Various masks on shufflevector instructions are recognizable as
specific PowerPC instructions (vector pack, vector merge, etc.).
There is existing code in PPCISelLowering.cpp to recognize the correct
patterns for big endian code. The masks for these instructions are
different for little endian code due to the big-endian numbering
employed by these instructions. This patch adds the recognition code
for little endian.
I've added a new test case test/CodeGen/PowerPC/vec_shuffle_le.ll for
this. The existing recognizer test (vec_shuffle.ll) is unnecessarily
verbose and difficult to read, so I felt it was better to add a new
test rather than modify the old one.
llvm-svn: 210536
The code in PPCTargetLowering::PerformDAGCombine() that handles
unaligned Altivec vector loads generates a lvsl followed by a vperm.
As we've seen in numerous other places, the vperm instruction has a
big-endian bias, and this is fixed for little endian by complementing
the permute control vector and swapping the input operands. In this
case the lvsl is providing the permute control vector. Rather than
generating an lvsl and a complement operation, it is sufficient to
generate an lvsr instruction instead. Thus for LE code generation we
will generate an lvsr rather than an lvsl, and swap the other input
arguments on the vperm.
The existing test/CodeGen/PowerPC/vec_misalign.ll is updated to test
the code generation for PPC64 and PPC64LE, in addition to the existing
PPC32/G5 testing.
llvm-svn: 210493
The existing code in PPCTargetLowering::LowerMUL() for multiplying two
v16i8 values assumes that vector elements are numbered in big-endian
order. For little-endian targets, the vector element numbering is
reversed, but the vmuleub, vmuloub, and vperm instructions still
assume big-endian numbering. To account for this, we must adjust the
permute control vector and reverse the order of the input registers on
the vperm instruction.
The existing test/CodeGen/PowerPC/vec_mul.ll is updated to be executed
on powerpc64 and powerpc64le targets as well as the original powerpc
(32-bit) target.
llvm-svn: 210474
This patch fixes a couple of lowering issues for little endian
PowerPC. The code for lowering BUILD_VECTOR contains a number of
optimizations that are only valid for big endian. For now, we disable
those optimizations for correctness. In the future, we will add
analogous optimizations that are correct for little endian.
When lowering a SHUFFLE_VECTOR to a VPERM operation, we again need to
make the now-familiar transformation of swapping the input operands
and complementing the permute control vector. Correctness of this
transformation is tested by the accompanying test case.
llvm-svn: 210336
Chandler correctly pointed out that I need an LLVM IR test for
r210282, which modified the vperm -> shuffle transform for little
endian PowerPC. This patch provides that test.
llvm-svn: 210297
This seems to match what gcc does for ppc and what every other llvm
backend does.
This is a fixed version of r209638. The difference is to avoid any change
in behavior for functions. The logic for using constant pools for function
addresseses is spread over a few places and we have to keep them in sync.
llvm-svn: 209821
This reverts commit r209638 because it broke self-hosting on ppc64/Linux. (the
Clang-compiled TableGen would segfault because it jumped to an invalid address
from within _ZNK4llvm17ManagedStaticBase21RegisterManagedStaticEPFPvvEPFvS1_E
(which is within the command-line parameter registration process)).
llvm-svn: 209745
In PPCISelLowering.cpp: PPCTargetLowering::LowerBUILD_VECTOR(), there
is an optimization for certain patterns to generate one or two vector
splats followed by a vector add or subtract. This operation is
represented by a VADD_SPLAT in the selection DAG. Prior to this
patch, it was possible for the VADD_SPLAT to be assigned the wrong
data type, causing incorrect code generation. This patch corrects the
problem.
Specifically, the code previously assigned the value type of the
BUILD_VECTOR node to the newly generated VADD_SPLAT node. This is
correct much of the time, but not always. The problem is that the
call to isConstantSplat() may return a SplatBitSize that is not the
same as the number of bits in the original element vector type. The
correct type to assign is a vector type with the same element bit size
as SplatBitSize.
The included test case shows an example of this, where the
BUILD_VECTOR node has a type of v16i8. The vector to be built is {0,
16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16}. isConstantSplat
detects that we can generate a splat of 16 for type v8i16, which is
the type we must assign to the VADD_SPLAT node. If we do not, we
generate a vspltisb of 8 and a vaddubm, which generates the incorrect
result {16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
16}. The correct code generation is a vspltish of 8 and a vadduhm.
