of normally binary shuffle instructions like PUNPCKL and MOVLHPS.
This detects cases where a single register is used for both operands
making the shuffle behave in a unary way. We detect this and adjust the
mask to use the unary form which allows the existing DAG combine for
shuffle instructions to actually work at all.
As a consequence, this uncovered a number of obvious bugs in the
existing DAG combine which are fixed. It also now canonicalizes several
shuffles even with the existing lowering. These typically are trying to
match the shuffle to the domain of the input where before we only really
modeled them with the floating point variants. All of the cases which
change to an integer shuffle here have something in the integer domain, so
there are no more or fewer domain crosses here AFAICT. Technically, it
might be better to go from a GPR directly to the floating point domain,
but detecting floating point *outputs* despite integer inputs is a lot
more code and seems unlikely to be worthwhile in practice. If folks are
seeing domain-crossing regressions here though, let me know and I can
hack something up to fix it.
Also as a consequence, a bunch of missed opportunities to form pshufb
now can be formed. Notably, splats of i8s now form pshufb.
Interestingly, this improves the existing splat lowering too. We go from
3 instructions to 1. Yes, we may tie up a register, but it seems very
likely to be worth it, especially if splatting the 0th byte (the
common case) as then we can use a zeroed register as the mask.
llvm-svn: 214625
Stop using ST registers for function returns and inline-asm instructions and use
FP registers instead. This allows removing a large amount of code in the
stackifier pass that was needed to track register liveness and handle copies
between ST and FP registers and function calls returning floating point values.
It also fixes a bug which manifests when an ST register defined by an
inline-asm instruction was live across another inline-asm instruction, as shown
in the following sequence of machine instructions:
1. INLINEASM <es:frndint> $0:[regdef], %ST0<imp-def,tied5>
2. INLINEASM <es:fldcw $0>
3. %FP0<def> = COPY %ST0
<rdar://problem/16952634>
llvm-svn: 214580
Currently when DAGCombine converts loads feeding a switch into a switch of
addresses feeding a load the new load inherits the isInvariant flag of the left
side. This is incorrect since invariant loads can be reordered in cases where it
is illegal to reoarder normal loads.
This patch adds an isInvariant parameter to getExtLoad() and updates all call
sites to pass in the data if they have it or false if they don't. It also
changes the DAGCombine to use that data to make the right decision when
creating the new load.
llvm-svn: 214449
Rename to allowsMisalignedMemoryAccess.
On R600, 8 and 16 byte accesses are mostly OK with 4-byte alignment,
and don't need to be split into multiple accesses. Vector loads with
an alignment of the element type are not uncommon in OpenCL code.
llvm-svn: 214055
instructions in the legalized DAG, and leverage it to combine long
sequences of instructions to PSHUFB.
Eventually, the other x86-instruction-specific shuffle combines will
probably all be driven out of this routine. But the real motivation is
to detect after we have fully legalized and optimized a shuffle to the
minimal number of x86 instructions whether it is profitable to replace
the chain with a fully generic PSHUFB instruction even though doing so
requires either a load from a constant pool or tying up a register with
the mask.
While the Intel manuals claim it should be used when it replaces 5 or
more instructions (!!!!) my experience is that it is actually very fast
on modern chips, and so I've gon with a much more aggressive model of
replacing any sequence of 3 or more instructions.
I've also taught it to do some basic canonicalization to special-purpose
instructions which have smaller encodings than their generic
counterparts.
There are still quite a few FIXMEs here, and I've not yet implemented
support for lowering blends with PSHUFB (where its power really shines
due to being able to zero out lanes), but this starts implementing real
PSHUFB support even when using the new, fancy shuffle lowering. =]
llvm-svn: 214042
The tale starts with r212808 which attempted to fix inversion of the low
and high bits when lowering MUL_LOHI. Sadly, that commit did not include
any positive test cases, and just removed some operations from a test
case where the actual logic being changed isn't fully visible from the
test.
What this commit did was two things. First, it reversed the low and high
results in the formation of the MERGE_VALUES node for the multiple
results. This is entirely correct.
Second it changed the shuffles for extracting the low and high
components from the i64 results of the multiplies to extract them
assuming a big-endian-style encoding of the multiply results. This
second change is wrong. There is no big-endian encoding in x86, the
results of the multiplies are normal v2i64s: when cast to v4i32, the low
i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3.
However, the first change wasn't enough to actually fix the bug, which
is (I assume) why the second change was also made. There was another bug
in the MERGE_VALUES formation: we weren't using a VTList, and so were
getting a single result node! When grabbing the *second* result from the
node, we got... well.. colud be anything. I think this *appeared* to
invert things, but had to be causing other problems as well.
Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should
have been fine, right? NOOOPE! Because the core bug was never addressed,
the test in vector-idiv failed when I fixed the MERGE_VALUES node.
Because there are essentially no docs for this node, I had to guess at
how to fix it and tried swapping the operands, restoring the order of
the original code before r212808. While this "fixed" the test case (in
that we produced the write instructions) we were still extracting the
wrong elements of the i64s, and thus PR20355 was still broken.
This commit essentially reverts the big-endian-style extraction part of
r212808 and goes back to the original masks which were correct. Now that
the MERGE_VALUES node formation is also correct, everything works. I've
also included a more detailed test from PR20355 to make sure this stays
fixed.
llvm-svn: 214011
The clever way to implement signed multiplication with unsigned *is
already implemented* and tested and working correctly. The bug is
somewhere else. Re-investigating.
This will teach me to not scroll far enough to read the code that did
what I thought needed to be done.
llvm-svn: 214009
signed multiplication is requested. While there is not a difference in
the *low* half of the result, the *high* half (used specifically to
implement the signed division by these constants) certainly is used. The
test case I've nuked was actively asserting wrong code.
There is a delightful solution to doing signed multiplication even when
we don't have it that Richard Smith has crafted, but I'll add the
machinery back and implement that in a follow-up patch. This at least
restores correctness.
llvm-svn: 214007
address of the stack guard was being spilled to the stack.
Previously the address of the stack guard would get spilled to the stack if it
was impossible to keep it in a register. This patch introduces a new target
independent node and pseudo instruction which gets expanded post-RA to a
sequence of instructions that load the stack guard value. Register allocator
can now just remat the value when it can't keep it in a register.
