Re-apply of r241928 which had to be reverted because of the r241926
revert.
This commit factors out common code from MergeBaseUpdateLoadStore() and
MergeBaseUpdateLSMultiple() and introduces a new function
MergeBaseUpdateLSDouble() which merges adds/subs preceding/following a
strd/ldrd instruction into an strd/ldrd instruction with writeback where
possible.
Differential Revision: http://reviews.llvm.org/D10676
llvm-svn: 242743
Re-apply r241926 with an additional check that r13 and r15 are not used
for LDRD/STRD. See http://llvm.org/PR24190. This also already includes
the fix from r241951.
Differential Revision: http://reviews.llvm.org/D10623
llvm-svn: 242742
whether register r9 should be reserved.
This change is needed because we cannot use a backend option to set
cl::opt "arm-reserve-r9" when doing LTO.
Out-of-tree projects currently using cl::opt option "-arm-reserve-r9" to
reserve r9 should make changes to add subtarget feature "reserve-r9" to
the IR.
rdar://problem/21529937
Differential Revision: http://reviews.llvm.org/D11320
llvm-svn: 242737
Even though this is just some hinting for the scheduler it doesn't make
sense to do that unless you know the target can perform the fusion.
llvm-svn: 242732
This patch does the following:
* Fix FIXME on `needsStackRealignment`: it is now shared between multiple targets, implemented in `TargetRegisterInfo`, and isn't `virtual` anymore. This will break out-of-tree targets, silently if they used `virtual` and with a build error if they used `override`.
* Factor out `canRealignStack` as a `virtual` function on `TargetRegisterInfo`, by default only looks for the `no-realign-stack` function attribute.
Multiple targets duplicated the same `needsStackRealignment` code:
- Aarch64.
- ARM.
- Mips almost: had extra `DEBUG` diagnostic, which the default implementation now has.
- PowerPC.
- WebAssembly.
- x86 almost: has an extra `-force-align-stack` option, which the default implementation now has.
The default implementation of `needsStackRealignment` used to just return `false`. My current patch changes the behavior by simply using the above shared behavior. This affects:
- AMDGPU
- BPF
- CppBackend
- MSP430
- NVPTX
- Sparc
- SystemZ
- XCore
- Out-of-tree targets
This is a breaking change! `make check` passes.
The only implementation of the `virtual` function (besides the slight different in x86) was Hexagon (which did `MF.getFrameInfo()->getMaxAlignment() > 8`), and potentially some out-of-tree targets. Hexagon now uses the default implementation.
`needsStackRealignment` was being overwritten in `<Target>GenRegisterInfo.inc`, to return `false` as the default also did. That was odd and is now gone.
Reviewers: sunfish
Subscribers: aemerson, llvm-commits, jfb
Differential Revision: http://reviews.llvm.org/D11160
llvm-svn: 242727
This is the first step toward supporting shrink-wrapping for this target.
The changes could be summarized by these items:
- Expand the tail-call return as part of the expand pseudo pass.
- Get rid of the assumptions that the epilogue is the exit block:
* Do not assume which registers are free in the epilogue. (This indirectly
improve the lowering of the code for the segmented stacks, see the test
cases.)
* Take into account that the basic block can be empty.
Related to <rdar://problem/20821730>
llvm-svn: 242714
Summary:
[NVPTX] make load on global readonly memory to use ldg
Summary:
As describe in [1], ld.global.nc may be used to load memory by nvcc when
__restrict__ is used and compiler can detect whether read-only data cache
is safe to use.
This patch will try to check whether ldg is safe to use and use them to
replace ld.global when possible. This change can improve the performance
by 18~29% on affected kernels (ratt*_kernel and rwdot*_kernel) in
S3D benchmark of shoc [2].
Patched by Xuetian Weng.
