MMX extraction often ends up as extract_i32(bitcast_v2i32(extract_i64(bitcast_v1i64(x86mmx v), 0)), 0) which fails to simplify on 32-bit targets as i64 isn't legal
llvm-svn: 296782
This patch reduces the stack frame size by not allocating the parameter area if
it is not required. In the current implementation LowerFormalArguments_64SVR4
already handles the parameter area, but LowerCall_64SVR4 does not
(when calculating the stack frame size). What this patch does is make
LowerCall_64SVR4 consistent with LowerFormalArguments_64SVR4.
Committing on behalf of Hiroshi Inoue.
Differential Revision: https://reviews.llvm.org/D29881
llvm-svn: 296771
This bug was introduced with:
https://reviews.llvm.org/rL296699
There may be a way to loosen the restriction, but for now just bail out
on any opaque constant.
The tests show that opacity is target-specific. This goes back to cost
calculations in ConstantHoisting based on TTI->getIntImmCost().
llvm-svn: 296768
The CallingConv.td rules allocate 8 bytes for these kinds of arguments
on AAPCS targets, but we were only recording the smaller amount. The
difference is theoretical on AArch64 because we don't actually store
more than the smaller amount, but it's still much better to have these
two components in agreement.
Based on Diana Picus's ARM equivalent patch (where it matters a lot
more).
llvm-svn: 296754
If dominator tree is not calculated or is invalidated, set corresponding
pointer in the pass state to nullptr. Such pointer value will indicate
that operations with dominator tree are not allowed. In particular, it
allows to skip verification for such pass state. The dominator tree is
not calculated if the machine dominator pass was skipped, it occures in
the case of entities with linkage available_externally.
The change fixes some test fails observed when expensive checks
are enabled.
Differential Revision: https://reviews.llvm.org/D29280
llvm-svn: 296742
Surprisingly, one of the three interference checks in LiveRegMatrix was
using the main live range instead of the apropriate subregister range
resulting in unnecessarily conservative results.
llvm-svn: 296722
Original commit message:
[ARM] Fix insert point for store rescheduling.
In ARMPreAllocLoadStoreOpt::RescheduleOps, LastOp should be the last
operation which we want to merge. If we break out of the loop because
an operation has the wrong offset, we shouldn't use that operation as
LastOp.
This patch fixes some cases where we would sink stores for no reason.
llvm-svn: 296718
Summary:
This can be used to optimize large multiplications after legalization.
Depends on D29565
Reviewers: mkuper, spatel, RKSimon, zvi, bkramer, aaboud, craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29587
llvm-svn: 296711
Until now, we've had to use -global-isel to enable GISel. But using
that on other targets that don't support it will result in an abort, as we
can't build a full pipeline.
Additionally, we want to experiment with enabling GISel by default for
some targets: we can't just enable GISel by default, even among those
target that do have some support, because the level of support varies.
This first step adds an override for the target to explicitly define its
level of support. For AArch64, do that using
a new command-line option (I know..):
-aarch64-enable-global-isel-at-O=<N>
Where N is the opt-level below which GISel should be used.
Default that to -1, so that we still don't enable GISel anywhere.
We're not there yet!
While there, remove a couple LLVM_UNLIKELYs. Building the pipeline is
such a cold path that in practice that shouldn't matter at all.
llvm-svn: 296710
In ARMPreAllocLoadStoreOpt::RescheduleOps, LastOp should be the last
operation which we want to merge. If we break out of the loop because
an operation has the wrong offset, we shouldn't use that operation as
LastOp.
This patch fixes some cases where we would sink stores for no reason.
Differential Revision: https://reviews.llvm.org/D30124
llvm-svn: 296708
This code starts from the high end of the sorted vector of offsets, and
works backwards: it tries to find contiguous offsets, process them, then
pops them from the end of the vector. Most of the code agrees with this
order of processing, but one loop doesn't: it instead processes elements
from the low end of the vector (which are nodes with unrelated offsets).
Fix that loop to process the correct elements.
This has a few implications. One, we don't incorrectly return early when
processing multiple groups of offsets in the same block (which allows
rescheduling prera-ldst-insertpt.mir). Two, we pick the correct insert
point for loads, so they're correctly sorted (which affects the
scheduling of vldm-liveness.ll). I think it might also impact some of
the heuristics slightly.
Differential Revision: https://reviews.llvm.org/D30368
llvm-svn: 296701
This is part of the ongoing attempt to improve select codegen for all targets and select
canonicalization in IR (see D24480 for more background). The transform is a subset of what
is done in InstCombine's FoldOpIntoSelect().
I first noticed a regression in the x86 avx512-insert-extract.ll tests with a patch that
hopes to convert more selects to basic math ops. This appears to be a general missing DAG
transform though, so I added tests for all standard binops in rL296621
(PowerPC was chosen semi-randomly; it has scripted FileCheck support, but so do ARM and x86).
The poor output for "sel_constants_shl_constant" is tracked with:
https://bugs.llvm.org/show_bug.cgi?id=32105
Differential Revision: https://reviews.llvm.org/D30502
llvm-svn: 296699
Summary:
Avoids tons of prologue boilerplate when arguments are passed in memory
and left in memory. This can happen in a debug build or in a release
build when an argument alloca is escaped. This will dramatically affect
the code size of x86 debug builds, because X86 fast isel doesn't handle
arguments passed in memory at all. It only handles the x86_64 case of up
to 6 basic register parameters.