<rdar://problem/12475629>
llvm-svn: 213967
SDValues, fixing the two bugs left in the regression suite.
The key for both of these was the use a single value type rather than
a VTList which caused an unintentionally single-result merge-value node.
Fix this by getting the appropriate VTList in place.
Doing this exposed that the comments in x86's code abouth how MUL_LOHI
operands are handle is wrong. The bug with the use of out-of-range
result numbers was hiding the bug about the order of operands here (as
best i can tell). There are more places where the code appears to get
this backwards still...
llvm-svn: 213931
vector operation legalization with support for custom target lowering
and fallback to expand when it fails, and use this to implement sext and
anyext load lowering for x86 in a more principled way.
Previously, the x86 backend relied on a target DAG combine to "combine
away" sextload and extload nodes prior to legalization, or would expand
them during legalization with terrible code. This is particularly
problematic because the DAG combine relies on running over non-canonical
DAG nodes at just the right time to match several common and important
patterns. It used a combine rather than lowering because we didn't have
good lowering support, and to expose some tricks being employed to more
combine phases.
With this change it becomes a proper lowering operation, the backend
marks that it can lower these nodes, and I've added support for handling
the canonical forms that don't have direct legal representations such as
sextload of a v4i8 -> v4i64 on AVX1. With this change, our test cases
for this behavior continue to pass even after the DAG combiner beigns
running more systematically over every node.
There is some noise caused by this in the test suite where we actually
use vector extends instead of subregister extraction. This doesn't
really seem like the right thing to do, but is unlikely to be a critical
regression. We do regress in one case where by lowering to the
target-specific patterns early we were able to combine away extraneous
legal math nodes. However, this regression is completely addressed by
switching to a widening based legalization which is what I'm working
toward anyways, so I've just switched the test to that mode.
Differential Revision: http://reviews.llvm.org/D4654
llvm-svn: 213897
GCC 4.8 detected a signed compare [-Wsign-compare]. Add a cast for the
destination index. Add an assert to catch a potential overflow however unlikely
it may be.
llvm-svn: 213878
When we had a vector_shuffle where we had an input from each vector, we
could miscompile it because we were assuming the input from V2 wouldn't
be moved from where it was on the vector.
Added a test case.
llvm-svn: 213826
The transform to constant fold unary operations with an AND across a
vector comparison applies when the constant is not a splat of a scalar
as well.
llvm-svn: 213800
The folding of unary operations through a vector compare and mask operation
is only safe if the unary operation result is of the same size as its input.
For example, it's not safe for [su]itofp from v4i32 to v4f64.
llvm-svn: 213799
This chang fully reverts r211771.
That revision added a canonicalization rule which has the potential to causes a
combine-cycle in the target-independent canonicalizing DAG combine.
The plan is to move the logic that forms target specific addsub nodes as part of
the lowering of shuffles.
llvm-svn: 213736
This patch removes function 'CommuteVectorShuffle' from X86ISelLowering.cpp
and moves its logic into SelectionDAG.cpp as method 'getCommutedVectorShuffles'.
This refactoring is in preperation of an upcoming change to the DAGCombiner.
llvm-svn: 213503
Since the result of a SETCC for X86 is 0 or -1 in each lane, we can
move unary operations, in this case [su]int_to_fp through the mask
operation and constant fold the operation away. Generally speaking:
UNARYOP(AND(VECTOR_CMP(x,y), constant))
--> AND(VECTOR_CMP(x,y), constant2)
where constant2 is UNARYOP(constant).
This implements the transform where UNARYOP is [su]int_to_fp.
For example, consider the simple function:
define <4 x float> @foo(<4 x float> %val, <4 x float> %test) nounwind {
%cmp = fcmp oeq <4 x float> %val, %test
%ext = zext <4 x i1> %cmp to <4 x i32>
%result = sitofp <4 x i32> %ext to <4 x float>
ret <4 x float> %result
}
Before this change, the SSE code is generated as:
LCPI0_0:
.long 1 ## 0x1
.long 1 ## 0x1
.long 1 ## 0x1
.long 1 ## 0x1
.section __TEXT,__text,regular,pure_instructions
.globl _foo
.align 4, 0x90
_foo: ## @foo
cmpeqps %xmm1, %xmm0
andps LCPI0_0(%rip), %xmm0
cvtdq2ps %xmm0, %xmm0
retq
After, the code is improved to:
LCPI0_0:
.long 1065353216 ## float 1.000000e+00
.long 1065353216 ## float 1.000000e+00
.long 1065353216 ## float 1.000000e+00
.long 1065353216 ## float 1.000000e+00
.section __TEXT,__text,regular,pure_instructions
.globl _foo
.align 4, 0x90
_foo: ## @foo
cmpeqps %xmm1, %xmm0
andps LCPI0_0(%rip), %xmm0
retq
The cvtdq2ps has been constant folded away and the floating point 1.0f
vector lanes are materialized directly via the ModRM operand of andps.
llvm-svn: 213342
Previously we asserted on this code. Currently compiler-rt doesn't
actually implement any of these new libcalls, but external help is
pretty much the only viable option for LLVM.
I've followed the much more generic "__truncST2" naming, as opposed to
the odd name for f32 -> f16 truncation. This can obviously be changed
later, or overridden by any targets that need to.
llvm-svn: 213252
x86 has no native ability to extend an f16 to f64, but the same result
is obtained if we expand it into two separate extensions: f16 -> f32
-> f64.
Unfortunately the same is not true for truncate, so that still results
in a compilation failure.
llvm-svn: 213251
This makes the two intrinsics @llvm.convert.from.f16 and
@llvm.convert.to.f16 accept types other than simple "float". This is
only strictly needed for the truncate operation, since otherwise
double rounding occurs and there's no way to represent the strict IEEE
conversion. However, for symmetry we allow larger types in the extend
too.
During legalization, we can expand an "fp16_to_double" operation into
two extends for convenience, but abort when the truncate isn't legal. A new
libcall is probably needed here.
