Summary:
Various SelectionDAG non-combine operations (e.g. the getNode smart
constructor and legalization) may leave dangling nodes by applying
optimizations without fully pruning unused result values. This results
in nodes that are never added to the worklist and therefore can not be
pruned.
Add a node inserter for the combiner to make sure such nodes have the
chance of being pruned. This allows a number of additional peephole
optimizations.
Reviewers: efriedma, RKSimon, craig.topper, jyknight
Reviewed By: jyknight
Subscribers: msearles, jyknight, sdardis, nemanjai, javed.absar, hiraditya, jrtc27, atanasyan, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D58068
llvm-svn: 357279
After investigating the examples from D59777 targeting an SSE4.1 machine,
it looks like a very different problem due to how we map illegal types (256-bit in these cases).
We're missing a shuffle simplification that maps elements of a vector back to a shuffled operand.
We have a more general version of this transform in DAGCombiner::visitVECTOR_SHUFFLE(), but that
generality means it is limited to patterns with a one-use constraint, and the examples here have
2 uses. We don't need any uses or legality limitations for a simplification (no new value is
created).
It looks like we miss this pattern in IR too.
In one of the zext examples here, we have shuffle masks like this:
Shuf0 = vector_shuffle<0,u,3,7,0,u,3,7>
Shuf = vector_shuffle<4,u,6,7,u,u,u,u>
...so that's moving the high half of the 1st vector into the low half. But the high half of the
1st vector is already identical to the low half.
Differential Revision: https://reviews.llvm.org/D59961
llvm-svn: 357258
This is a sibling to rL357178 that I noticed we'd hit if we chose
an alternate transform in D59818.
%z = zext i8 %x to i32
%dec = add i32 %z, -1
%r = sext i32 %dec to i64
=>
%z2 = zext i8 %x to i64
%r = add i64 %z2, -1
https://rise4fun.com/Alive/kPP
The x86 vector diffs show a slight regression, so there's a chance
that we should limit this and the previous transform to scalars.
But given that we allowed vectors before, I'm matching that behavior
here. We should change both transforms together if that's the right
thing to do.
llvm-svn: 357254
In the example below, we would previously emit two range checks, one for cases
1--3 and one for 4--6. This patch makes us exploit the fact that the
fall-through is unreachable and only one range check is necessary.
switch i32 %i, label %default [
i32 1, label %bb1
i32 2, label %bb1
i32 3, label %bb1
i32 4, label %bb2
i32 5, label %bb2
i32 6, label %bb2
]
default: unreachable
llvm-svn: 357252
Some DAG mutations can only be applied to `ScheduleDAGMI`, and have to
internally cast a `ScheduleDAGInstrs` to `ScheduleDAGMI`.
There is nothing actually specific to `ScheduleDAGMI` in `Topo`.
llvm-svn: 357239
Even if the interleaving transform would otherwise be legal, we shouldn't
introduce an interleaved load that is wider than the original load: it might
have undefined behavior.
It might be possible to perform some sort of mask-narrowing transform in
some cases (using a narrower interleaved load, then extending the
results using shufflevectors). But I haven't tried to implement that,
at least for now.
Fixes https://bugs.llvm.org/show_bug.cgi?id=41245 .
Differential Revision: https://reviews.llvm.org/D59954
llvm-svn: 357212
If scalar truncates are free, attempt to pre-truncate build_vectors source operands.
Only attempt to do this before legalization as we often end up with truncations/extensions during build_vector lowering.
Differential Revision: https://reviews.llvm.org/D59654
llvm-svn: 357161
When lowering a load or store for TypeWidenVector, the type legalizer
would use a single load or store if the associated integer type was legal
or promoted. E.g. it loads a v4i8 as an i32 if i32 is legal/promotable.
(See https://reviews.llvm.org/rL236528 for reference.)
This applies that behaviour to vector types. If the vector type is
TypePromoteInteger, the element type is going to be TypePromoteInteger
as well, which will lead to have a single promoting load rather than N
individual promoting loads. For instance, if we have a v3i1, we would
now have a load of v4i1 instead of 3 loads of i1.
Patch by Guillaume Marques. Thanks!
Differential Revision: https://reviews.llvm.org/D56201
llvm-svn: 357120
Summary:
A recent fix (r355751) caused a compile time regression because setting
the ModifiedDT flag in optimizeSelectInst means that each time a select
instruction is optimized the function walk in runOnFunction stops and
restarts again (which was needed to build a new DT before we started
building it lazily in r356937). Now that the DT is built lazily, a
simple fix is to just reset the DT at this point, rather than restarting
the whole function walk.