This patch also corrected code generation for
CodeGen/PowerPC/2008-07-10-SplatMiscompile.ll, which had been marked
as an XFAIL, so we can remove the XFAIL from the test case.
llvm-svn: 209662
The SplitIndexingFromLoad changes exposed a latent isel bug in the PowerPC64
backend. We matched an immediate offset with STWX8 even though it only
supports register offset.
The culprit is the complex-pattern predicate, SelectAddrIdx, which decides
that if the offset is not ISD::Constant it must be a register.
Many thanks to Bill Schmidt for testing this.
llvm-svn: 209219
Many old tests using prior schemas still had some brokenness here (both
indirect arrays and arrays with single bogus elements). Fixed those up
so they don't hit the new assertions.
Also reduced nesting in some places, etc.
llvm-svn: 208817
Support for the intrinsics that read from and write to global named registers
is added for r1, r2 and r13 (depending on the subtarget).
llvm-svn: 208509
The counter-loops formation pass needs to know what operations might be
function calls (because they can't appear in counter-based loops). On PPC32,
128-bit shifts might be runtime calls (even though you can't use __int128 on
PPC32, it seems that SROA might form them).
Fixes PR19709.
llvm-svn: 208501
We had been using the known-zero values of the operand of the or to construct
the mask for an rlwimi; this is not quite correct, but fine when the mask is
constant. When the mask is constant, then the known zeros of the operand must
be a superset of the zeros in the mask. However, when the mask is not a
constant, then there might be bits in the operand that are not known to be zero
that, at runtime, might be zero in the mask. Therefore, we check that any bits
not known to be zero *are* known to be one in the mask. Otherwise, we can't
fold the mask with the or and shift.
This was revealed as a miscompile of
MultiSource/Benchmarks/BitBench/drop3/drop3 when I started experimenting with
constant hoisting.
llvm-svn: 206136
We had disabled use of TBAA during CodeGen (even when otherwise using AA)
because the ptrtoint/inttoptr used by CGP for address sinking caused BasicAA to
miss basic type punning that it should catch (and, thus, we'd fail to override
TBAA when we should).
However, when AA is in use during CodeGen, CGP now uses normal GEPs and
bitcasts, instead of ptrtoint/inttoptr, when doing address sinking. As a
result, BasicAA should be able to make us do the right thing in the face of
type-punning, and it seems safe to enable use of TBAA again. self-hosting seems
fine on PPC64/Linux on the P7, with TBAA enabled and -misched=shuffle.
Note: We still don't update TBAA when merging stack slots, although because
BasicAA should now catch all such cases, this is no longer a blocking issue.
Nevertheless, I plan to commit code to deal with this properly in the near
future.
llvm-svn: 206093
The current memory-instruction optimization logic in CGP, which sinks parts of
the address computation that can be adsorbed by the addressing mode, does this
by explicitly converting the relevant part of the address computation into
IR-level integer operations (making use of ptrtoint and inttoptr). For most
targets this is currently not a problem, but for targets wishing to make use of
IR-level aliasing analysis during CodeGen, the use of ptrtoint/inttoptr is a
problem for two reasons:
1. BasicAA becomes less powerful in the face of the ptrtoint/inttoptr
2. In cases where type-punning was used, and BasicAA was used
to override TBAA, BasicAA may no longer do so. (this had forced us to disable
all use of TBAA in CodeGen; something which we can now enable again)
This (use of GEPs instead of ptrtoint/inttoptr) is not currently enabled by
default (except for those targets that use AA during CodeGen), and so aside
from some PowerPC subtargets and SystemZ, there should be no change in
behavior. We may be able to switch completely away from the ptrtoint/inttoptr
sinking on all targets, but further testing is required.
I've doubled-up on a number of existing tests that are sensitive to the
address sinking behavior (including some store-merging tests that are
sensitive to the order of the resulting ADD operations at the SDAG level).
llvm-svn: 206092
gcc inline asm supports specifying "cc" as a clobber of all condition
registers. Add just enough modeling of the full register to make this work.
Fixed PR19326.
llvm-svn: 205630
If we have two unique values for a v2i64 build vector, this will always result
in two vector loads if we expand using shuffles. Only one is necessary.
llvm-svn: 205231
When the loop vectorizer vectorizes code that uses the loop induction variable,
we often end up with IR like this:
%b1 = insertelement <2 x i32> undef, i32 %v, i32 0
%b2 = shufflevector <2 x i32> %b1, <2 x i32> undef, <2 x i32> zeroinitializer
%i = add <2 x i32> %b2, <i32 2, i32 3>
If the add in this example is not legal (as is the case on PPC with VSX), it
will be scalarized, and we'll end up with a number of extract_vector_elt nodes
with the vector shuffle as the input operand, and that vector shuffle is fed by
one or more build_vector nodes. By the time that vector operations are
expanded, visitEXTRACT_VECTOR_ELT will not create new extract_vector_elt by
looking through the vector shuffle (to make sure that no illegal operations are
created), and so the extract_vector_elt -> vector shuffle -> build_vector is
never simplified to an operand of the build vector.
