VSX is an ISA extension supported on the POWER7 and later cores that enhances
floating-point vector and scalar capabilities. Among other things, this adds
<2 x double> support and generally helps to reduce register pressure.
The interesting part of this ISA feature is the register configuration: there
are 64 new 128-bit vector registers, the 32 of which are super-registers of the
existing 32 scalar floating-point registers, and the second 32 of which overlap
with the 32 Altivec vector registers. This makes things like vector insertion
and extraction tricky: this can be free but only if we force a restriction to
the right register subclass when needed. A new "minipass" PPCVSXCopy takes care
of this (although it could do a more-optimal job of it; see the comment about
unnecessary copies below).
Please note that, currently, VSX is not enabled by default when targeting
anything because it is not yet ready for that. The assembler and disassembler
are fully implemented and tested. However:
- CodeGen support causes miscompiles; test-suite runtime failures:
MultiSource/Benchmarks/FreeBench/distray/distray
MultiSource/Benchmarks/McCat/08-main/main
MultiSource/Benchmarks/Olden/voronoi/voronoi
MultiSource/Benchmarks/mafft/pairlocalalign
MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4
SingleSource/Benchmarks/CoyoteBench/almabench
SingleSource/Benchmarks/Misc/matmul_f64_4x4
- The lowering currently falls back to using Altivec instructions far more
than it should. Worse, there are some things that are scalarized through the
stack that shouldn't be.
- A lot of unnecessary copies make it past the optimizers, and this needs to
be fixed.
- Many more regression tests are needed.
Normally, I'd fix these things prior to committing, but there are some
students and other contributors who would like to work this, and so it makes
sense to move this development process upstream where it can be subject to the
regular code-review procedures.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203768 91177308-0d34-0410-b5e6-96231b3b80d8
There are currently two schemes for mapping instruction operands to
instruction-format variables for generating the instruction encoders and
decoders for the assembler and disassembler respectively: a) to map by name and
b) to map by position.
In the long run, we'd like to remove the position-based scheme and use only
name-based mapping. Unfortunately, the name-based scheme currently cannot deal
with complex operands (those with suboperands), and so we currently must use
the position-based scheme for those. On the other hand, the position-based
scheme cannot deal with (register) variables that are split into multiple
ranges. An upcoming commit to the PowerPC backend (adding VSX support) will
require this capability. While we could teach the position-based scheme to
handle that, since we'd like to move away from the position-based mapping
generally, it seems silly to teach it new tricks now. What makes more sense is
to allow for partial transitioning: use the name-based mapping when possible,
and only use the position-based scheme when necessary.
Now the problem is that mixing the two sensibly was not possible: the
position-based mapping would map based on position, but would not skip those
variables that were mapped by name. Instead, the two sets of assignments would
overlap. However, I cannot currently change the current behavior, because there
are some backends that rely on it [I think mistakenly, but I'll send a message
to llvmdev about that]. So I've added a new TableGen bit variable:
noNamedPositionallyEncodedOperands, that can be used to cause the
position-based mapping to skip variables mapped by name.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203767 91177308-0d34-0410-b5e6-96231b3b80d8
Support to the IAS was added to actually parse and handle the complex SO
expressions. However, the object file lowering was not updated to compensate
for the fact that the shift operand may be an absolute expression.
When trying to assemble to an object file, the lowering would fail while
succeeding when emitting purely assembly. Add an appropriate test.
The test case is inspired by the test case provided by Jiangning Liu who also
brought the issue to light.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203762 91177308-0d34-0410-b5e6-96231b3b80d8
Extend what's currently done for shift because the HW performs this masking
implicitly:
(rotl:i32 x, (and y, 31)) -> (rotl:i32 x, y)
I use the newly factored out multiclass that was only supporting shifts so
far.
For testing I extended my testcase for the new rotation idiom.
<rdar://problem/15295856>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203718 91177308-0d34-0410-b5e6-96231b3b80d8
The peephole (shift x, (and y, 31)) -> (shift x, y) is repeated for each
integer type and each shift variant.
To improve this a new multiclass is added that covers all integer types. The
shift patterns are now instantiated from this. I am planning to add new
instances for rotates as well.
No functional change intended:
* test/CodeGen/X86/shift-and.ll provides coverage
* Compared the expanded tablegen output and matched up the defs for these
Pat<>s before and after
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203685 91177308-0d34-0410-b5e6-96231b3b80d8
When printing assembly we don't have a Layout object, but we can still
try to fold some constants.