In the future other places that set ModifiedDT may want to switch to
just resetting the DT directly. But that will require an evaluation to
ensure that they don't otherwise need to restart the function walk.
Reviewers: spatel
Subscribers: jdoerfert, llvm-commits, xur
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59889
llvm-svn: 357111
Split out from D59749. The current implementation of isWrappedSet()
doesn't do what it says on the tin, and treats ranges like
[X, Max] as wrapping, because they are represented as [X, 0) when
using half-inclusive ranges. This also makes it inconsistent with
the semantics of isSignWrappedSet().
This patch renames isWrappedSet() to isUpperWrapped(), in preparation
for the introduction of a new isWrappedSet() method with corrected
behavior.
llvm-svn: 357107
If there were only dbg_values in the block, recede would hit the
beginning of the block and try to use thet dbg_value as a real
instruction.
llvm-svn: 357105
The artifact combiners push instructions which have been marked for deletion
onto an list for the legalizer to deal with on return. However, for trunc(ext)
combines the combiner routine recursively calls itself. When it does this the
dead instructions list may not be empty, and the other combiners don't expect
to be dealing with essentially invalid MIR (multiple vreg defs etc).
This change fixes it by ensuring that the dead instructions are processed on
entry into tryCombineInstruction.
As a result, this fix exposed a few places in tests where G_TRUNC instructions
were not being deleted even though they were dead.
Differential Revision: https://reviews.llvm.org/D59892
llvm-svn: 357101
This patch removes an overly conservative check that would prevent
simplifying copies when the value we were tracking would go through
several subregister indices.
Indeed, the intend of this check was to not track values whenever
we have to compose subregister, but actually what the check was
doing was bailing anytime we see a second subreg, even if that
second subreg would actually be the new source of truth (as opposed
to a part of that subreg).
Differential Revision: https://reviews.llvm.org/D59891
llvm-svn: 357095
Currently this is called before the frame size is set on the
function. For AMDGPU, the scavenger is used for large frames where
part of the offset needs to be materialized in a register, so
estimating the frame size is useful for knowing whether the scavenger
is useful.
llvm-svn: 357087
The AMDGPU implementation of getReservedRegs depends on
MachineFunctionInfo fields that are parsed from the YAML section. This
was reserving the wrong register since it was setting the reserved
regs before parsing the correct one.
Some tests were relying on the default reserved set for the assumed
default calling convention.
llvm-svn: 357083
Rework BaseIndexOffset and isAlias to fully work with lifetime nodes
and fold in lifetime alias analysis.
This is mostly NFC.
Reviewers: courbet
Reviewed By: courbet
Subscribers: hiraditya, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59794
llvm-svn: 357070
Original commit by Ayonam Ray.
This commit adds a regression test for the issue discovered in the
previous commit: that the range check for the jump table can only be
omitted if the fall-through destination of the jump table is
unreachable, which isn't necessarily true just because the default of
the switch is unreachable.
This addresses the missing optimization in PR41242.
> During the lowering of a switch that would result in the generation of a
> jump table, a range check is performed before indexing into the jump
> table, for the switch value being outside the jump table range and a
> conditional branch is inserted to jump to the default block. In case the
> default block is unreachable, this conditional jump can be omitted. This
> patch implements omitting this conditional branch for unreachable
> defaults.
>
> Differential Revision: https://reviews.llvm.org/D52002
> Reviewers: Hans Wennborg, Eli Freidman, Roman Lebedev
llvm-svn: 357067
getAsCarry() checks that the input argument is a carry-producing node before
allowing a transformation to addcarry. This patch adds a check to make sure
that the carry-producing node is legal. If it is not, it may not remain in a
form that is manageable by the target backend. The test case caused a
compilation failure during instruction selection for this reason on SystemZ.
Patch by Ulrich Weigand.
Review: Sanjay Patel
https://reviews.llvm.org/D59822
llvm-svn: 357052
A section containing metadata on remark diagnostics will be emitted if
the flag (-mllvm) -remarks-section is present.
For now, the metadata is:
* a magic number for remarks: "REMARKS\0"
* the version number: a little-endian uint64_t
* the absolute file path to the serialized remark diagnostics: a
null-terminated string.
Differential Revision: https://reviews.llvm.org/D59571
llvm-svn: 357043
When splitting a subrange we end up with two different subranges covering
two different, non overlapping, lanes.
As part of this splitting the VNIs of the original live-range need
to be dispatched to the subranges according to which lanes they are
actually defining.