[1] http://docs.nvidia.com/cuda/kepler-tuning-guide/#read-only-data-cache
[2] https://github.com/vetter/shoc
Test Plan: test/CodeGen/NVPTX/load-with-non-coherent-cache.ll
Reviewers: jholewinski, jingyue
Subscribers: jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D11314
llvm-svn: 242713
Summary:
This change generalizes the implicit null checks pass to work with
instructions that don't have any explicit register defs. This lets us
use X86's `cmp` against memory as faulting load instructions.
Reviewers: reames, JosephTremoulet
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11286
llvm-svn: 242703
Summary:
The MUBUF addr64 bit has been removed on VI, so we must use FLAT
instructions when the pointer is stored in VGPRs.
Reviewers: arsenm
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11067
llvm-svn: 242673
Reordered the data tables at the top and placed the lookups after. The first stage in the yak shaving necessary to get more accurate costs for a variety of targets given the recent improvements to SINT_TO_FP/UINT_TO_FP/SIGN_EXTEND vector lowering.
llvm-svn: 242643
canFoldMemoryOperand is not actually used anywhere in the codebase - all existing users instead call foldMemoryOperand directly when they wish to fold and can correctly deduce what they need from the return value.
This patch removes the canFoldMemoryOperand base function and the target implementations; only x86 had a real (bit-rotted) implementation, although AMDGPU had a preparatory stub that had never needed to be completed.
Differential Revision: http://reviews.llvm.org/D11331
llvm-svn: 242638
SKX supports conversion for all FP types. Integer types include doublewords and quardwords.
I added "Legal" status for these nodes and a bunch of tests.
I added "NoVLX" for AVX DAG selection to force VLX instructions selection when VLX is supported.
Differential Revision: http://reviews.llvm.org/D11255
llvm-svn: 242637
The standard containers are not designed to be inherited from, as
illustrated by the MSVC hacks for NodeOrdering. No functional change
intended.
llvm-svn: 242616
Reapply r242500 now that the swift schedmodel includes LDRLIT.
This is mostly done to disable the PostRAScheduler which optimizes for
instruction latencies which isn't a good fit for out-of-order
architectures. This also allows to leave out the itinerary table in
swift in favor of the SchedModel ones.
This change leads to performance improvements/regressions by as much as
10% in some benchmarks, in fact we loose 0.4% performance over the
llvm-testsuite for reasons that appear to be unknown or out of the
compilers control. rdar://20803802 documents the investigation of
these effects.
While it is probably a good idea to perform the same switch for the
other ARM out-of-order CPUs, I limited this change to swift as I cannot
perform the benchmark verification on the other CPUs.
Differential Revision: http://reviews.llvm.org/D10513
llvm-svn: 242588
These pseudo instructions are only lowered after register allocation and
are therefore still present when the machine scheduler runs.
Add a run: line to a testcase that uses the uncommon flags necessary to
actually produce a LDRLIT instruction on swift.
llvm-svn: 242587
This is mostly done to disable the PostRAScheduler which optimizes for
instruction latencies which isn't a good fit for out-of-order
architectures. This also allows to leave out the itinerary table in
swift in favor of the SchedModel ones.
This change leads to performance improvements/regressions by as much as
10% in some benchmarks, in fact we loose 0.4% performance over the
llvm-testsuite for reasons that appear to be unknown or out of the
compilers control. rdar://20803802 documents the investigation of
these effects.
While it is probably a good idea to perform the same switch for the
other ARM out-of-order CPUs, I limited this change to swift as I cannot
perform the benchmark verification on the other CPUs.
Differential Revision: http://reviews.llvm.org/D10513
llvm-svn: 242500
Constructing a name based on the function name didn't give us a unique
symbol if we had more than one setjmp in a function. Using
MCContext::createTempSymbol() always gives us a unique name.