Even after this commit, various target tweaks are needed to actually use the
extended intrinsics. I've put these into separate commits for clarity, so there
are no actual tests of f64 conversion here.
llvm-svn: 213248
It turns out that in most cases (the main exception being i1-related
types) once these operations are formed we cannot separate them and
the targets end up having to deal with them whether they want to or
not.
This is not a good situation, and a more reasonable default can be
formed by ackowledging this and having targets leave them as Legal.
Only x86 seems to be affected (other targets don't even try marking
the operation Expand).
Mostly there's no visible change here yet, but it will be useful to
have truly expanded EXTLOADS for MVT::f16 softening support.
llvm-svn: 213162
Before this change, method 'isShuffleMaskLegal' didn't know that shuffles
implementing a 'movhlps' operation were perfectly legal for SSE targets.
This patch adds the missing check for 'isMOVHLPSMask' inside method
'isShuffleMaskLegal' to fix the problem.
The reason why it is important to do this is because the DAGCombiner
conservatively avoids combining a pair of shuffles if the resulting shuffle
node has an illegal mask. Before this patch, shuffles with a MOVHLPS mask were
wrongly considered not to be legal. This was the root cause of some poor-code
generation bugs.
llvm-svn: 213137
shuffle lowering: match shuffle patterns equivalent to an unpcklwd or
unpckhwd instruction.
This allows us to use generic lowering code for v8i16 shuffles and match
the unpack pattern late.
llvm-svn: 212705
combine into half-shuffles through unpack instructions that expand the
half to a whole vector without messing with the dword lanes.
This fixes some redundant instructions in splat-like lowerings for
v16i8, which are now getting to be *really* nice.
llvm-svn: 212695
that splat i8s into i16s.
Previously, we would try much too hard to arrange a sequence of i8s in
one half of the input such that we could unpack them into i16s and
shuffle those into place. This isn't always going to be a cheaper i8
shuffle than our other strategies. The case where it is always going to
be cheaper is when we can arrange all the necessary inputs into one half
using just i16 shuffles. It happens that viewing the problem this way
also makes it much easier to produce an efficient set of shuffles to
move the inputs into one half and then unpack them.
With this, our splat code gets one step closer to being not terrible
with the new experimental lowering strategy. It also exposes two
combines missing which I will add next.
llvm-svn: 212692
shuffles specifically for cases where a small subset of the elements in
the input vector are actually used.
This is specifically targetted at improving the shuffles generated for
trunc operations, but also helps out splat-like operations.
There is still some really low-hanging fruit here that I want to address
but this is a huge step in the right direction.
llvm-svn: 212680
don't need to set it manually.
This is based on feedback from Tom who pointed out that if every target
needs to handle this we need to reach out to those maintainers. In fact,
it doesn't make sense to duplicate everything when anything other than
expand seems unlikely at this stage.
llvm-svn: 212661
Turns out my trick of using the same masks for SSE4.1 and AVX2 didn't work out
as we have to blend two vectors. While there remove unecessary cross-lane moves
from the shuffles so the backend can lower it to palignr instead of vperm.
Fixes PR20118, a miscompilation of vector sdiv by constant on AVX2.
llvm-svn: 212611
vector types to be legal and a ZERO_EXTEND node is encountered.
When we use widening to legalize vector types, extend nodes are a real
challenge. Either the input or output is likely to be legal, but in many
cases not both. As a consequence, we don't really have any way to
represent this situation and the prior code in the widening legalization
framework would just scalarize the extend operation completely.
This patch introduces a new DAG node to represent doing a zero extend of
a vector "in register". The core of the idea is to allow legal but
different vector types in the input and output. The output vector must
have fewer lanes but wider elements. The operation is defined to zero
extend the low elements of the input to the size of the output elements,
and drop all of the high elements which don't have a corresponding lane
in the output vector.
It also includes generic expansion of this node in terms of blending
a zero vector into the high elements of the vector and bitcasting
across. This in turn yields extremely nice code for x86 SSE2 when we use
the new widening legalization logic in conjunction with the new shuffle
lowering logic.
There is still more to do here. We need to support sign extension, any
extension, and potentially int-to-float conversions. My current plan is
to continue using similar synthetic nodes to model each of these
transitions with generic lowering code for each one.
However, with this patch LLVM already reaches performance parity with
GCC for the core C loops of the x264 code (assuming you disable the
hand-written assembly versions) when compiling for SSE2 and SSE3
architectures and enabling the new widening and lowering logic for
vectors.
Differential Revision: http://reviews.llvm.org/D4405
llvm-svn: 212610
has settled without incident, removing the x86-specific and overly
strict 'isVectorSplat' routine in favor of generic and more powerful
splat detection.
The primary motivation and result of this is that the x86 backend can
now see through splats which contain undef elements. This is essential
if we are using a widening form of legalization and I've updated a test
case to also run in that mode as before this change the generated code
for the test case was completely scalarized.
This version of the patch much more carefully handles the undef lanes.
- We aren't overly conservative about them in the shift lowering
(where we will never use the splat itself).
- One place where the splat would have been re-used by the existing code
now explicitly constructs a new constant splat that will be safe.
- The broadcast lowering is much more reasonable with undefs by doing
a correct check of whether the splat is the only user of a loaded
value, checking that the splat actually crosses multiple lanes before
using a broadcast, and handling broadcasts of non-constant splats.
As a consequence of the last bullet, the weird usage of vpshufd instead
of vbroadcast is gone, and we actually can lower an AVX splat with
vbroadcastss where before we emitted a really strange pattern of
a vector load and a manual splat across the vector.
llvm-svn: 212602
aggressively from the x86 shuffle lowering to the generic SDAG vector
shuffle formation code.
This code already tried to fold away shuffles of splats! It just had
lots of bugs and couldn't handle the case my new x86 shuffle lowering
needed.
First, it failed to correctly compute whether N2 was undef because it
pre-computed this, then did transformations which could *make* N2 undef,
then failed to ever re-consider the precomputed state.
Second, it didn't look through bitcasts at all, even in the safe cases
where they are just element-type bitcasts with no change to the number
of elements.
Third, it didn't handle all-zero bit casts nicely the way my code in the
x86 side of things did, which is essential to getting good zext-shuffle
lowerings.