Testcase by Ulrich Weigand.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203677 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
This is a white lie to workaround a widespread bug in the -mfp64
implementation.
The problem is that none of the 32-bit fpu ops mention the fact that they
clobber the upper 32-bits of the 64-bit FPR. This allows MTHC1 to be
scheduled on the wrong side of most 32-bit FPU ops, particularly MTC1.
Fixing that requires a major overhaul of the FPU implementation which can't
be done right now due to time constraints.
The testcase is SingleSource/Benchmarks/Misc/oourafft.c when given
TARGET_CFLAGS='-mips32r2 mfp64 -mmsa'.
Also correct the comment added in r203464 to indicate that two
instructions were affected.
Reviewers: matheusalmeida, jacksprat
Reviewed By: matheusalmeida
Differential Revision: http://llvm-reviews.chandlerc.com/D3029
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203659 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
Correct the match patterns and the lowerings that made the CodeGen tests pass despite the mistakes.
The original testcase that discovered the problem was SingleSource/UnitTests/SignlessType/factor.c in test-suite.
During review, we also found that some of the existing CodeGen tests were incorrect and fixed them:
* bitwise.ll: In bsel_v16i8 the IfSet/IfClear were reversed because bsel and bmnz have different operand orders and the test didn't correctly account for this. bmnz goes 'IfClear, IfSet, CondMask', while bsel goes 'CondMask, IfClear, IfSet'.
* vec.ll: In the cases where a bsel is emitted as a bmnz (they are the same operation with a different input tied to the result) the operands were in the wrong order.
* compare.ll and compare_float.ll: The bsel operand order was correct for a greater-than comparison, but a greater-than comparison instruction doesn't exist. Lowering this operation inverts the condition so the IfSet/IfClear need to be swapped to match.
The differences between BSEL, BMNZ, and BMZ and how they map to/from vselect are rather confusing. I've therefore added a note to MSA.txt to explain this in a single place in addition to the comments that explain each case.
Reviewers: matheusalmeida, jacksprat
Reviewed By: matheusalmeida
Differential Revision: http://llvm-reviews.chandlerc.com/D3028
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203657 91177308-0d34-0410-b5e6-96231b3b80d8
When the list of VFP registers to be saved was non-contiguous (so multiple
vpush/vpop instructions were needed) these were being ordered oddly, as in:
vpush {d8, d9}
vpush {d11}
This led to the layout in memory being [d11, d8, d9] which is ugly and doesn't
match the CFI_INSTRUCTIONs we're generating either (so Dwarf info would be
broken).
This switches the order of vpush/vpop (in both prologue and epilogue,
obviously) so that the Dwarf locations are correct again.
rdar://problem/16264856
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203655 91177308-0d34-0410-b5e6-96231b3b80d8
The function hasReliableSymbolDifference had exactly one use in the MachO
writer. It is also only true for X86_64. In fact, the comments refers to
"Darwin x86_64" and everything else, so this makes the code match the
comment.
If this is to be abstracted again, it should be a property of
TargetObjectWriter, like useAggressiveSymbolFolding.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203605 91177308-0d34-0410-b5e6-96231b3b80d8
Use the options in the ARMISelLowering to control whether tail calls are
optimised or not. Previously, this option was entirely ignored on the ARM
target and only honoured on x86.
This option is mostly useful in profiling scenarios. The default remains that
tail call optimisations will be applied.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203577 91177308-0d34-0410-b5e6-96231b3b80d8
This option is from 2010, designed to work around a linker issue on Darwin for
ARM. According to grosbach this is no longer an issue and this option can
safely be removed.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203576 91177308-0d34-0410-b5e6-96231b3b80d8
Tail call optimisation was previously disabled on all targets other than
iOS5.0+. This enables the tail call optimisation on all Thumb 2 capable
platforms.
The test adjustments are to remove the IR hint "tail" to function invocation.
The tests were designed assuming that tail call optimisations would not kick in
which no longer holds true.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203575 91177308-0d34-0410-b5e6-96231b3b80d8
ATOMIC_STORE operations always get here as a lowered ATOMIC_SWAP, so there's no
need for any code to handle them specially.