This is implemented by analyzing the entry block before ISel to identify
copy elision candidates. A copy elision candidate is an argument that is
used to fully initialize an alloca before any other possibly escaping
uses of that alloca. If an argument is a copy elision candidate, we set
a flag on the InputArg. If the the target generates loads from a fixed
stack object that matches the size and alignment requirements of the
alloca, the SelectionDAG builder will delete the stack object created
for the alloca and replace it with the fixed stack object. The load is
left behind to satisfy any remaining uses of the argument value. The
store is now dead and is therefore elided. The fixed stack object is
also marked as mutable, as it may now be modified by the user, and it
would be invalid to rematerialize the initial load from it.
Supersedes D28388
Fixes PR26328
Reviewers: chandlerc, MatzeB, qcolombet, inglorion, hans
Subscribers: igorb, llvm-commits
Differential Revision: https://reviews.llvm.org/D29668
llvm-svn: 296683
This patch adds a MachineSSA pass that coalesces blocks that branch
on the same condition.
Committing on behalf of Lei Huang.
Differential Revision: https://reviews.llvm.org/D28249
llvm-svn: 296670
Add check that deleted nodes do not get added to worklist. This can
occur when a node's operand is simplified to an existing node.
This fixes PR32108.
Reviewers: jyknight, hfinkel, chandlerc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30506
llvm-svn: 296668
Resubmit r295336 after the bug with non-zero offset patterns on BE targets is fixed (r296336).
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 296651
On Hexagon, values of type i1 are passed in registers of type i32,
even though i1 is not a legal value for these registers. This is a
special case and needs special handling to maintain consistency of
the lowering information.
This fixes PR32089.
llvm-svn: 296645
Lower i1, i8 and i16 call parameters by extending them before storing them on
the stack. Also make sure we encode the correct, extended size in the
corresponding memory operand, and that we compute the correct stack size in the
end.
The latter is a bit more complicated because we used to compute the stack size
in the getStackAddress method, based on the Size and Offset of the parameters.
However, if the last parameter is sign extended, we'd be using the wrong,
non-extended size, and we'd end up with a smaller stack than we need to hold the
extended value. Instead of hacking this up based on the value of Size in
getStackAddress, we move our stack size handling logic to assignArg, where we
have access to the CCState which knows everything we could possibly want to know
about the stack. This way we don't need to duplicate any knowledge or resort to
any ugly hacks.
On this same occasion, update the IRTranslator test to check the sizes of the
stores everywhere, not just for sign extended paramteres.
llvm-svn: 296631
Modify the test so that it is still testing something
closer to what it was intended to originally.
I think the original intent was to test the situation where
there was a branch on execz and then unconditional branch
required relaxing.With the change in r296539,
there was no longer and execz branch.
Change the test so that there is now an execz branch inserted.
There is no longer an unconditional branch after the execz branch,
so this might need to be tricked in some other way to keep that
there.
llvm-svn: 296574
When SDAGISel (top-down) selects a tail-call, it skips the remainder
of the block.
If, before that, FastISel (bottom-up) selected some of the (no-op) next
few instructions, we can end up with dead instructions following the
terminator (selected by SDAGISel).
We need to erase them, as we know they aren't necessary (in addition to
being incorrect).
We already do this when FastISel falls back on the tail-call itself.
Also remove the FastISel-emitted code if we fallback on the
instructions between the tail-call and the return.
llvm-svn: 296552
Iterating on the use-list we're modifying doesn't work: after the first
iteration, the use-list iterator will point to a MachineOperand
referencing the new register. This caused us to skip the other uses to
replace.
Instead, use MRI.replaceRegWith(), which accounts for this behavior.
llvm-svn: 296551
To facilitate this, add a new hidden command-line option to disable
the explicit-locals pass. That causes llc to emit invalid code that doesn't
have all locals converted to get_local/set_local, however it simplifies
testwriting in many cases.
llvm-svn: 296540
This prevents generating stm r1!, {r0, r1} on Thumb1, where value
stored for r1 is UNKONWN.
Patch by Zhaoshi Zheng.
Differential Revision: https://reviews.llvm.org/D27910
llvm-svn: 296538
Requesting DWARF v5 will now get you the new compile-unit and
type-unit headers. llvm-dwarfdump will also recognize them.
Differential Revision: http://reviews.llvm.org/D30206
llvm-svn: 296514
This recovers a test case that was severely broken by r296476, my making sure we don't create ADD/ADC that loads and stores when there is also a flag dependency.
llvm-svn: 296486
Stack Smash Protection is not completely free, so in hot code, the overhead it causes can cause performance issues. By adding diagnostic information for which functions have SSP and why, a user can quickly determine what they can do to stop SSP being applied to a specific hot function.
This change adds a remark that is reported by the stack protection code when an instruction or attribute is encountered that causes SSP to be applied.