But all of these are generic. I just ported the code down to this layer
and fixed the surrounding bugs. Tests exercising this in the x86 backend
still pass and some silly code in widen_cast-6.ll gets better. I updated
that test to be a bit more precise but it's still pretty unclear what
the value of the test is in this day and age.
llvm-svn: 212517
When combining a sequence of two PSHUFD dag nodes into a single PSHUFD,
make sure that we assign the correct type to the resulting PSHUFD.
X86ISD::PSHUFD dag nodes can be either MVT::v4i32 or MVT::v4f32.
Before this change, an assertion failure was triggered in method
'DAGCombinerInfo::CombineTo' when trying to combine the shuffles from the test
below into a single PSHUFD.
define <4 x float> @test1(<4 x float> %V) {
%1 = shufflevector <4 x float> %V, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
%2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
ret <4 x float> %2
}
llvm-svn: 212498
lanes in vector splats.
The core problem here is that undef lanes can't *unilaterally* be
considered to contribute to splats. Their handling needs to be more
cautious. There is also a reported failure of the nightly testers
(thanks Tobias!) that may well stem from the same core issue. I'm going
to fix this theoretical issue, factor the APIs a bit better, and then
verify that I don't see anything bad with Tobias's reduction from the
test suite before recommitting.
Original commit message for r212324:
[x86] Generalize BuildVectorSDNode::getConstantSplatValue to work for
any constant, constant FP, or undef splat and to tolerate any undef
lanes in a splat, then replace all uses of isSplatVector in X86's
lowering with it.
This fixes issues where undef lanes in an otherwise splat vector would
prevent the splat logic from firing. It is a touch more awkward to use
this interface, but it is much more accurate. Suggestions for better
interface structuring welcome.
With this fix, the code generated with the widening legalization
strategy for widen_cast-4.ll is *dramatically* improved as the special
lowering strategies for a v16i8 SRA kick in even though the high lanes
are undef.
We also get a slightly different choice for broadcasting an aligned
memory location, and use vpshufd instead of vbroadcastss. This looks
like a minor win for pipelining and domain crossing, but a minor loss
for the number of micro-ops. I suspect its a wash, but folks can
easily tweak the lowering if they want.
llvm-svn: 212475
essentially a DAG combine that never gets a chance to run.
We might typically expect DAG combining to remove shuffles-of-splats and
other similar patterns, but we don't get a chance to run the DAG
combiner when we recursively form sub-shuffles during the lowering of
a shuffle. So instead hand-roll a really important combine directly into
the lowering code to detect shuffles-of-splats, especially shuffles of
an all-zero splat which needn't even have the same element width, etc.
This lets the new vector shuffle lowering handle shuffles which
implement things like zero-extension really nicely. This will become
even more important when I wire the legalization of zero-extension to
vector shuffles with the new widening legalization strategy.
llvm-svn: 212444
any constant, constant FP, or undef splat and to tolerate any undef
lanes in a splat, then replace all uses of isSplatVector in X86's
lowering with it.
This fixes issues where undef lanes in an otherwise splat vector would
prevent the splat logic from firing. It is a touch more awkward to use
this interface, but it is much more accurate. Suggestions for better
interface structuring welcome.
With this fix, the code generated with the widening legalization
strategy for widen_cast-4.ll is *dramatically* improved as the special
lowering strategies for a v16i8 SRA kick in even though the high lanes
are undef.
We also get a slightly different choice for broadcasting an aligned
memory location, and use vpshufd instead of vbroadcastss. This looks
like a minor win for pipelining and domain crossing, but a minor loss
for the number of micro-ops. I suspect its a wash, but folks can easily
tweak the lowering if they want.
llvm-svn: 212324
This patch adds tablegen patterns to select F16C float-to-half-float
conversion instructions from 'f32_to_f16' and 'f16_to_f32' dag nodes.
If the target doesn't have F16C, then 'f32_to_f16' and 'f16_to_f32'
are expanded into library calls.
llvm-svn: 212293
mode.
This also runs the test in that mode which would reproduce the crash.
What I love is that *every single FIXME* in the test is addressed by
switching to widening.
llvm-svn: 212254
Finkel, Eric Christopher, and a bunch of other people I'm probably
forgetting (sorry), add an option to the x86 backend to widen vectors
during type legalization rather than promote them.
This still would promote vNi1 vectors to get the masks right, but would
widen other vectors. A lot of experiments are piling up right now
showing that widening should probably be the default legalization
strategy outside of vNi1 cases, but it is very hard to test the
rammifications of that and fix bugs in widening-based legalization
without an option that enables it. I'll be checking in tests shortly
that use this option to exercise cases where widening doesn't work well
and hopefully we'll be able to switch fully to this soon.
llvm-svn: 212249
The argument list vector is never used after it has been passed to the
CallLoweringInfo and moving it to the CallLoweringInfo is cleaner and
pretty much as cheap as keeping a pointer to it.
llvm-svn: 212135
On targets without cmpxchg16b or cmpxchg8b, the borderline atomic
operations were slipping through the gaps.
X86AtomicExpand.cpp was delegating to ISelLowering. Generic
ISelLowering was delegating to X86ISelLowering and X86ISelLowering was
asserting. The correct behaviour is to expand to a libcall, preferably
in generic ISelLowering.
This can be achieved by X86ISelLowering deciding it doesn't want the
faff after all.
llvm-svn: 212134
The logic for expanding atomics that aren't natively supported in
terms of cmpxchg loops is much simpler to express at the IR level. It
also allows the normal optimisations and CodeGen improvements to help
out with atomics, instead of using a limited set of possible
instructions..
rdar://problem/13496295
llvm-svn: 212119
This patch adds support for a new builtin instruction called
__builtin_ia32_rdpmc.
Builtin '__builtin_ia32_rdpmc' is defined as a 'GCC builtin'; on X86, it can
be used to read performance monitoring counters. It takes as input the index
of the performance counter to read, and returns the value of the specified
performance counter as a 64-bit number.
Calls to this new builtin will map to instruction RDPMC.
The index in input to the builtin call is moved to register %ECX. The result
of the builtin call is the value of the specified performance counter (RDPMC
would return that quantity in registers RDX:RAX).