Differential Revision: http://reviews.llvm.org/D9314
llvm-svn: 242482
llvm.eh.sjlj.setjmp was used as part of the SjLj exception handling
style but is also used in clang to implement __builtin_setjmp. The ARM
backend needs to output additional dispatch tables for the SjLj
exception handling style, these tables however can't be emitted if
llvm.eh.sjlj.setjmp is simply used for __builtin_setjmp and no actual
landing pad blocks exist.
To solve this issue a new llvm.eh.sjlj.setup_dispatch intrinsic is
introduced which is used instead of llvm.eh.sjlj.setjmp in the SjLj
exception handling lowering, so we can differentiate between the case
where we actually need to setup a dispatch table and the case where we
just need the __builtin_setjmp semantic.
Differential Revision: http://reviews.llvm.org/D9313
llvm-svn: 242481
C11 leaves the choice on whether round-to-integer operations set the inexact
flag implementation-defined. Darwin does expect it to be set, but this seems to
be against the intent of the IEEE document and slower to implement anyway. So
it should be opt-in.
llvm-svn: 242446
I was looking at some vector code generation and kept seeing
unnecessary vector copies into the Altivec half of the VSX registers.
I discovered that we overlooked v4i32 when adding the register classes
for VSX; we only added v4f32 and v2f64. This means that anything that
canonicalizes into v4i32 (which is a LOT of stuff) ends up being
forced into VRRC on its way to VSRC.
The fix is one line. The rest of the patch is fixing up some test
cases whose code generation has changed as a result.
This seems like it would be a good candidate for backport to 3.7.
llvm-svn: 242442
Summary:
SpeculativeExecution enables a series straight line optimizations (such
as SLSR and NaryReassociate) on conditional code. For example,
if (...)
... b * s ...
if (...)
... (b + 1) * s ...
speculative execution can hoist b * s and (b + 1) * s from then-blocks,
so that we have
... b * s ...
if (...)
...
... (b + 1) * s ...
if (...)
...
Then, SLSR can rewrite (b + 1) * s to (b * s + s) because after
speculative execution b * s dominates (b + 1) * s.
The performance impact of this change is significant. It speeds up the
benchmarks running EigenFloatContractionKernelInternal16x16
(ba68f42fa6/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h (cl-526))
by roughly 2%. Some internal benchmarks that have the above code pattern
are improved by up to 40%. No significant slowdowns are observed on
Eigen CUDA microbenchmarks.
Reviewers: jholewinski, broune, eliben
Subscribers: llvm-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D11201
llvm-svn: 242437
This is a new iteration of the reverted r238793 /
http://reviews.llvm.org/D8232 which wrongly assumed that any and/or
trees can be represented by conditional compare sequences, however there
are some restrictions to that. This version fixes this and adds comments
that explain exactly what types of and/or trees can actually be
implemented as conditional compare sequences.
Related to http://llvm.org/PR20927, rdar://18326194
Differential Revision: http://reviews.llvm.org/D10579
llvm-svn: 242436
Summary:
We can safely assume that the high bit of scratch offsets will never
be set, because this would require at least 128 GB of GPU memory.
Reviewers: arsenm
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11225
llvm-svn: 242433
This reverts commit r242300.
This is causing buildbot failures which we are investigating.
I'll reapply once we know whats going on, but for now want to
get the bots green.
llvm-svn: 242428
Summary:
This fixes an issue on MIPS where the infinite-loop-evergreen.ll test
was failing to terminate.
Fixes PR24147.
Reviewers: arsenm, dsanders
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11260
llvm-svn: 242410
This allows more call sequences to use pushes instead of movs when optimizing for size.
In particular, calling conventions that pass some parameters in registers (e.g. thiscall) are now supported.
This should no longer cause miscompiles, now that a bug in emitPrologue was fixed in r242395.
llvm-svn: 242398
When X86FrameLowering::emitPrologue() looks for where to insert the %esp subtraction
to allocate stack space for local allocations, it assumes that any sequence of push
instructions that starts at function entry consists purely of spills of callee-save
registers.
This may be false, since from some point forward, the pushes may pushing arguments
to a subsequent function call.