Patch by: James Henderson
Differential Revision: https://reviews.llvm.org/D29023
llvm-svn: 296483
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* 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: 296476
Lower i32, float and double parameters that need to live on the stack. This
boils down to creating some G_GEPs starting from the stack pointer and storing
the values there. During the process we also keep track of the stack size and
use the final value in the ADJCALLSTACKDOWN/UP instructions.
We currently assert for smaller types, since they usually require extensions.
They will be handled in a separate patch.
llvm-svn: 296473
Summary:
With this change ImplicitNullCheck optimization uses alias analysis
and can use load/store memory access for implicit null check if there
are other load/store before but memory accesses do not alias.
Patch by Serguei Katkov!
Reviewers: sanjoy
Reviewed By: sanjoy
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30331
llvm-svn: 296440
This is a patch for the outliner described in the RFC at:
http://lists.llvm.org/pipermail/llvm-dev/2016-August/104170.html
The outliner is a code-size reduction pass which works by finding
repeated sequences of instructions in a program, and replacing them with
calls to functions. This is useful to people working in low-memory
environments, where sacrificing performance for space is acceptable.
This adds an interprocedural outliner directly before printing assembly.
For reference on how this would work, this patch also includes X86
target hooks and an X86 test.
The outliner is run like so:
clang -mno-red-zone -mllvm -enable-machine-outliner file.c
Patch by Jessica Paquette<jpaquette@apple.com>!
rdar://29166825
Differential Revision: https://reviews.llvm.org/D26872
llvm-svn: 296418
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296416
The transform in question claims to be doing:
// fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
...starting in PerformADDCombineWithOperands(), but it wasn't actually checking for a setcc node
for the sext/zext patterns.
This is exactly the opposite of a transform I'd like to add to DAGCombiner's foldSelectOfConstants(),
so I was seeing infinite loops with my draft of a patch applied.
The changes in select_const.ll look positive (less instructions). The change in arm-and-tst-peephole.ll
is unrelated. We're changing the input IR in that test to preserve the intent of the test, but that's
not affected by this code change.
Differential Revision:
https://reviews.llvm.org/D30355
llvm-svn: 296389
DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered.
This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it.
Differential Revision: https://reviews.llvm.org/D30176
llvm-svn: 296381
Summary: Existing implementation of duplicateSimpleBB function drops DebugLoc metadata of branch instructions during the transformation. This patch addresses this issue by making newly created branch instructions to keep the metadata of replaced branch instructions.
Reviewers: qcolombet, craig.topper, aprantl, MatzeB, sanjoy, dblaikie
Reviewed By: dblaikie
Subscribers: dblaikie, llvm-commits
Differential Revision: https://reviews.llvm.org/D30026
llvm-svn: 296371
This pattern is essentially a i16 load from p+1 address:
%p1.i16 = bitcast i8* %p to i16*
%p2.i8 = getelementptr i8, i8* %p, i64 2
%v1 = load i16, i16* %p1.i16
%v2.i8 = load i8, i8* %p2.i8
%v2 = zext i8 %v2.i8 to i16
%v1.shl = shl i16 %v1, 8
%res = or i16 %v1.shl, %v2
Current implementation would identify %v1 load as the first byte load and would mistakenly emit a i16 load from %p1.i16 address. This patch adds a check that the first byte is loaded from a non-zero offset of the first load address. This way this address can be used as the base address for the combined value. Otherwise just give up combining.
llvm-svn: 296336
Summary:
While collecting operands we make copies of the LiveReg objects which are stored in the LiveRegs array. If the instruction uses the same register multiple times we end up with multiple copies. Later we iterate through the collected list of LiveReg objects and merge DomainValues. In the process of doing this the merge function can change the contents of the original LiveReg object in the LiveRegs array, but not the copies that have been made. So when we get to the second usage of the register we end up seeing a stale copy of the LiveReg object.
To fix this I've stopped copying and now just store a pointer to the original LiveReg object. Another option might be to avoid adding the same register to the Regs array twice, but this approach seemed simpler.
The included test case exposes this bug due to an AVX-512 masked OR instruction using the same register for the passthru operand and one of the inputs to the OR operation.
Fixes PR30284.
Reviewers: RKSimon, stoklund, MatzeB, spatel, myatsina
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30242
llvm-svn: 296260
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* 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: 296252
With the "wasm32-unknown-unknown-wasm" triple, this allows writing out
simple wasm object files, and is another step in a larger series toward
migrating from ELF to general wasm object support. Note that this code
and the binary format itself is still experimental.
llvm-svn: 296190
This reverts commit r296009. It broke one out of tree target and also
does not account for all partial lines added or removed when calculating
PressureDiff.
llvm-svn: 296182
All G_CONSTANTS created by the MachineIRBuilder have an operand of type CImm
(i.e. a ConstantInt), so that's what the selector needs to look for.
llvm-svn: 296176
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296149
Provide a 64-bit pattern to use SUBFIC for subtracting from a 16-bit immediate.
The corresponding pattern already exists for 32-bit integers.
Committing on behalf of Hiroshi Inoue.
Differential Revision: https://reviews.llvm.org/D29387
llvm-svn: 296144
Emit clrrdi (extended mnemonic for rldicr) for AND-ing with masks that
clear bits from the right hand size.