There should be no functionality change so no tests.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203567 91177308-0d34-0410-b5e6-96231b3b80d8
The syntax for "cmpxchg" should now look something like:
cmpxchg i32* %addr, i32 42, i32 3 acquire monotonic
where the second ordering argument gives the required semantics in the case
that no exchange takes place. It should be no stronger than the first ordering
constraint and cannot be either "release" or "acq_rel" (since no store will
have taken place).
rdar://problem/15996804
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203559 91177308-0d34-0410-b5e6-96231b3b80d8
When the MOVBE instructions are available, use them for 16-bit endian
swapping as well as for 32 and 64 bit.
The patterns were already present on the instructions, but weren't being
matched because the operation was unconditionally marked to 'Expand.'
Change that to be conditional on whether the MOVBE instructions are
available. Use 'rolw' to implement the in-register version (32 and 64
bit have the dedicated 'bswap' instruction for that).
Patch by Louis Gerbarg <lgg@apple.com>.
rdar://15479984
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203524 91177308-0d34-0410-b5e6-96231b3b80d8
NVPTX, like the other backends, relies on generic symbol name sanitizing done by
MCSymbol. However, the ptxas assembler is more stringent and disallows some
additional characters in symbol names.
See PR19099 for more details.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203483 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
This is a white lie to workaround a widespread bug in the -mfp64
implementation.
The problem is that none of the 32-bit fpu ops mention the fact that they
clobber the upper 32-bits of the 64-bit FPR. This allows MFHC1 to be
scheduled on the wrong side of most 32-bit FPU ops. Fixing that requires a
major overhaul of the FPU implementation which can't be done right now due to
time constraints.
MFHC1 is one of two affected instructions. These instructions are the only
FPU instructions that don't read or write the lower 32-bits. We therefore
pretend that it reads the bottom 32-bits to artificially create a dependency and
prevent the scheduler changing the behaviour of the code.
The other instruction is MTHC1 which will be fixed once I've have found a failing
test case for it.
The testcase is test-suite/SingleSource/UnitTests/Vector/simple.c when
given TARGET_CFLAGS="-mips32r2 -mfp64 -mmsa".
Reviewers: jacksprat, matheusalmeida
Reviewed By: jacksprat
Differential Revision: http://llvm-reviews.chandlerc.com/D2966
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203464 91177308-0d34-0410-b5e6-96231b3b80d8
The function was making too many assumptions about its input:
1. The NEON_VDUP optimisation was far too aggressive, assuming (I
think) that the input would always be BUILD_VECTOR.
2. We were treating most unknown concats as legal (by returning Op
rather than SDValue()). I think only concats of pairs of vectors are
actually legal.
http://llvm.org/PR19094
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203450 91177308-0d34-0410-b5e6-96231b3b80d8
the stack of the analysis group because they are all immutable passes.
This is made clear by Craig's recent work to use override
systematically -- we weren't overriding anything for 'finalizePass'
because there is no such thing.
This is kind of a lame restriction on the API -- we can no longer push
and pop things, we just set up the stack and run. However, I'm not
invested in building some better solution on top of the existing
(terrifying) immutable pass and legacy pass manager.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203437 91177308-0d34-0410-b5e6-96231b3b80d8
This requires a number of steps.
1) Move value_use_iterator into the Value class as an implementation
detail
2) Change it to actually be a *Use* iterator rather than a *User*
iterator.
3) Add an adaptor which is a User iterator that always looks through the
Use to the User.
4) Wrap these in Value::use_iterator and Value::user_iterator typedefs.
5) Add the range adaptors as Value::uses() and Value::users().
6) Update *all* of the callers to correctly distinguish between whether
they wanted a use_iterator (and to explicitly dig out the User when
needed), or a user_iterator which makes the Use itself totally
opaque.
Because #6 requires churning essentially everything that walked the
Use-Def chains, I went ahead and added all of the range adaptors and
switched them to range-based loops where appropriate. Also because the
renaming requires at least churning every line of code, it didn't make
any sense to split these up into multiple commits -- all of which would
touch all of the same lies of code.
The result is still not quite optimal. The Value::use_iterator is a nice
regular iterator, but Value::user_iterator is an iterator over User*s
rather than over the User objects themselves. As a consequence, it fits
a bit awkwardly into the range-based world and it has the weird
extra-dereferencing 'operator->' that so many of our iterators have.
I think this could be fixed by providing something which transforms
a range of T&s into a range of T*s, but that *can* be separated into
another patch, and it isn't yet 100% clear whether this is the right
move.
However, this change gets us most of the benefit and cleans up
a substantial amount of code around Use and User. =]
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203364 91177308-0d34-0410-b5e6-96231b3b80d8