Prior to this patch we were assuming that all values were defining
all lanes. This was wrong as demonstrated by llvm.org/PR40835.
Differential Revision: https://reviews.llvm.org/D59731
llvm-svn: 357032
We have the folds for fadd/fsub/fmul already in DAGCombiner,
so it may be possible to remove that code if we can guarantee that
these ops are zapped before they can exist.
llvm-svn: 357029
This reverts commit rL357020.
The commit broke the test llvm/test/tools/llvm-objdump/embedded-source.test
on some builds including clang-ppc64be-linux-multistage,
clang-s390x-linux, clang-with-lto-ubuntu, clang-x64-windows-msvc,
llvm-clang-lld-x86_64-scei-ps4-windows10pro-fast (and others).
llvm-svn: 357026
Reapply rL356941 after regenerating the object file in the failing test
llvm/test/tools/llvm-objdump/embedded-source.test from source.
Original commit message:
[llvm] Prevent duplicate files in debug line header in dwarf 5.
Motivation: In previous dwarf versions, file name indexes started from 1, and
the primary source file was not explicit. Dwarf 5 standard (6.2.4) prescribes
the primary source file to be explicitly given an entry with an index number 0.
The current implementation honors the specification by just duplicating the
main source file, once with index number 0, and later maybe with another
index number. While this is compliant with the letter of the standard, the
duplication causes problems for consumers of this information such as lldb.
(Some files are duplicated, where only some of them have a line table although
all refer to the same file)
With this change, dwarf 5 debug line section files always start from 0, and
the zeroth entry is not duplicated whenever possible. This requires different
handling of dwarf 4 and dwarf 5 during generation (e.g. when a function returns
an index zero for a file name, it signals an error in dwarf 4, but not in dwarf 5)
However, I think the minor complication is worth it, because it enables all
consumers (lldb, gdb, dwarfdump, objdump, and so on) to treat all files in the
file name list homogenously.
Tags: #llvm, #debug-info
Differential Revision: https://reviews.llvm.org/D59515
llvm-svn: 357018
Various SelectionDAG non-combine operations (e.g. the getNode smart
constructor and legalization) may leave dangling nodes by applying
optimizations or not fully pruning unused result values. This can
result in nodes that are never added to the worklist and therefore can
not be pruned.
Add a node inserter as the current node deleter to make sure such
nodes have the chance of being pruned.
Many minor changes, mostly positive.
llvm-svn: 356996
This helps us relax the extension of a lot of scalar elements before they are inserted into a vector.
Its exposes an issue in DAGCombiner::convertBuildVecZextToZext as some/all the zero-extensions may be relaxed to ANY_EXTEND, so we need to handle that case to avoid a couple of AVX2 VPMOVZX test regressions.
Once this is in it should be easier to fix a number of remaining failures to fold loads into VBROADCAST nodes.
Differential Revision: https://reviews.llvm.org/D59484
llvm-svn: 356989
DenseMap iteration order is not guaranteed, use MapVector instead.
Fix provided by srhines.
Differential Revision: https://reviews.llvm.org/D59807
llvm-svn: 356988
Summary:
Motivation: In previous dwarf versions, file name indexes started from 1, and
the primary source file was not explicit. Dwarf 5 standard (6.2.4) prescribes
the primary source file to be explicitly given an entry with an index number 0.
The current implementation honors the specification by just duplicating the
main source file, once with index number 0, and later maybe with another
index number. While this is compliant with the letter of the standard, the
duplication causes problems for consumers of this information such as lldb.
(Some files are duplicated, where only some of them have a line table although
all refer to the same file)
With this change, dwarf 5 debug line section files always start from 0, and
the zeroth entry is not duplicated whenever possible. This requires different
handling of dwarf 4 and dwarf 5 during generation (e.g. when a function returns
an index zero for a file name, it signals an error in dwarf 4, but not in dwarf 5)
However, I think the minor complication is worth it, because it enables all
consumers (lldb, gdb, dwarfdump, objdump, and so on) to treat all files in the
file name list homogenously.
Reviewers: dblaikie, probinson, aprantl, espindola
Reviewed By: probinson
Subscribers: emaste, jvesely, nhaehnle, aprantl, javed.absar, arichardson, hiraditya, MaskRay, rupprecht, jdoerfert, llvm-commits
Tags: #llvm, #debug-info
Differential Revision: https://reviews.llvm.org/D59515
llvm-svn: 356941
First half of PR40800, this patch adds DAG undef handling to icmp instructions to match the behaviour in llvm::ConstantFoldCompareInstruction and SimplifyICmpInst, this permits constant folding of vector comparisons where some elements had been reduced to UNDEF (by SimplifyDemandedVectorElts etc.).