Committing on behalf of Hiroshi Inoue.
Differential Revision: https://reviews.llvm.org/D29388
llvm-svn: 296143
The motivation for filling out these select-of-constants cases goes back to D24480,
where we discussed removing an IR fold from add(zext) --> select. And that goes back to:
https://reviews.llvm.org/rL75531https://reviews.llvm.org/rL159230
The idea is that we should always canonicalize patterns like this to a select-of-constants
in IR because that's the smallest IR and the best for value tracking. Note that we currently
do the opposite in some cases (like the cases in *this* patch). Ie, the proposed folds in
this patch already exist in InstCombine today:
https://github.com/llvm-mirror/llvm/blob/master/lib/Transforms/InstCombine/InstCombineSelect.cpp#L1151
As this patch shows, most targets generate better machine code for simple ext/add/not ops
rather than a select of constants. So the follow-up steps to make this less of a patchwork
of special-case folds and missing IR canonicalization:
1. Have DAGCombiner convert any select of constants into ext/add/not ops.
2 Have InstCombine canonicalize in the other direction (create more selects).
Differential Revision: https://reviews.llvm.org/D30180
llvm-svn: 296137
This time with the missing files.
Similar to PR/25526, fast-regalloc introduces spills at the end of basic
blocks. When this occurs in between an ll and sc, the store can cause the
atomic sequence to fail.
This patch fixes the issue by introducing more pseudos to represent atomic
operations and moving their lowering to after the expansion of postRA
pseudos.
This resolves PR/32020.
Thanks to James Cowgill for reporting the issue!
Reviewers: slthakur
Differential Revision: https://reviews.llvm.org/D30257
llvm-svn: 296134
Similar to PR/25526, fast-regalloc introduces spills at the end of basic
blocks. When this occurs in between an ll and sc, the store can cause the
atomic sequence to fail.
This patch fixes the issue by introducing more pseudos to represent atomic
operations and moving their lowering to after the expansion of postRA
pseudos.
This resolves PR/32020.
Thanks to James Cowgill for reporting the issue!
Reviewers: slthakur
Differential Revision: https://reviews.llvm.org/D30257
llvm-svn: 296132
Summary:
This isn't testable for AArch64 by itself so this patch also adds
support for constant immediates in the pattern and physical
register uses in the result.
The new IntOperandMatcher matches the constant in patterns such as
'(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold
immediates into an instruction so this is the first rule that will match
across multiple BB's.
The Renderer hierarchy is responsible for adding operands to the result
instruction. Renderers can copy operands (CopyRenderer) or add physical
registers (in particular %wzr and %xzr) to the result instruction
in any order (OperandMatchers now import the operand names from
SelectionDAG to allow renderers to access any operand). This allows us to
emit the result instruction for:
%1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0
%1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0
although the latter is untested since the matcher/importer has not been
taught about commutativity yet.
Added BuildMIAction which can build new instructions and mutate them where
possible. W.r.t the mutation aspect, MatchActions are now told the name of
an instruction they can recycle and BuildMIAction will emit mutation code
when the renderers are appropriate. They are appropriate when all operands
are rendered using CopyRenderer and the indices are the same as the matcher.
This currently assumes that all operands have at least one matcher.
Finally, this change also fixes a crash in
AArch64InstructionSelector::select() caused by an immediate operand
passing isImm() rather than isCImm(). This was uncovered by the other
changes and was detected by existing tests.
Depends on D29711
Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar
Reviewed By: rovka
Subscribers: aemerson, dberris, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D29712
llvm-svn: 296131
The Fuchsia ABI defines slots from the thread pointer where the
stack-guard value for stack-protector, and the unsafe stack pointer
for safe-stack, are stored. This parallels the Android ABI support.
Patch by Roland McGrath
Differential Revision: https://reviews.llvm.org/D30237
llvm-svn: 296081
clang will generate IR like this for input using packed bitfields;
very simple semantically, but it's a bit tricky to actually
generate good code.
llvm-svn: 296080
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296060
This patch enables support for .f16x2 operations.
Added new register type Float16x2.
Added support for .f16x2 instructions.
Added handling of vectorized loads/stores of v2f16 values.
Differential Revision: https://reviews.llvm.org/D30057
Differential Revision: https://reviews.llvm.org/D30310
llvm-svn: 296032
FastISel wasn't checking the isFPOnlySP subtarget feature before emitting
double-precision operations, so it got completely invalid CodeGen for doubles
on Cortex-M4F.
The normal ISel testing wasn't spectacular either so I added a second RUN line
to improve that while I was in the area.
llvm-svn: 296031
Having more fine-grained information on the specific construct that
caused us to fallback is valuable for large-scale data collection.
We still have the fallback warning, that's also used for FastISel.
We still need to remove the fallback warning, and teach FastISel to also
emit remarks (it currently has a combination of the warning, stats, and
debug prints: the remarks could unify all three).
The abort-on-fallback path could also be better handled using remarks:
one could imagine a "-Rpass-error", analoguous to "-Werror", which would
promote missed/failed remarks to errors. It's not clear whether that
would be useful for other remarks though, so we're not there yet.
llvm-svn: 296013
If a subreg is used in an instruction it counts as a whole superreg
for the purpose of register pressure calculation. This patch corrects
improper register pressure calculation by examining operand's lane mask.