By looking at build_vectors through a shuffle we fix this particular situation,
preventing a vector from being built, only to be deconstructed again (for the
scalarized add) -- an expensive proposition when this all needs to be done via
the stack. We probably want a more comprehensive fix here where we look back
recursively through any shuffles to any build_vectors or scalar_to_vectors,
etc. but that can come later.
llvm-svn: 205179
When expanding EXTRACT_VECTOR_ELT and EXTRACT_SUBVECTOR using
SelectionDAGLegalize::ExpandExtractFromVectorThroughStack, we store the entire
vector and then load the piece we want. This is fine in isolation, but
generating a new store (and corresponding stack slot) for each extraction ends
up producing code of poor quality. When we scalarize a vector operation (using
SelectionDAG::UnrollVectorOp for example) we generate one EXTRACT_VECTOR_ELT
for each element in the vector. This used to generate one stored copy of the
vector for each element in the vector. Now we search the uses of the vector for
a suitable store before generating a new one, which results in much more
efficient scalarization code.
llvm-svn: 205153
sitofp from v2i32 to v2f64 ends up generating a SIGN_EXTEND_INREG v2i64 node
(and similarly for v2i16 and v2i8). Even though there are no sign-extension (or
algebraic shifts) for v2i64 types, we can handle v2i32 sign extensions by
converting two and from v2i64. The small trick necessary here is to shift the
i32 elements into the right lanes before the i32 -> f64 step. This is because
of the big Endian nature of the system, we need the i32 portion in the high
word of the i64 elements.
For v2i16 and v2i8 we can do the same, but we first use the default Altivec
shift-based expansion from v2i16 or v2i8 to v2i32 (by casting to v4i32) and
then apply the above procedure.
llvm-svn: 205146
v2i64 is a legal type under VSX, however we don't have native vector
comparisons. We can handle eq/ne by casting it to an Altivec type, but
everything else must be expanded.
llvm-svn: 205106
We had stored both f64 values and v2f64, etc. values in the VSX registers. This
worked, but was suboptimal because we would always spill 16-byte values even
through we almost always had scalar 8-byte values. This resulted in an
increase in stack-size use, extra memory bandwidth, etc. To fix this, I've
added 64-bit subregisters of the Altivec registers, and combined those with the
existing scalar floating-point registers to form a class of VSX scalar
floating-point registers. The ABI code has also been enhanced to use this
register class and some other necessary improvements have been made.
llvm-svn: 205075
v2[fi]64 values need to be explicitly passed in VSX registers. This is because
the code in TRI that finds the minimal register class given a register and a
value type will assert if given an Altivec register and a non-Altivec type.
llvm-svn: 205041
As explained in r204976, because of how the allocation of VSX registers
interacts with the call-lowering code, we sometimes end up generating self VSX
copies. Specifically, things like this:
%VSL2<def> = COPY %F2, %VSL2<imp-use,kill>
(where %F2 is really a sub-register of %VSL2, and so this copy is a nop)
This adds a small cleanup pass to remove these prior to post-RA scheduling.
llvm-svn: 204980
First, v2f64 vector extract had not been declared legal (and so the existing
patterns were not being used). Second, the patterns for that, and for
scalar_to_vector, should really be a regclass copy, not a subregister
operation, because the VSX registers directly hold both the vector and scalar data.
llvm-svn: 204971
These operations need to be expanded during legalization so that isel does not
crash. In theory, we might be able to custom lower some of these. That,
however, would need to be follow-up work.
llvm-svn: 204963
These instructions have access to the complete VSX register file. In addition,
they "swap" the order of the elements so that element 0 (the scalar part) comes
first in memory and element 1 follows at a higher address.
llvm-svn: 204838
v2i64 needs to be a legal VSX type because it is the SetCC result type from
v2f64 comparisons. We need to expand all non-arithmetic v2i64 operations.
This fixes the lowering for v2f64 VSELECT.
llvm-svn: 204828
With VSX there is a real vector select instruction, and so we should use it.