This caused a miscompile that was exposed by r240257, and is not easily testable
since r240257 was reverted. A test will be committed separately after r240257 is
reapplied.
llvm-svn: 242395
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.
This patch is quite boring overall, except for some uglyness in
ASMPrinter which has a getDataLayout function but has some clients
that use it without a Module (llmv-dsymutil, llvm-dwarfdump), so
some methods are taking a DataLayout as parameter.
Reviewers: echristo
Subscribers: yaron.keren, rafael, llvm-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D11090
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 242386
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: yaron.keren, rafael, llvm-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D11079
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 242385
pairs for 32-bit immediates.
This change is needed to avoid emitting movt/movw pairs when doing LTO
and do so on a per-function basis.
Out-of-tree projects currently using cl::opt option -arm-use-movt=0 or
false to avoid emitting movt/movw pairs should make changes to add
subtarget feature "+no-movt" (see the changes made to clang in r242368).
rdar://problem/21529937
Differential Revision: http://reviews.llvm.org/D11026
llvm-svn: 242369
The pass here was clearing kill flags on instructions which had
their sources killed in the instruction being combined. But
given that the new instruction is inserted after the existing ones,
any existing instructions with kill flags will lead to the verifier
complaining that we are reading an undefined physreg.
For example, what we had prior to this optimization is
t2STRi12 %R1, %SP, 12
t2STRi12 %R1<kill>, %SP, 16
t2STRi12 %R0<kill>, %SP, 8
and prior to this fix that would generate
t2STRi12 %R1<kill>, %SP, 16
t2STRDi8 %R0<kill>, %R1, %SP, 8
This is clearly incorrect as it didn't clear the kill flag on R1
used with offset 16 because there was no kill flag on the instruction
with offset 12.
After this change we clear the kill flag on the offset 16 instruction
because we know it will be used afterwards in the new instruction.
I haven't provided a test case. I have a small test, but even it is
very sensitive to register allocation order which isn't ideal.
llvm-svn: 242359
Pass a const reference to LiveRegMatrix to getRegAllocationHints()
because some targets can prodive better hints if they can test whether a
physreg has been used for register allocation yet.
llvm-svn: 242340
These were the cause of a verifier error when building 7zip with
-verify-machineinstrs. Running 'make check' with the verifier
triggered the same error on the test here so i've updated the test
to run the verifier on one of its runs instead of adding a new one.
While looking at this code, there was a stale comment that these
instructions were only used for disassembly. This probably used to
be the case, but they are now used in the 'ARM load / store optimization pass' too.
llvm-svn: 242300
The vec_sld interface provides access to the vsldoi instruction.
Unlike most of the vec_* interfaces, we do not attempt to change the
generated code for vec_sld based on the endian mode. It is too
difficult to correctly infer the desired semantics because of
different element types, and the corrected instruction sequence is
expensive, involving loading a permute control vector and performing a
generalized permute.
For GCC, this was implemented as "Don't touch the vec_sld"
implementation. When it came time for the LLVM implementation, I did
the same thing. However, this was hasty and incorrect. In LLVM's
version of altivec.h, vec_sld was previously defined in terms of the
vec_perm interface. Because vec_perm semantics are adjusted for
little endian, this means that leaving vec_sld untouched causes it to
generate something different for LE than for BE. Not good.
This back-end patch accompanies the changes to altivec.h that change
vec_sld's behavior for little endian. Those changes mean that we see
slightly different code in the back end when trying to recognize a
VSLDOI instruction in isVSLDOIShuffleMask. In particular, a
ShuffleKind of 1 (where the two inputs are identical) must now be
treated the same way as a ShuffleKind of 2 (little endian with
different inputs) when little endian mode is in force. This is
because ShuffleKind of 1 is defined using big-endian numbering.