Differential Revision: https://reviews.llvm.org/D29835
llvm-svn: 296009
Introduce a common ValueHandler for call returns and formal arguments, and
inherit two different versions for handling the differences (at the moment the
only difference is the way physical registers are marked as used).
llvm-svn: 295973
Add support for lowering calls with parameters than can fit into regs. Use the
same ValueHandler that we used for function returns, but rename it to match its
new, extended purpose.
llvm-svn: 295971
The ARMConstantIslandPass didn't have support for handling accesses to
constant island objects through ARM::t2LDRBpci instructions. This adds
support for that.
This fixes PR31997.
llvm-svn: 295964
AVX versions of the converts work on f32/f64 types, while AVX512 version work on vectors.
Differential Revision: https://reviews.llvm.org/D29988
llvm-svn: 295940
The manual is unclear on the details of this. It's not
clear to me if denormals are not allowed with clamp,
or if that is only omod. Not allowing denorms for
fp16 or fp64 isn't useful so I also question if that
is really a restriction. Same with whether this is valid
without IEEE mode enabled.
llvm-svn: 295905
This should avoid reporting any stack needs to be allocated in the
case where no stack is truly used. An unused stack slot is still
left around in other cases where there are real stack objects
but no spilling occurs.
llvm-svn: 295891
This allows us to ensure that 0 is never a valid pointer
to a user object, and ensures that the offset is always legal
without needing a register to access it. This comes at the cost
of usable offsets and wasted stack space.
llvm-svn: 295877
Summary:
Extend AArch64RedundantCopyElimination to catch cases where the register
that is known to be zero is COPY'd in the predecessor block. Before
this change, this pass would catch cases like:
CBZW %W0, <BB#1>
BB#1:
%W0 = COPY %WZR // removed
After this change, cases like the one below are also caught:
%W0 = COPY %W1
CBZW %W1, <BB#1>
BB#1:
%W0 = COPY %WZR // removed
This change results in a 4% increase in static copies removed by this
pass when compiling the llvm test-suite. It also fixes regressions
caused by doing post-RA copy propagation (a separate change to be put up
for review shortly).
Reviewers: junbuml, mcrosier, t.p.northover, qcolombet, MatzeB
Subscribers: aemerson, rengolin, llvm-commits
Differential Revision: https://reviews.llvm.org/D30113
llvm-svn: 295863
The pass tries to fix a spill of LR that turns out to be unnecessary.
So it removes the tPOP but forgets to remove tPUSH.
This causes the stack be misaligned upon returning the function.
Thus, remove the tPUSH as well in this case.
Differential Revision: https://reviews.llvm.org/D30207
llvm-svn: 295816
This patch adds missing sched classes for Thumb2 instructions.
This has been missing so far, and as a consequence, machine
scheduler models for individual sub-targets have tended to
be larger than they needed to be. These patches should help
write schedulers better and faster in the future
for ARM sub-targets.
Reviewer: Diana Picus
Differential Revision: https://reviews.llvm.org/D29953
llvm-svn: 295811
This patch introduces new X86ISD::FMAXS and X86ISD::FMINS opcodes. The legacy intrinsics now lower to this node. As do the AVX-512 masked intrinsics when the rounding mode is CUR_DIRECTION.
I've merged a copy of the tablegen multiclass avx512_fp_scalar into avx512_fp_scalar_sae. avx512_fp_scalar still needs to support CUR_DIRECTION appearing as a rounding mode for X86ISD::FADD_ROUND and others.
Differential revision: https://reviews.llvm.org/D30186
llvm-svn: 295810
Change implementation to use max instead of add.
min/max/med3 do not flush denormals regardless of the mode,
so it is OK to use it whether or not they are enabled.
Also allow using clamp with f16, and use knowledge
of dx10_clamp.
llvm-svn: 295788
Original code only used vector loads/stores for explicit vector arguments.
It could also do more loads/stores than necessary (e.g v5f32 would
touch 8 f32 values). Aggregate types were loaded one element at a time,
even the vectors contained within.
This change attempts to generalize (and simplify) parameter space
loads/stores so that vector loads/stores can be used more broadly.
Functionality of the patch has been verified by compiling thrust
test suite and manually checking the differences between PTX
generated by llvm with and without the patch.
General algorithm:
* ComputePTXValueVTs() flattens input/output argument into a flat list
of scalars to load/store and returns their types and offsets.
* VectorizePTXValueVTs() uses that data to create vectorization plan
which returns an array of flags marking boundaries of vectorized
load/stores. Scalars are represented as 1-element vectors.
* Code that generates loads/stores implements a simple state machine
that constructs a vector according to the plan.
Differential Revision: https://reviews.llvm.org/D30011
llvm-svn: 295784
Before frame offsets are calculated, try to eliminate the
frame indexes used by SGPR spills. Then we can delete them
after.
I think for now we can be sure that no other instruction
will be re-using the same frame indexes. It should be easy
to notice if this assumption ever breaks since everything
asserts if it tries to use a dead frame index later.