This involved a lot of tweaking to reduced tests as bugpoint loves to reduce icmp arguments to undef........
Differential Revision: https://reviews.llvm.org/D59363
llvm-svn: 356938
Summary:
In r355512 CGP was changed to build the DominatorTree only once per
function traversal, to avoid repeatedly building it each time it was
accessed. This solved one compile time issue but introduced another. In
the second case, we now were building the DT unnecessarily many times
when we performed many function traversals (i.e. more than once per
function when running CGP because of changes made each time).
Change to saving the DT in the CodeGenPrepare object, and building it
lazily when needed. It is reset whenever we need to rebuild it.
The case that exposed the issue there are 617 functions, and we walk
them (i.e. execute the "while (MadeChange)" loop in runOnFunction) a
total of 12083 times (so previously we were building the DT 12083
times). With this patch we only build the DT 844 times (average of 1.37
times per function). We dropped the total time to compile this file from
538.11s without this patch to 339.63s with it.
There is still an issue as CGP is taking much longer than all other
passes even with this patch, and before a recent compiler release cut at
r355392 the total time to this compile was only 97 sec with a huge
reduction in CGP time. I suspect that one of the other recent changes to
CGP led to iterating each function many more times on average, but I
need to do some more investigation.
Reviewers: spatel
Subscribers: jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59696
llvm-svn: 356937
I think this is correct, but may not necessarily be the correct fix
for the assertion I'm really trying to solve. If a scheduling region
was found that only has dbg_value instructions, the RegPressure
tracker would end up in an inconsistent state because it would skip
over any debug instructions and point to an instruction outside of the
scheduling region. It may still be possible for this to happen if
there are some real schedulable instructions between dbg_values, but I
haven't managed to break this.
The testcase is extremely sensitive and I'm not sure how to make it
more resistent to future scheduler changes that would avoid stressing
this situation.
llvm-svn: 356926
An i16 bswap can be implemented with an i16 rotate by 8. We previously emitted
a shift and OR sequence that DAG combine should be able to turn back into
rotate. But we might as well go there directly. If rotate isn't legal,
LegalizeDAG should further legalize it to either the opposite rotate, or the
shift and OR pattern.
I don't know of any way to get the existing DAG combine reliance to fail. So
I don't know any way to add new tests for this that wouldn't have worked
previously.
llvm-svn: 356860
The AArch64 test was broken since the result register already had a
set register class, so this test was a no-op. The mapping verify call
would fail because the result size is not the same as the inputs like
in a copy or phi.
The AMDGPU testcases are half broken and introduce illegal VGPR->SGPR
copies which need much more work to handle correctly (same for phis),
but add them as a baseline.
llvm-svn: 356713
This adds support for scalarizing these intrinsics as well the X86TargetTransformInfo support to avoid scalarizing them in the cases X86 can handle.
I've omitted handling special cases for constant masks for this first pass. Though CodeGenPrepare can constant fold the branch conditions and remove some of the control flow anyway.
Fixes PR40994 and is covers most of PR3666. Might want to implement constant masks to close that.
Differential Revision: https://reviews.llvm.org/D59180
llvm-svn: 356687
SDNodes can only have 64k operands and for some inputs (e.g. large
number of stores), we can reach this limit when creating TokenFactor
nodes. This patch is a follow up to D56740 and updates a few more places
that potentially can create TokenFactors with too many operands.
Reviewers: efriedma, craig.topper, aemerson, RKSimon
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D59156
llvm-svn: 356668
This is probably a bigger limitation than necessary, but since we don't have any evidence yet
that this transform led to real-world perf improvements rather than regressions, I'm making a
quick, blunt fix.
In the motivating x86 example from:
https://bugs.llvm.org/show_bug.cgi?id=41129
...and shown in the regression test, we want to avoid an extra instruction in the dominating
block because that could be costly.
The x86 LSR test diff is reversing the changes from D57789. There's no evidence that 1 version
is any better than the other yet.
Differential Revision: https://reviews.llvm.org/D59602
llvm-svn: 356665
Machine DCE cannot remove a dead definition if there are non-dbg uses.
A use however can be in the same instruction:
dead %0 = INST %0
Such instructions sometimes created by Detect dead lanes pass.
Allow this instruction to be deleted despite the use if the only use
belongs to the same instruction.
Differential Revision: https://reviews.llvm.org/D59565
llvm-svn: 356619