This has a ripple effect on LowerBUILD_VECTOR, where we create our own
internal VSLDOI instructions. Because these are a ShuffleKind of 1,
they will now have their shift amounts subtracted from 16 when
recognizing the shuffle mask. To avoid problems we have to subtract
them from 16 again before creating the VSLDOI instructions.
There are a couple of other uses of BuildVSLDOI, but these do not need
to be modified because the shift amount is 8, which is unchanged when
subtracted from 16.
llvm-svn: 242296
- Teaches the ValueTracker in the PeepholeOptimizer to look through PHI
instructions.
- Add findNextSourceAndRewritePHI method to lookup into multiple sources
returnted by the ValueTracker and rewrite PHIs with new sources.
With these changes we can find more register sources and rewrite more
copies to allow coaslescing of bitcast instructions. Hence, we eliminate
unnecessary VR64 <-> GR64 copies in x86, but it could be extended to
other archs by marking "isBitcast" on target specific instructions. The
x86 example follows:
A:
psllq %mm1, %mm0
movd %mm0, %r9
jmp C
B:
por %mm1, %mm0
movd %mm0, %r9
jmp C
C:
movd %r9, %mm0
pshufw $238, %mm0, %mm0
Becomes:
A:
psllq %mm1, %mm0
jmp C
B:
por %mm1, %mm0
jmp C
C:
pshufw $238, %mm0, %mm0
Differential Revision: http://reviews.llvm.org/D11197
rdar://problem/20404526
llvm-svn: 242295
This is a direct port of the code from the X86 backend (r239486/r240361), which
uses the MachineCombiner to reassociate (floating-point) adds/muls to increase
ILP, to the PowerPC backend. The rationale is the same.
There is a lot of copy-and-paste here between the X86 code and the PowerPC
code, and we should extract at least some of this into CodeGen somewhere.
However, I don't want to do that until this code is enhanced to handle FMAs as
well. After that, we'll be in a better position to extract the common parts.
llvm-svn: 242279
If the source of the copy that defines the addend is a physical register, then
its existing live range may not extend to the FMA being mutated. Make sure we
extend the live range of the register to meet the FMA because it will become
its operand in this case.
I don't have an independent test case, but it will be exposed by change to be
committed shortly enabling the use of the machine combiner to do fadd/fmul
reassociation, and will be covered by one of the associated regression tests.
llvm-svn: 242278
Bitpatterns rejected by the decoder method of `MSR (immediate)` should be
decoded as the `extended MSR (register)` instruction.
Differential Revision: http://reviews.llvm.org/D7174
llvm-svn: 242276
This code was breaking from the case statement if the getStoreSizeInBits()
value was not a multiple of 0. Given that the implementation returns
getStoreSize() * 8, it can only be a multiple of 8.
llvm-svn: 242255
Summary:
processFunctionBeforeCalleeSavedScan was renamed to determineCalleeSaves and now takes a BitVector parameter as of rL242165, reviewed in http://reviews.llvm.org/D10909
WebAssembly is still marked as experimental and therefore doesn't build by default. It does, however, grep by default! I notice that processFunctionBeforeCalleeSavedScan is still mentioned in a few comments and error messages, which I also fixed.
Reviewers: qcolombet, sunfish
Subscribers: jfb, dsanders, hfinkel, MatzeB, llvm-commits
Differential Revision: http://reviews.llvm.org/D11199
llvm-svn: 242242
Follow-up r235483, with the corresponding support in PPC. We use a regular call
for symbolic targets (because they're much cheaper than indirect calls).
llvm-svn: 242239
We used to take the address specified as the direct target of the patchpoint
and did no TOC-pointer handling. This, however, as not all that useful,
because MCJIT tends to create a lot of modules, and they have their own TOC
sections. Thus, to call from the generated code to other generated code, you
really need to switch TOC pointers. Make this work as expected, and under
ELFv1, tread the address as the function descriptor address so that the correct
TOC pointer can be loaded.
llvm-svn: 242217
SelectionDAG already had begin/end methods for iterating over all
the nodes, but didn't define an iterator_range for us in foreach
loops.