This patch:
- Adds builtin int_x86_rdpmc as a GCCBuiltin;
- Adds a new x86 DAG node called 'RDPMC_DAG';
- Teaches how to lower this new builtin;
- Adds an ISel pattern to select instruction RDPMC;
- Fixes the definition of instruction RDPMC adding %RAX and %RDX as
implicit definitions, and adding %ECX as implicit use;
- Adds a LLVM test to verify that the new builtin is correctly selected.
llvm-svn: 212049
lowering for v16i8.
ASan and some bots caught this bug with existing test cases. Fixing it
even fixed a miscompile with one of the test cases. I'm still a bit
suspicious of this test case as I've not taken a proper amount of time
to think about it, but the fix here is strict goodness.
llvm-svn: 211976
These show up really frequently, not the least with actual splats. =] We
lowered these quite badly before. The new code path tries to widen i8
shuffles to i16 shuffles in a splat-like way. There are still some
inefficiencies in our i16 splat logic though, so we aren't really done
here.
Also, for certain patterns (bit of a gather-and-splat) we still
generate pretty silly code, and I've left a fixme for addressing it.
However, I'm not actually worried about this code pattern as much. The
old shuffle lowering generates a 29 instruction monstrosity for it that
should execute much more slowly.
llvm-svn: 211974
lowering.
For maximum irony, I had already discovered this bug, diagnosed it, and
left FIXMEs about it in the test cases. =[ I just failed to go back over
those until after i had reduced a bootstrap miscompile down to a single
TU, stared at the assembly for an hour, and figured out the bug. Again.
Oh well.
llvm-svn: 211955
a bootstrap.
I managed to mis-remember how PACKUS worked on x86, and was using undef
for the high bytes instead of zero. The fix is fairly obvious.
llvm-svn: 211922
Summary:
This allows it to fold pshufd instructions across intervening
half-shuffles and other noise. This pattern actually shows up in the
generic lowering tests, but I've also added direct tests using
intrinsics to make sure that the specific desired functionality is
working even if the lowering stuff changes in the future.
Differential Revision: http://reviews.llvm.org/D4292
llvm-svn: 211892
half-shuffles, even looking through intervening instructions in a chain.
Summary:
This doesn't happen to show up with any test cases I've found for the current
shuffle lowering, but previous attempts would benefit from this and it seems
generally useful. I've tested it directly using intrinsics, which also shows
that it will work with hand vectorized code as well.
Note that even though pshufd isn't directly used in these tests, it gets
exercised because we combine some of the half shuffles into a pshufd
first, and then merge them.
Differential Revision: http://reviews.llvm.org/D4291
llvm-svn: 211890
trivially redundant.
This fixes several cases in the new vector shuffle lowering algorithm
which would generate redundant shuffle instructions for the sake of
simplicity.
I'm also deleting a testcase which was somewhat ridiculous. It was
checking for a bug in 2007 about incorrectly transforming shuffles by
looking for the string "-86" in the output of a pretty substantial
function. This test case doesn't seem to have any value at this point.
Differential Revision: http://reviews.llvm.org/D4240
llvm-svn: 211889
x86 backend.
This sketches out a new code path for vector lowering, hidden behind an
off-by-default flag while it is under development. The fundamental idea
behind the new code path is to aggressively break down the problem space
in ways that ease selecting the odd set of instructions available on
x86, and carefully avoid scalarizing code even when forced to use older
ISAs. Notably, this starts off restricting itself to SSE2 and implements
the complete vector shuffle and blend space for 128-bit vectors in SSE2
without scalarizing. The plan is to layer on top of this ISA extensions
where we can bail out of the complex SSE2 lowering and opt for
a cheaper, specialized instruction (or set of instructions). It also
needs to be generalized to AVX and AVX512 vector widths.
Currently, this does a decent but not perfect job for SSE2. There are
some specific shortcomings that I plan to address:
- We need a peephole combine to fold together shuffles where possible.
There are cases where a previous shuffle could be modified slightly to
arrange for elements to be in the correct position and a later shuffle
eliminated. Doing this eagerly added quite a bit of complexity, and
so my plan is to combine away these redundancies afterward.
- There are a lot more clever ways to use unpck and pack that need to be
added. This is essential for real world shuffles as it turns out...
Once SSE2 is polished a bit I should be able to get interesting numbers
on performance improvements on benchmarks conducive to vectorization.
All of this will be off by default until it is functionally equivalent
of course.
Differential Revision: http://reviews.llvm.org/D4225
llvm-svn: 211888
This patch teaches the backend how to canonicalize a shuffle vectors
according to the rule:
- (shuffle (FADD A, B), (FSUB A, B), Mask) ->
(shuffle (FSUB A, -B), (FADD A, -B), Mask)
Where 'Mask' is:
<0,5,2,7> ;; for v4f32 and v4f64 shuffles.
<0,3> ;; for v2f64 shuffles.
<0,9,2,11,4,13,6,15> ;; for v8f32 shuffles.
In general, ISel only knows how to pattern-match a canonical
'fadd + fsub + blendi' dag node sequence into an ADDSUB instruction.
This new rule allows to convert a non-canonical dag sequence into a
canonical one that will be matched by a single ADDSUB at ISel stage.
The idea of converting a non-canonical ADDSUB into a canonical one by
swapping the first two operands of the shuffle, and then negating the
second operand of the FADD and FSUB, was originally proposed by Hal Finkel.
llvm-svn: 211771
This patch teaches method 'LowerVECTOR_SHUFFLE' to give higher precedence to
the check for 'isBlendMask'; the idea is that, when possible, we should firstly
check if a shuffle performs a blend, and in case, try to lower it into a BLENDI
instead of selecting a SHUFP or (worse) a VPERM2X128.
In general:
- AVX VBLENDPS/D always have better latency and throughput than VPERM2F128;
- BLENDPS/D instructions tend to always have better 'reciprocal throughput'
than the equivalent SHUFPS/D;
- Both BLENDPS/D and SHUFPS/D are often decoded into the same number of
m-ops; however, a m-op obtained from a BLENDPS/D can be scheduled to more
than one execution port.