The unused emergency stack slot seems to still be left behind,
so an additional 4 bytes is still wasted.
llvm-svn: 295753
Summary:
Rework the code that was sinking/duplicating (icmp and, 0) sequences
into blocks where they were being used by conditional branches to form
more tbz instructions on AArch64. The new code is more general in that
it just looks for 'and's that have all icmp 0's as users, with a target
hook used to select which subset of 'and' instructions to consider.
This change also enables 'and' sinking for X86, where it is more widely
beneficial than on AArch64.
The 'and' sinking/duplicating code is moved into the optimizeInst phase
of CodeGenPrepare, where it can take advantage of the fact the
OptimizeCmpExpression has already sunk/duplicated any icmps into the
blocks where they are used. One minor complication from this change is
that optimizeLoadExt needed to be updated to always mark 'and's it has
determined should be in the same block as their feeding load in the
InsertedInsts set to avoid an infinite loop of hoisting and sinking the
same 'and'.
This change fixes a regression on X86 in the tsan runtime caused by
moving GVNHoist to a later place in the optimization pipeline (see
PR31382).
Reviewers: t.p.northover, qcolombet, MatzeB
Subscribers: aemerson, mcrosier, sebpop, llvm-commits
Differential Revision: https://reviews.llvm.org/D28813
llvm-svn: 295746
As i64 isn't a value type on 32-bit targets, we fail to fold the VZEXT_LOAD into VPBROADCASTQ.
Also shows that we're not decoding VPERMIV3 shuffles very well....
llvm-svn: 295729
This matches what is already done during shuffle lowering and helps prevent the need for a zero-vector in cases where shuffles match both patterns.
llvm-svn: 295723
Currently just contains one case where we combine to VZEXT_MOVL instead of VZEXT which would avoid the need for a zero vector to be generated
llvm-svn: 295721
They are all covered by the SSE4.2 intrinsics test with SSE4.2, AVX, and AVX512 command lines.
Merge sse42.ll into the other intrinsics test. Rename sse42_64.ll to be named like other intrinsic tests.
llvm-svn: 295707
They are all covered by the SSE2 intrinsics test with SSE2, AVX, and AVX512 command lines.
Also remove an unneeded lfence intrinsic test since it was already covered.
llvm-svn: 295700
They are all covered by the SSE intrinsics test with SSE, AVX, and AVX512 command lines.
Also remove an unneeded sfence intrinsic test since it was already covered.
llvm-svn: 295699
Summary:
Sandy Bridge and later CPUs have better throughput using a SHLD to implement rotate versus the normal rotate instructions. Additionally it saves one uop and avoids a partial flag update dependency.
This patch implements this change on any Sandy Bridge or later processor without BMI2 instructions. With BMI2 we will use RORX as we currently do.
Reviewers: zvi
Reviewed By: zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30181
llvm-svn: 295697
Summary:
Currently, BranchFolder drops DebugLoc for branch instructions in some places. For example, for the test code attached, the branch instruction of 'entry' block has a DILocation of
```
!12 = !DILocation(line: 6, column: 3, scope: !11)
```
, but this information is gone when then block is lowered because BranchFolder misses it. This patch is a fix for this issue.
Reviewers: qcolombet, aprantl, craig.topper, MatzeB
Reviewed By: aprantl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29902
llvm-svn: 295684
Its more profitable to go through memory (1 cycles throughput)
than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index.
IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer)
For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles.
Removing the VINSERT node, we don't need it any more.
Differential Revision: https://reviews.llvm.org/D29690
llvm-svn: 295660
Summary:
This file was missed in the commit for Cortex-M23 and Cortex-M33
support. See https://reviews.llvm.org/D29073?id=85814 .
Reviewers: rengolin, javed.absar, samparker
Reviewed By: samparker
Subscribers: llvm-commits, aemerson
Differential Revision: https://reviews.llvm.org/D30162
llvm-svn: 295655
Replaces existing approach that could only search BUILD_VECTOR nodes.
Requires getTargetConstantBitsFromNode to discriminate cases with all/partial UNDEF bits in each element - this should also be useful when we get around to supporting getTargetShuffleMaskIndices with UNDEF elements.
llvm-svn: 295613
As discussed on D27692, this permits another domain to be used to combine a shuffle at high depths.
We currently set the required depth at 4 or more combined shuffles, this is probably too high for most targets but is a good starting point and already helps avoid a number of costly variable shuffles.
llvm-svn: 295608
The instructions are marked commutable, but without special handling we don't get the immediate correct.
While here also remove the masked memory forms that aren't commutable.
llvm-svn: 295602
It seems we were already upgrading 128-bit VPCMOV, but the intrinsic was still defined and being used in isel patterns. While I was here I also simplified the tablegen multiclasses.
llvm-svn: 295564
- Adapt MachineBasicBlock::getName() to have the same behavior as the IR
BasicBlock (Value::getName()).
- Add it to lib/CodeGen/CodeGen.cpp::initializeCodeGen so that it is linked in
the CodeGen library.
- MachineRegionInfoPass's name conflicts with RegionInfoPass's name ("region").
- MachineRegionInfo should depend on MachineDominatorTree,
MachinePostDominatorTree and MachineDominanceFrontier instead of their
respective IR versions.
