that we correctly triggers
the new rules when combining shuffles.
llvm-svn: 212748
Summary:
On MIPS32r6/MIPS64r6, floating point comparisons return 0 or -1 but integer
comparisons return 0 or 1.
Updated the various uses of getBooleanContents. Two simplifications had to be
disabled when float and int boolean contents differ:
- ScalarizeVecRes_VSELECT except when the kind of boolean contents is trivially
discoverable (i.e. when the condition of the VSELECT is a SETCC node).
- visitVSELECT (select C, 0, 1) -> (xor C, 1).
Come to think of it, this one could test for the common case of 'C'
being a SETCC too.
Preserved existing behaviour for all other targets and updated the affected
MIPS32r6/MIPS64r6 tests. This also fixes the pi benchmark where the 'low'
variable was counting in the wrong direction because it thought it could simply
add the result of the comparison.
Reviewers: hfinkel
Reviewed By: hfinkel
Subscribers: hfinkel, jholewinski, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D4389
llvm-svn: 212697
tracks which elements of the build vector are in fact undef.
This should make actually inpsecting them (likely in my next patch)
reasonably pretty. Also makes the output parameter optional as it is
clear now that *most* users are happy with undefs in their splats.
llvm-svn: 212581
This patch teaches how to fold a shuffle according to rule:
shuffle (shuffle (x, undef, M0), undef, M1) -> shuffle(x, undef, M2)
We do this only if the resulting mask M2 is legal; this is to avoid introducing
illegal shuffles that are potentially expanded into a sub-optimal sequence
of target specific dag nodes.
This patch has the advantage of being target independent, since it works on ISD
nodes. Therefore, all targets (not only x86) can take advantage of this rule.
The idea behind this patch is that most shuffle pairs can be safely combined
before we run the legalizer on vector operations. This allows us to
combine/simplify dag nodes earlier in the process and not only immediately
before instruction selection stage.
That said. This patch is not meant to replace any existing target specific
combine rules; backends might still introduce new shuffles during legalization
stage. Also, this rule is very simple and avoids to aggressively optimize
shuffles.
llvm-svn: 212539
nodes about whether they are splats. This is factored out and improved
from r212324 which got reverted as it was far too aggressive. The new
API should help more conservatively handle buildvectors that are
a mixture of splatted and undef values.
No functionality change at this point. The hope is to slowly
re-introduce the undef-tolerant optimization of splats, but each time
being forced to make a concious decision about how to handle the undefs
in a way that doesn't lead to contradicting assumptions about the
collapsed value.
Hal has pointed out in discussions that this may not end up being the
desired API and instead it may be more convenient to get a mask of the
undef elements or something similar. I'm starting simple and will expand
the API as I adapt actual callers and see exactly what they need.
llvm-svn: 212514
lanes in vector splats.
The core problem here is that undef lanes can't *unilaterally* be
considered to contribute to splats. Their handling needs to be more
cautious. There is also a reported failure of the nightly testers
(thanks Tobias!) that may well stem from the same core issue. I'm going
to fix this theoretical issue, factor the APIs a bit better, and then
verify that I don't see anything bad with Tobias's reduction from the
test suite before recommitting.
Original commit message for r212324:
[x86] Generalize BuildVectorSDNode::getConstantSplatValue to work for
any constant, constant FP, or undef splat and to tolerate any undef
lanes in a splat, then replace all uses of isSplatVector in X86's
lowering with it.
This fixes issues where undef lanes in an otherwise splat vector would
prevent the splat logic from firing. It is a touch more awkward to use
this interface, but it is much more accurate. Suggestions for better
interface structuring welcome.
With this fix, the code generated with the widening legalization
strategy for widen_cast-4.ll is *dramatically* improved as the special
lowering strategies for a v16i8 SRA kick in even though the high lanes
are undef.
We also get a slightly different choice for broadcasting an aligned
memory location, and use vpshufd instead of vbroadcastss. This looks
like a minor win for pipelining and domain crossing, but a minor loss
for the number of micro-ops. I suspect its a wash, but folks can
easily tweak the lowering if they want.
llvm-svn: 212475
any constant, constant FP, or undef splat and to tolerate any undef
lanes in a splat, then replace all uses of isSplatVector in X86's
lowering with it.
This fixes issues where undef lanes in an otherwise splat vector would
prevent the splat logic from firing. It is a touch more awkward to use
this interface, but it is much more accurate. Suggestions for better
interface structuring welcome.
With this fix, the code generated with the widening legalization
strategy for widen_cast-4.ll is *dramatically* improved as the special
lowering strategies for a v16i8 SRA kick in even though the high lanes
are undef.
We also get a slightly different choice for broadcasting an aligned
memory location, and use vpshufd instead of vbroadcastss. This looks
like a minor win for pipelining and domain crossing, but a minor loss
for the number of micro-ops. I suspect its a wash, but folks can easily
tweak the lowering if they want.
llvm-svn: 212324
The PowerPC 128-bit long double data type (ppcf128 in LLVM) is in fact a
pair of two doubles, where one is considered the "high" or
more-significant part, and the other is considered the "low" or
less-significant part. When a ppcf128 value is stored in memory or a
register pair, the high part always comes first, i.e. at the lower
memory address or in the lower-numbered register, and the low part
always comes second. This is true both on big-endian and little-endian
PowerPC systems. (Similar to how with a complex number, the real part
always comes first and the imaginary part second, no matter the byte
order of the system.)