This patch:
- Moves the check for 'isBlendMask' immediately before the check for
'isSHUFPMask' within method 'LowerVECTOR_SHUFFLE';
- Updates existing tests for sse/avx shuffle/blend instructions to verify
that we select (v)blendps/d when possible (instead of (v)shufps/d or
vperm2f128).
llvm-svn: 211720
--
This patch enables LLVM to emit Win64-native unwind info rather than
DWARF CFI. It handles all corner cases (I hope), including stack
realignment.
Because the unwind info is not flexible enough to describe stack frames
with a gap of unknown size in the middle, such as the one caused by
stack realignment, I modified register spilling code to place all spills
into the fixed frame slots, so that they can be accessed relative to the
frame pointer.
Patch by Vadim Chugunov!
Reviewed By: rnk
Differential Revision: http://reviews.llvm.org/D4081
llvm-svn: 211691
This patch teaches the backend how to combine a build_vector that implements
an 'addsub' between packed float vectors into a sequence of vector add
and vector sub followed by a VSELECT.
The new VSELECT is expected to be lowered into a BLENDI.
At ISel stage, the sequence 'vector add + vector sub + BLENDI' is
pattern-matched against ISel patterns added at r211427 to select
'addsub' instructions.
Added three more ISel patterns for ADDSUB.
Added test sse3-avx-addsub-2.ll to verify that we correctly emit 'addsub'
instructions.
llvm-svn: 211679
This patch enables LLVM to emit Win64-native unwind info rather than
DWARF CFI. It handles all corner cases (I hope), including stack
realignment.
Because the unwind info is not flexible enough to describe stack frames
with a gap of unknown size in the middle, such as the one caused by
stack realignment, I modified register spilling code to place all spills
into the fixed frame slots, so that they can be accessed relative to the
frame pointer.
Patch by Vadim Chugunov!
Reviewed By: rnk
Differential Revision: http://reviews.llvm.org/D4081
llvm-svn: 211399
instructions available as synthetic SDNodes PACKSS and PACKUS that will
select to the correct instruction variants based on the return type.
This allows us to use these rather important instructions when lowering
vector shuffles.
Also moves the relevant instruction definitions to be split out from
the fully generic multiclasses to allow them to match these new SDNodes
in the same way that the UNPCK instructions do.
No functionality should actually be changed here.
llvm-svn: 211332
Before this change, the backend was unable to fold a build_vector dag
node with UNDEF operands into a single horizontal add/sub.
This patch teaches how to combine a build_vector with UNDEF operands into a
horizontal add/sub when possible. The algorithm conservatively avoids to combine
a build_vector with only a single non-UNDEF operand.
Added test haddsub-undef.ll to verify that we correctly fold horizontal binop
even in the presence of UNDEFs.
llvm-svn: 211265
Lowering this new node allows us to fold the almost universal
comparison for success before it's even formed. Instead we can create
a copy from EFLAGS and an X86ISD::SETCC operation since all "cmpxchg"
instructions set the zero-flag to the correct value.
rdar://problem/13201607
llvm-svn: 210923
This commit adds a weak variant of the cmpxchg operation, as described
in C++11. A cmpxchg instruction with this modifier is permitted to
fail to store, even if the comparison indicated it should.
As a result, cmpxchg instructions must return a flag indicating
success in addition to their original iN value loaded. Thus, for
uniformity *all* cmpxchg instructions now return "{ iN, i1 }". The
second flag is 1 when the store succeeded.
At the DAG level, a new ATOMIC_CMP_SWAP_WITH_SUCCESS node has been
added as the natural representation for the new cmpxchg instructions.
It is a strong cmpxchg.
By default this gets Expanded to the existing ATOMIC_CMP_SWAP during
Legalization, so existing backends should see no change in behaviour.
If they wish to deal with the enhanced node instead, they can call
setOperationAction on it. Beware: as a node with 2 results, it cannot
be selected from TableGen.
Currently, no use is made of the extra information provided in this
patch. Test updates are almost entirely adapting the input IR to the
new scheme.
Summary for out of tree users:
------------------------------
+ Legacy Bitcode files are upgraded during read.
+ Legacy assembly IR files will be invalid.
+ Front-ends must adapt to different type for "cmpxchg".
+ Backends should be unaffected by default.
llvm-svn: 210903
When I originally added node RDTSCP_DAG (r207127) I forgot to add
a string name for it in method 'getTargetNodeName'.
No functional change intended.
llvm-svn: 210769
This patch adds target combine rules to match:
- [AVX] Horizontal add/sub of packed single/double precision floating point
values from 256-bit vectors;
- [AVX2] Horizontal add/sub of packed integer values from 256-bit vectors.
llvm-svn: 210761
This patch moves part of the logic implemented by the target specific
combine rules added at r210477 to a separate helper function.
This should make easier to add more rules for matching AVX/AVX2 horizontal
adds/subs.
This patch also fixes a problem caused by a wrong check performed on indices
of extract_vector_elt dag nodes in input to the scalar adds/subs.
New tests have been added to verify that we correctly check indices of
extract_vector_elt dag nodes when selecting a horizontal operation.
llvm-svn: 210644
This patch slightly changes the algorithm introduced at revision 210477
to fix a problem where the algorithm was producing incorrect code for
the VEX.256 encoded versions of horizontal add/sub.
For these cases, we now try to split the two 256-bit vectors into
128-bit chunks before emitting horizontal add/sub dag nodes.
Added a new test case into haddsub-2.ll.
llvm-svn: 210545
The SelectionDAG bad a special case for ISD::SELECT_CC, where it would
allow targets to specify:
setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
to indicate that they wanted to expand ISD::SELECT_CC for all types.
This wasn't applied correctly everywhere, and it makes writing new
DAG patterns with ISD::SELECT_CC difficult.
llvm-svn: 210541
The C++ and C semantics of the compare_and_swap operations actually
require us to return a boolean "success" value. In LLVM terms this
means a second comparison of the output of "cmpxchg" against the input
desired value.
However, x86's "cmpxchg" instruction sets all flags for the comparison
formed, so we can skip any secondary comparison. (N.b. this isn't true
for cmpxchg8b/16b, which only set ZF).
rdar://problem/13201607
llvm-svn: 210523
This patch adds new target specific combine rules to identify horizontal
add/sub idioms from BUILD_VECTOR dag nodes.