- Since there were no tests for this, add a X86 MIR test.
Patch by Francis Visoiu Mistrih<fvisoiumistrih@apple.com>
llvm-svn: 295518
A line number doesn't make much sense if you don't say where it's
from. Add a verifier check for this and update some tests that had
bogus debug info.
llvm-svn: 295516
When promoting the Load of a Store-Load pair to a COPY all kill flags
between the store and the load need to be cleared.
rdar://30402435
Differential Revision: https://reviews.llvm.org/D30110
llvm-svn: 295512
Removed the HasT2ExtractPack feature and replaced its references
with HasDSP. This then allows the Thumb2 extend instructions to be
selected for ARMv8M +dsp. These instruction descriptions have also
been refactored and more target tests have been added for their isel.
Differential Revision: https://reviews.llvm.org/D29623
llvm-svn: 295452
During legalization we are often creating shuffles (via a build_vector scalarization stage) that are "any_extend_vector_inreg" style masks, and also other masks that are the equivalent of "truncate_vector_inreg" (if we had such a thing).
This patch is an attempt to match these cases to help undo the effects of just leaving shuffle lowering to handle it - which typically means we lose track of the undefined elements of the shuffles resulting in an unnecessary extension+truncation stage for widened illegal types.
The 2011-10-21-widen-cmp.ll regression will be fixed by making SIGN_EXTEND_VECTOR_IN_REG legal in SSE instead of lowering them to X86ISD::VSEXT (PR31712).
Differential Revision: https://reviews.llvm.org/D29454
llvm-svn: 295451
Start using the Subtarget to make decisions about what's legal. In particular,
we only mark floating point operations as legal if we have VFP2, which is
something we should've done from the very start.
llvm-svn: 295439
The original commit was reverted in r283329 due to a miscompile in
Chromium. That turned out to be the same issue as PR31257, which was
fixed in r295262.
llvm-svn: 295357
The existing code always saves the xmm registers for 64-bit targets even if the
target doesn't support SSE (which is common for kernels). Thus, the compiler
inserts movaps instructions which lead to CPU exceptions when an interrupt
handler is invoked.
This commit fixes this bug by returning a register set without xmm registers
from getCalleeSavedRegs and getCallPreservedMask for such targets.
Patch by Philipp Oppermann.
Differential Revision: https://reviews.llvm.org/D29959
llvm-svn: 295347
Resubmit -r295314 with PowerPC and AMDGPU tests updated.
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295336
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295314
Since they're only used for passing around double precision floating point
values into the general purpose registers, we'll lower them to VMOVDRR and
VMOVRRD.
llvm-svn: 295310
For now we just mark them as legal all the time and let the other passes bail
out if they can't handle it. In the future, we'll want to move more of the
brains into the legalizer.
llvm-svn: 295300
For the hard float calling convention, we just use the D registers.
For the soft-fp calling convention, we use the R registers and move values
to/from the D registers by means of G_SEQUENCE/G_EXTRACT. While doing so, we
make sure to honor the endianness of the target, since the CCAssignFn doesn't do
that for us.
For pure soft float targets, we still bail out because we don't support the
libcalls yet.
llvm-svn: 295295
The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO.
llvm-svn: 295290
This reverts r294348, which removed support for conditional tail calls
due to the PR above. It fixes the PR by marking live registers as
implicitly used and defined by the now predicated tailcall. This is
similar to how IfConversion predicates instructions.
Differential Revision: https://reviews.llvm.org/D29856
llvm-svn: 295262
Uses a Custom implementation because the slot sizes being a multiple of the
pointer size isn't really universal, even for the architectures that do have a
simple "void *" va_list.
llvm-svn: 295255
Fixes PR 31921
Summary:
Predicateinfo requires an ugly workaround to try to avoid literal
struct types due to the intrinsic mangling not being implemented.
This workaround actually does not work in all cases (you can hit the
assert by bootstrapping with -print-predicateinfo), and can't be made
to work without DFS'ing the type (IE copying getMangledStr and using a
version that detects if it would crash).
Rather than do that, i just implemented the mangling. It seems
simple, since they are unified structurally.
Looking at the overloaded-mangling testcase we have, it actually turns
out the gc intrinsics will *also* crash if you try to use a literal
struct. Thus, the testcase added fails before this patch, and works
after, without needing to resort to predicateinfo.
Reviewers: chandlerc, davide
Subscribers: llvm-commits, sanjoy
Differential Revision: https://reviews.llvm.org/D29925
llvm-svn: 295253
Lay out trellis-shaped CFGs optimally.
A trellis of the shape below:
A B
|\ /|
| \ / |
| X |
| / \ |
|/ \|
C D
would be laid out A; B->C ; D by the current layout algorithm. Now we identify
trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an
increasing number of predecessors. A trellis is a a group of 2 or more
predecessor blocks that all have the same successors.
because of this we can tail duplicate to extend existing trellises.
As an example consider the following CFG:
B D F H
/ \ / \ / \ / \
A---C---E---G---Ret
Where A,C,E,G are all small (Currently 2 instructions).
The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret.