This was implemented incorrectly for little-endian systems in LLVM.
This commit fixes three related issues:
- When printing an immediate ppcf128 constant to assembler output
in emitGlobalConstantFP, emit the high part first on both big-
and little-endian systems.
- When lowering a ppcf128 type to a pair of f64 types in SelectionDAG
(which is used e.g. when generating code to load an argument into a
register pair), use correct low/high part ordering on little-endian
systems.
- In a related issue, because lowering ppcf128 into a pair of f64 must
operate differently from lowering an int128 into a pair of i64,
bitcasts between ppcf128 and int128 must not be optimized away by the
DAG combiner on little-endian systems, but must effect a word-swap.
Reviewed by Hal Finkel.
llvm-svn: 212274
The SelectionDAG bad a special case for ISD::SELECT_CC, where it would
allow targets to specify:
setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
to indicate that they wanted to expand ISD::SELECT_CC for all types.
This wasn't applied correctly everywhere, and it makes writing new
DAG patterns with ISD::SELECT_CC difficult.
llvm-svn: 210541
This patch adds new target specific combine rules to identify horizontal
add/sub idioms from BUILD_VECTOR dag nodes.
This patch also teaches the DAGCombiner how to canonicalize sequences of
insert_vector_elt dag nodes according to the following rule:
(insert_vector_elt (insert_vector_elt A, I0), I1) ->
(insert_vecto_elt (insert_vector_elt A, I1), I0)
This new canonicalization rule only triggers if the inner insert_vector
dag node has exactly one use; also, both indices must be known constants,
and I1 < I0.
This last rule made it possible to write a simpler algorithm to identify
horizontal add/sub patterns because now we don't have to worry about the
ordering of insert_vector_elt dag nodes.
llvm-svn: 210477
This patch modifies SelectionDAGBuilder to construct SDNodes with associated
NoSignedWrap, NoUnsignedWrap and Exact flags coming from IR BinaryOperator
instructions.
Added a new SDNode type called 'BinaryWithFlagsSDNode' to allow accessing
nsw/nuw/exact flags during codegen.
Patch by Marcello Maggioni.
llvm-svn: 210467
This patch teaches the backend how to simplify/canonicalize dag node
sequences normally introduced by the backend when promoting certain dag nodes
with illegal vector type.
This patch adds two new combine rules:
1) fold (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
(shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
2) fold (BINOP (shuffle (A, Undef, <Mask>)), (shuffle (B, Undef, <Mask>))) ->
(shuffle (BINOP A, B), Undef, <Mask>).
Both rules are only triggered on the type-legalized DAG.
In particular, rule 1. is a target specific combine rule that attempts
to sink a bitconvert into the operands of a binary operation.
Rule 2. is a target independet rule that attempts to move a shuffle
immediately after a binary operation.
llvm-svn: 209930
Summary:
If both vector args to vselect are concat_vectors and the condition is
constant and picks half a vector from each argument, convert the vselect
into a concat_vectors.
Added a test.
The ConvertSelectToConcatVector is assuming it doesn't get vselects with
arguments of, for example, <undef, undef, true, true>. Those get taken
care of in the checks above its call.
Reviewers: nadav, delena, grosbach, hfinkel
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3916
llvm-svn: 209929
An address only use of an extract element of a load can be simplified to a
load. Without this the result of the extract element is spilled to the
stack so that an address is available.
llvm-svn: 209788
This reverts r208640 (I've just XFAILed the test) because it broke ppc64/Linux
self-hosting. Because nearly every regression test triggers a segfault, I hope
this will be easy to fix.
llvm-svn: 209747
Right now the load may not get DCE'd because of the side-effect of updating
the base pointer.
This can happen if we lower a read-modify-write of an illegal larger type
(e.g. i48) such that the modification only affects one of the subparts (the
lower i32 part but not the higher i16 part). See the testcase.
In order to spot the dead load we need to revisit it when SimplifyDemandedBits
decided that the value of the load is masked off. This is the
CommitTargetLoweringOpt piece.
I checked compile time with ARM64 by sending SPEC bitcode files through llc.
No measurable change.
Fixes <rdar://problem/16031651>
llvm-svn: 208640
For pattern like ((x >> C1) & Mask) << C2, DAG combiner may convert it
into (x >> (C1-C2)) & (Mask << C2), which makes pattern matching of ubfx
more difficult.
For example:
Given
%shr = lshr i64 %x, 4
%and = and i64 %shr, 15
%arrayidx = getelementptr inbounds [8 x [64 x i64]]* @arr, i64 0, %i64 2, i64 %and
%0 = load i64* %arrayidx
With current shift folding, it takes 3 instrs to compute base address:
lsr x8, x0, #1
and x8, x8, #0x78
add x8, x9, x8
If using ubfx, it only needs 2 instrs:
ubfx x8, x0, #4, #4
add x8, x9, x8, lsl #3
This fixes bug 19589
llvm-svn: 207702
Artur Pilipenko