This patch also teaches the DAGCombiner how to canonicalize sequences of
insert_vector_elt dag nodes according to the following rule:
(insert_vector_elt (insert_vector_elt A, I0), I1) ->
(insert_vecto_elt (insert_vector_elt A, I1), I0)
This new canonicalization rule only triggers if the inner insert_vector
dag node has exactly one use; also, both indices must be known constants,
and I1 < I0.
This last rule made it possible to write a simpler algorithm to identify
horizontal add/sub patterns because now we don't have to worry about the
ordering of insert_vector_elt dag nodes.
llvm-svn: 210477
This patch teaches the backend how to check for the 'NoSignedWrap' flag on
binary operations to improve the emission of 'test' instructions.
If the result of a binary operation is known not to overflow we know that
resetting the Overflow flag is unnecessary and so we can avoid emitting
the test instruction.
Patch by Marcello Maggioni.
llvm-svn: 210468
According to Intel Software Optimization Manual
on Silvermont INC or DEC instructions require
an additional uop to merge the flags.
As a result, a branch instruction depending
on an INC or a DEC instruction incurs a 1 cycle penalty.
Differential Revision: http://reviews.llvm.org/D3990
llvm-svn: 210466
Summary:
We were being too strict and not accounting for undefs.
Added a test case and fixed another one where we improved codegen.
Reviewers: grosbach, nadav, delena
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D4039
llvm-svn: 210361
When lowering a ISD::BRCOND into a test+branch, make sure that we
always use the correct condition code to emit the test operation.
This fixes PR19858: "i8 checked mul is wrong on x86".
Patch by Keno Fisher!
llvm-svn: 210032
This patch teaches the backend how to simplify/canonicalize dag node
sequences normally introduced by the backend when promoting certain dag nodes
with illegal vector type.
This patch adds two new combine rules:
1) fold (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
(shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
2) fold (BINOP (shuffle (A, Undef, <Mask>)), (shuffle (B, Undef, <Mask>))) ->
(shuffle (BINOP A, B), Undef, <Mask>).
Both rules are only triggered on the type-legalized DAG.
In particular, rule 1. is a target specific combine rule that attempts
to sink a bitconvert into the operands of a binary operation.
Rule 2. is a target independet rule that attempts to move a shuffle
immediately after a binary operation.
llvm-svn: 209930
Summary:
Separate the check for blend shuffle_vector masks into isBlendMask.
This function will also be used to check if a vector shuffle is legal. No
change in functionality was intended, but we ended up improving codegen on
two tests, which were being (more) optimized only if the resulting shuffle
was legal.
Reviewers: nadav, delena, andreadb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3964
llvm-svn: 209923
This patch teaches the x86 backend how to efficiently lower ISD::BITCAST dag
nodes from MVT::f64 to MVT::v4i16 (and vice versa), and from MVT::f64 to
MVT::v8i8 (and vice versa).
This patch extends the logic from revision 208107 to also handle MVT::v4i16
and MVT::v8i8. Also, this patch correctly propagates Undef values when
performing the widening of a vector (example: when widening from v2i32 to
v4i32, the upper 64bits of the resulting vector are 'undef').
llvm-svn: 209451
ISD::VSELECT mask uses 1 to identify the first argument and 0 to identify the
second argument.
On the other hand, BLENDI uses 0 to identify the first argument and 1 to
identify the second argument.
Fix the generation of the blend mask to account for this difference.
The bug did not show up with r209043, because we were not checking for the
actual arguments of the blend instruction!
This commit also fixes the test cases.
Note: The same mask works for the BLENDr variant because the arguments are
swapped during instruction selection (see the BLENDXXrr patterns).
<rdar://problem/16975435>
llvm-svn: 209324
Summary:
When inserting an element that's coming from a vector load or a broadcast
of a vector (or scalar) load, combine the load into the insertps
instruction.
Added PerformINSERTPSCombine for the case where we need to fix the load
(load of a vector + insertps with a non-zero CountS).
Added patterns for the broadcasts.
Also added tests for SSE4.1, AVX, and AVX2.
Reviewers: delena, nadav, craig.topper
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3581
llvm-svn: 209156
- On ARM/ARM64 we get a vrev because the shuffle matching code is really smart. We still unroll anything that's not v4i32 though.
- On X86 we get a pshufb with SSSE3. Required more cleverness in isShuffleMaskLegal.
- On PPC we get a vperm for v8i16 and v4i32. v2i64 is unrolled.
llvm-svn: 209123
This is mostly a mechanical change changing all the call sites to the newer
chained-function construction pattern. This removes the horrible 15-parameter
constructor for the CallLoweringInfo in favour of setting properties of the call
via chained functions. No functional change beyond the removal of the old
constructors are intended.
llvm-svn: 209082
were added in SSE2, no SSSE3. Found this while auditing all uses of
SSSE3 in the X86 target. I don't actually expect this to make
a significant difference on anything and I don't have any detailed test
cases but I updated the existing test cases that already covered some of
this code path.
llvm-svn: 209056
vselects with constant masks, after legalization, will get turned into
specialized shuffle_vectors so they can be matched to blend+imm
instructions.
Fixed some tests.
llvm-svn: 209044
LowerVSELECT will, if possible, generate a X86ISD::BLENDI DAG node if the
condition is constant and we can emit that instruction, given the
subtarget.
This is not enough for all cases. An additional SELECTCombine optimization
will be committed.
Fixed tests that were expecting variable blends but where a blend+imm can
be generated.
Added test where we can't emit blend+immediate.
Added avx2 blend+imm tests.
llvm-svn: 209043
No functionality change intended. The types that previously were set to
lower as Expand or Legal are doing the same thing with this lowering
function.
llvm-svn: 209042
Added target specific combine rules to fold blend intrinsics according
to the following rules:
1) fold(blend A, A, Mask) -> A;
2) fold(blend A, B, <allZeros>) -> A;
3) fold(blend A, B, <allOnes>) -> B.
Added two new tests to verify that the new folding rules work for all
the optimized blend intrinsics.
llvm-svn: 208895
r208453 added support for having sret on the second parameter. In that
change, the code for copying sret into a virtual register was hoisted
into the loop that lowers formal parameters. This caused a "Wrong
topological sorting" assertion failure during scheduling when a
parameter is passed in memory. This change undoes that by creating a
second loop that deals with sret.