The current code will copy C into B, E into D and G into F and yield the layout
A,C,B(C),E,D(E),F(G),G,H,ret
define void @straight_test(i32 %tag) {
entry:
br label %test1
test1: ; A
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %optional1
optional1: ; B
call void @a()
br label %test2
test2: ; C
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %test3, label %optional2
optional2: ; D
call void @b()
br label %test3
test3: ; E
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
br i1 %tagbit3eq0, label %test4, label %optional3
optional3: ; F
call void @c()
br label %test4
test4: ; G
%tagbit4 = and i32 %tag, 8
%tagbit4eq0 = icmp eq i32 %tagbit4, 0
br i1 %tagbit4eq0, label %exit, label %optional4
optional4: ; H
call void @d()
br label %exit
exit:
ret void
}
here is the layout after D27742:
straight_test: # @straight_test
; ... Prologue elided
; BB#0: # %entry ; A (merged with test1)
; ... More prologue elided
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_2
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_3
b .LBB0_4
.LBB0_2: # %optional1 ; B (copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_4
.LBB0_3: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_5
b .LBB0_6
.LBB0_4: # %optional2 ; D (copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_5: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
b .LBB0_7
.LBB0_6: # %optional3 ; F (copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit ; Ret
ld 30, 96(1) # 8-byte Folded Reload
addi 1, 1, 112
ld 0, 16(1)
mtlr 0
blr
The tail-duplication has produced some benefit, but it has also produced a
trellis which is not laid out optimally. With this patch, we improve the layouts
of such trellises, and decrease the cost calculation for tail-duplication
accordingly.
This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have
back edges, which is a negative, but it has a bigger compensating
positive, which is that it handles the case where there are long strings
of skipped blocks much better than the original layout. Both layouts
handle runs of executed blocks equally well. Branch prediction also
improves if there is any correlation between subsequent optional blocks.
Here is the resulting concrete layout:
straight_test: # @straight_test
; BB#0: # %entry ; A (merged with test1)
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_4
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_5
.LBB0_2: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_3: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
bne 0, .LBB0_7
b .LBB0_8
.LBB0_4: # %optional1 ; B (Copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_2
.LBB0_5: # %optional2 ; D (Copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_3
.LBB0_6: # %optional3 ; F (Copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit
Differential Revision: https://reviews.llvm.org/D28522
llvm-svn: 295223
We currently can't legalize those, but we should really not be creating
them in the first place, since legalization would probably look similar to the
way we legalize CONCAT_VECTORS - basically replace the INSERT with a BUILD.
This fixes PR311956.
Differential Revision: https://reviews.llvm.org/D29961
llvm-svn: 295213
This patch reverts region's scheduling to the original untouched state
in case if we have have decreased occupancy.
In addition it switches to use TargetRegisterInfo occupancy callback
for pressure limits instead of gradually increasing limits which were
just passed by. We are going to stay with the best schedule so we do
not need to tolerate worsened scheduling anymore.
Differential Revision: https://reviews.llvm.org/D29971
llvm-svn: 295206
Summary:
We don't seem to have great rules on what a valid VBROADCAST node looks like. And as a consequence we end up with a lot of patterns to try to catch everything. We have patterns with scalar inputs, 128-bit vector inputs, 256-bit vector inputs, and 512-bit vector inputs.
As you can see from the things improved here we are currently missing patterns for 128-bit loads being extended to 256-bit before the vbroadcast.
I'd like to propose that VBROADCAST should always take a 128-bit vector type as input. As a first step towards that this patch adds an EXTRACT_SUBVECTOR in front of VBROADCAST when the input is 256 or 512-bits. In the future I would like to add scalar_to_vector around all the scalar operations. And maybe we should consider adding a VBROADCAST+load node to avoid separating loads from the broadcasting operation when the load itself isn't foldable.
This requires an additional change in target shuffle combining to look for the extract subvector and look through it to find the original operand. I'm sure this change isn't perfect but was enough to fix a few test failures that were being caused.
Another interesting thing I noticed is that the changes in masked_gather_scatter.ll show cases were we don't remove a useless insert into element 1 before broadcasting element 0.
Reviewers: delena, RKSimon, zvi
Reviewed By: zvi
Subscribers: igorb, llvm-commits
Differential Revision: https://reviews.llvm.org/D28747
llvm-svn: 295155
This patch corrects the maximum workgroups per CU if we have big
workgroups (more than 128). This calculation contributes to the
occupancy calculation in respect to LDS size.
Differential Revision: https://reviews.llvm.org/D29974
llvm-svn: 295134
Summary:
Blocks ending in unreachable are typically cold because they end the
program or throw an exception, so merging them with other identical
blocks is usually profitable because it reduces the size of cold code.
MachineBlockPlacement generally does not arrange to fall through to such
blocks, so commoning these blocks will not introduce additional
unconditional branches.
Reviewers: hans, iteratee, haicheng
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29153
llvm-svn: 295105
This instruction clears the low bits of a pointer without requiring (possibly
dodgy if pointers aren't ints) conversions to and from an integer. Since (as
far as I'm aware) all masks are statically known, the instruction takes an
immediate operand rather than a register to specify the mask.
llvm-svn: 295103
Simon Pilgrim