Note that VSELECT will still scalarize for v2f64 because the corresponding
SetCC result type (v2i64) is not currently a legal type.
llvm-svn: 204801
The VSX instruction set has two types of FMA instructions: A-type (where the
addend is taken from the output register) and M-type (where one of the product
operands is taken from the output register). This adds a small pass that runs
just after MI scheduling (and, thus, just before register allocation) that
mutates A-type instructions (that are created during isel) into M-type
instructions when:
1. This will eliminate an otherwise-necessary copy of the addend
2. One of the product operands is killed by the instruction
The "right" moment to make this decision is in between scheduling and register
allocation, because only there do we know whether or not one of the product
operands is killed by any particular instruction. Unfortunately, this also
makes the implementation somewhat complicated, because the MIs are not in SSA
form and we need to preserve the LiveIntervals analysis.
As a simple example, if we have:
%vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
...
%vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19,
%RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19
...
We can eliminate the copy by changing from the A-type to the
M-type instruction. This means:
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
is replaced by:
%vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9,
%RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9
and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
llvm-svn: 204768
When VSX is available, these instructions should be used in preference to the
older variants that only have access to the scalar floating-point registers.
llvm-svn: 204559
When converting a signed 32-bit integer to double-precision floating point on
hardware without a lfiwax instruction, we have to instead use a lfd followed
by fcfid. We were erroneously offsetting the address by 4 bytes in
preparation for either a lfiwax or lfiwzx when generating the lfd. This fixes
that silly error.
This was not caught in the test suite since the conversion tests were run with
-mcpu=pwr7, which implies availability of lfiwax. I've added another test
case for older hardware that checks the code we expect in the absence of
lfiwax and other flavors of fcfid. There are fewer tests in this test case
because we punt to DAG selection in more cases on older hardware. (We must
generate complex fiddly sequences in those cases, and there is marginal
benefit in duplicating that logic in fast-isel.)
llvm-svn: 204155
Commit r181723 introduced code to avoid placing initialized variables
needing relocations into the .rodata section, which avoid copy relocs
that do not work as expected on ppc64 function references.
The same treatment is also needed for *named* .rodata.XXX sections.
This patch changes PPC64LinuxTargetObjectFile::SelectSectionForGlobal
to modify "Kind" *before* calling the default SelectSectionForGlobal
routine, instead of first calling the default routine and then just
checking for the (main) .rodata section afterwards.
llvm-svn: 203921
These linkages were introduced some time ago, but it was never very
clear what exactly their semantics were or what they should be used
for. Some investigation found these uses:
* utf-16 strings in clang.
* non-unnamed_addr strings produced by the sanitizers.
It turns out they were just working around a more fundamental problem.
For some sections a MachO linker needs a symbol in order to split the
section into atoms, and llvm had no idea that was the case. I fixed
that in r201700 and it is now safe to use the private linkage. When
the object ends up in a section that requires symbols, llvm will use a
'l' prefix instead of a 'L' prefix and things just work.
With that, these linkages were already dead, but there was a potential
future user in the objc metadata information. I am still looking at
CGObjcMac.cpp, but at this point I am convinced that linker_private
and linker_private_weak are not what they need.
The objc uses are currently split in
* Regular symbols (no '\01' prefix). LLVM already directly provides
whatever semantics they need.
* Uses of a private name (start with "\01L" or "\01l") and private
linkage. We can drop the "\01L" and "\01l" prefixes as soon as llvm
agrees with clang on L being ok or not for a given section. I have two
patches in code review for this.
* Uses of private name and weak linkage.
The last case is the one that one could think would fit one of these
linkages. That is not the case. The semantics are
* the linker will merge these symbol by *name*.
* the linker will hide them in the final DSO.
Given that the merging is done by name, any of the private (or
internal) linkages would be a bad match. They allow llvm to rename the
symbols, and that is really not what we want. From the llvm point of
view, these objects should really be (linkonce|weak)(_odr)?.
For now, just keeping the "\01l" prefix is probably the best for these
symbols. If we one day want to have a more direct support in llvm,
IMHO what we should add is not a linkage, it is just a hidden_symbol
attribute. It would be applicable to multiple linkages. For example,
on weak it would produce the current behavior we have for objc
metadata. On internal, it would be equivalent to private (and we
should then remove private).
llvm-svn: 203866
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.
llvm-svn: 203768
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
llvm-svn: 203559
When copying an i1 value into a GPR for a vaarg call, we need to explicitly
zero-extend the i1 value (otherwise an invalid CRBIT -> GPR copy will be
generated).
llvm-svn: 203041