This adds such a method and uses it in some of the eligible places
throughout the backends.
llvm-svn: 242212
Summary: This patch has the most basic instruction codegen for 32 and 64 bit int/fp.
Reviewers: sunfish
Subscribers: llvm-commits, jfb
Differential Revision: http://reviews.llvm.org/D11193
llvm-svn: 242201
MOVSDto64rr and MOV64toSDrr are defined to convert between FR64 (%xmm)
<-> GR64 registers, not VR64 (%mm) <-> GR64. This is wrong.
I found this by inspection and could not find a suitable testcase for it
since (1) we don't handle MMX bitcasts in Peephole optimizer as to
generate COPYs that (2) could be expanded back to the appropriate x86
instruction in ExpandPostRA.
Switch to use the appropriate instructions: MMX_MOVD64from64rr and
MMX_MOVD64to64rr here.
llvm-svn: 242191
PowerPC uses itineraries to describe processor pipelines (and dispatch-group
restrictions for P7/P8 cores). Unfortunately, the target-independent
implementation of TII.getInstrLatency calls ItinData->getStageLatency, and that
looks for the largest cycle count in the pipeline for any given instruction.
This, however, yields the wrong answer for the PPC itineraries, because we
don't encode the full pipeline. Because the functional units are fully
pipelined, we only model the initial stages (there are no relevant hazards in
the later stages to model), and so the technique employed by getStageLatency
does not really work. Instead, we should take the maximum output operand
latency, and that's what PPCInstrInfo::getInstrLatency now does.
This caused some test-case churn, including two unfortunate side effects.
First, the new arrangement of copies we get from function parameters now
sometimes blocks VSX FMA mutation (a FIXME has been added to the code and the
test cases), and we have one significant test-suite regression:
SingleSource/Benchmarks/BenchmarkGame/spectral-norm
56.4185% +/- 18.9398%
In this benchmark we have a loop with a vectorized FP divide, and it with the
new scheduling both divides end up in the same dispatch group (which in this
case seems to cause a problem, although why is not exactly clear). The grouping
structure is hard to predict from the bottom of the loop, and there may not be
much we can do to fix this.
Very few other test-suite performance effects were really significant, but
almost all weakly favor this change. However, in light of the issues
highlighted above, I've left the old behavior available via a
command-line flag.
llvm-svn: 242188
This can be done only with moves which theoretically
will optimize better later.
Although this transform increases the instruction count,
it should be code size / cycle count neutral in the worst
VALU case. It also seems to slightly improve a couple
of testcases due to other DAG combines this exposes.
This is probably slightly worse for the SALU case, so
it might be better to handle this during moveToVALU,
although then you lose some simplifications like
the load width reducing in the simple testcase.
llvm-svn: 242177
If the read2 produced was supposed to be writing into a
super register, it would use the wrong subregister indices.
Fix this by inserting copies, so we only ever write to a vreg_64.
Run the register coalescer again to clean this up, although this
isn't ideal and often does result in an extra move.
Also remove the assert that offset1 > offset0.
There isn't a real reason to not allow this other than a minor
convenience in the compiler, and it doesn't seem worth the effort
of avoiding it.
llvm-svn: 242174
We have a detailed def/use lists for every physical register in
MachineRegisterInfo anyway, so there is little use in maintaining an
additional bitset of which ones are used.
Removing it frees us from extra book keeping. This simplifies
VirtRegMap.
Differential Revision: http://reviews.llvm.org/D10911
llvm-svn: 242173
This changes TargetFrameLowering::processFunctionBeforeCalleeSavedScan():
- Rename the function to determineCalleeSaves()
- Pass a bitset of callee saved registers by reference, thus avoiding
the function-global PhysRegUsed bitset in MachineRegisterInfo.
- Without PhysRegUsed the implementation is fine tuned to not save
physcial registers which are only read but never modified.
Related to rdar://21539507
Differential Revision: http://reviews.llvm.org/D10909
llvm-svn: 242165
Summary:
- Signed 16-bit should have priority over unsigned.
- For la, unsigned 16-bit must use ori+addu rather than directly use ori.
- Correct tests on 32-bit immediates with 64-bit predicates by
sign-extending the immediate beforehand. For example, isInt<16>(0xffff8000)
should be true and use addiu.
Also split li/la testing into separate files due to their size.
Reviewers: vkalintiris
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10967
llvm-svn: 242139
This patch allows VSX swap optimization to succeed more frequently.
Specifically, it is concerned with common code sequences that occur
when copying a scalar floating-point value to a vector register. This
patch currently handles cases where the floating-point value is
already in a register, but does not yet handle loads (such as via an
LXSDX scalar floating-point VSX load). That will be dealt with later.
A typical case is when a scalar value comes in as a floating-point
parameter. The value is copied into a virtual VSFRC register, and
then a sequence of SUBREG_TO_REG and/or COPY operations will convert
it to a full vector register of the class required by the context. If
this vector register is then used as part of a lane-permuted
computation, the original scalar value will be in the wrong lane. We
can fix this by adding a swap operation following any widening
SUBREG_TO_REG operation. Additional COPY operations may be needed
around the swap operation in order to keep register assignment happy,
but these are pro forma operations that will be removed by coalescing.
If a scalar value is otherwise directly referenced in a computation
(such as by one of the many XS* vector-scalar operations), we
currently disable swap optimization. These operations are
lane-sensitive by definition. A MentionsPartialVR flag is added for
use in each swap table entry that mentions a scalar floating-point
register without having special handling defined.
A common idiom for PPC64LE is to convert a double-precision scalar to
a vector by performing a splat operation. This ensures that the value
can be referenced as V[0], as it would be for big endian, whereas just
converting the scalar to a vector with a SUBREG_TO_REG operation
leaves this value only in V[1]. A doubleword splat operation is one
form of an XXPERMDI instruction, which takes one doubleword from a
first operand and another doubleword from a second operand, with a
two-bit selector operand indicating which doublewords are chosen. In
the general case, an XXPERMDI can be permitted in a lane-swapped
region provided that it is properly transformed to select the
corresponding swapped values. This transformation is to reverse the
order of the two input operands, and to reverse and complement the
bits of the selector operand (derivation left as an exercise to the
reader ;).
A new test case that exercises the scalar-to-vector and generalized
XXPERMDI transformations is added as CodeGen/PowerPC/swaps-le-5.ll.
The patch also requires a change to CodeGen/PowerPC/swaps-le-3.ll to
use CHECK-DAG instead of CHECK for two independent instructions that
now appear in reverse order.
There are two small unrelated changes that are added with this patch.
First, the XXSLDWI instruction was incorrectly omitted from the list
of lane-sensitive instructions; this is now fixed. Second, I observed
that the same webs were being rejected over and over again for
different reasons. Since it's sufficient to reject a web only once, I
added a check for this to speed up the compilation time slightly.
llvm-svn: 242081
Enable partial and runtime loop unrolling for NVPTX backend via
TTI::UnrollingPreferences with a small threshold. This partially unrolls
small loops which are often unrolled by the PTX to SASS compiler
and unrolling earlier can be beneficial.
llvm-svn: 242049
The two-address instruction pass will convert these back to v_mad_f32
if necessary.
Differential Revision: http://reviews.llvm.org/D11060
llvm-svn: 242038
The 64/128-bit vector types are legal if NEON instructions are
available. However, there was no matching patterns for @llvm.cttz.*()
intrinsics and result in fatal error.
This commit fixes the problem by lowering cttz to:
a. ctpop((x & -x) - 1)
b. width - ctlz(x & -x) - 1
llvm-svn: 242037