I'm worried that this fix is incomplete. I don't fully understand the
dependence issues. However, with this change we produce the same DAGs
we used to produce, so if they are broken, they are just as broken as
they have always been.
llvm-svn: 208637
1) Changed gather and scatter intrinsics. Now they are aligned with GCC built-ins. There is no more non-masked form. Masked intrinsic receives -1 if all lanes are executed.
2) I changed the function that works with intrinsics inside X86ISelLowering.cpp. I put all intrinsics in one table. I did it for INTRINSICS_W_CHAIN and plan to put all intrinsics from WO_CHAIN set to the same table in order to avoid the long-long "switch". (I wanted to use static map initialization that allowed by C++11 but I wasn't able to compile it on VS2012).
3) I added gather/scatter prefetch intrinsics.
4) I fixed MRMm encoding for masked instructions.
llvm-svn: 208522
We must validate the value type in TLI::getRegisterByName, because if we
don't and the wrong type was used with the IR intrinsic, then we'll assert
(because we won't be able to find a valid register class with which to
construct the requested copy operation). For PPC64, additionally, the type
information is necessary to decide between the 64-bit register and the 32-bit
subregister.
No functionality change.
llvm-svn: 208508
When lowering build_vector to an insertps, we would still lower it, even
if the source vectors weren't v4x32. This would break on avx if the source
was a v8x32. We now check the type of the source vectors.
llvm-svn: 208487
MSVC always places the implicit sret parameter after the implicit this
parameter of instance methods. We used to handle this for
x86_thiscallcc by allocating the sret parameter on the stack and leaving
the this pointer in ecx, but that doesn't handle alternative calling
conventions like cdecl, stdcall, fastcall, or the win64 convention.
Instead, change the verifier to allow sret on the second parameter.
This also requires changing the Mips and X86 backends to return the
argument with the sret parameter, instead of assuming that the sret
parameter comes first.
The Sparc backend also returns sret parameters in a register, but I
wasn't able to update it to handle secondary sret parameters. It
currently calls report_fatal_error if you feed it an sret in the second
parameter.
Reviewers: rafael.espindola, majnemer
Differential Revision: http://reviews.llvm.org/D3617
llvm-svn: 208453
Summary:
Also ran clang-format on the function. The code added is the last else
if block.
Reviewers: nadav, craig.topper, delena
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3518
llvm-svn: 208372
This patch teaches the backend how to combine packed SSE2/AVX2 arithmetic shift
intrinsics.
The rules are:
- Always fold a packed arithmetic shift by zero to its first operand;
- Convert a packed arithmetic shift intrinsic dag node into a ISD::SRA only if
the shift count is known to be smaller than the vector element size.
This patch also teaches to function 'getTargetVShiftByConstNode' how fold
target specific vector shifts by zero.
Added two new tests to verify that the DAGCombiner is able to fold
sequences of SSE2/AVX2 packed arithmetic shift calls.
llvm-svn: 208342
Summary:
Vectors built with zeros and elements in the same order as another
(source) vector are optimized to be built using a single insertps
instruction.
Also optimize when we move one element in a vector to a different place
in that vector while zeroing out some of the other elements.
Further optimizations are possible, described in TODO comments.
I will be implementing at least some of them in the near future.
Added some tests for different cases where this optimization triggers.
Reviewers: nadav, delena, craig.topper
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3521
llvm-svn: 208271
Before this patch, the backend always emitted a store+load sequence to
bitconvert from f64 to i64 the input operand of a ISD::BITCAST dag node that
performed a bitconvert from type MVT::f64 to type MVT::v2i32. The resulting
i64 node was then used to build a v2i32 vector.
With this patch, the backend now produces a cheaper SCALAR_TO_VECTOR from
MVT::f64 to MVT::v2f64. That SCALAR_TO_VECTOR is then followed by a "free"
bitcast to type MVT::v4i32. The elements of the resulting
v4i32 are then extracted to build a v2i32 vector (which is illegal and
therefore promoted to MVT::v2i64).
This is in general cheaper than emitting a stack store+load sequence
to bitconvert the operand from type f64 to type i64.
llvm-svn: 208107
This patch implements the infrastructure to use named register constructs in
programs that need access to specific registers (bare metal, kernels, etc).
So far, only the stack pointer is supported as a technology preview, but as it
is, the intrinsic can already support all non-allocatable registers from any
architecture.
llvm-svn: 208104
The Win64 docs are very clear that anything larger than 8 bytes is
passed by reference, and GCC MinGW64 honors that for __modti3 and
friends.
Patch by Jameson Nash!
llvm-svn: 208029
Summary:
Also ran clang-format on the function. The code added is the last else
if block.
Reviewers: nadav, craig.topper
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3518
llvm-svn: 207992
Currently, musttail codegen is relying on sibcall optimization, and
reporting a fatal error if fails. Sibcall optimization fails when stack
arguments need to be modified, which is insufficient for musttail.
The logic for moving arguments in memory safely is already implemented
for GuaranteedTailCallOpt. This change merely arranges for musttail
calls to use it.
No functional change for GuaranteedTailCallOpt.
Reviewers: espindola
Differential Revision: http://reviews.llvm.org/D3493
llvm-svn: 207598
Scaling factors are not free on X86 because every "complex" addressing mode
breaks the related instruction into 2 allocations instead of 1.
<rdar://problem/16730541>
llvm-svn: 207301
Summary:
If we're doing a v4f32/v4i32 shuffle on x86 with SSE4.1, we can lower
certain shufflevectors to an insertps instruction:
When most of the shufflevector result's elements come from one vector (and
keep their index), and one element comes from another vector or a memory
operand.
Added tests for insertps optimizations on shufflevector.
Added support and tests for v4i32 vector optimization.
Reviewers: nadav
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D3475
llvm-svn: 207291
This is similar to the 'tail' marker, except that it guarantees that
tail call optimization will occur. It also comes with convervative IR
verification rules that ensure that tail call optimization is possible.
Reviewers: nicholas
Differential Revision: http://llvm-reviews.chandlerc.com/D3240
llvm-svn: 207143
This patch:
- Adds two new X86 builtin intrinsics ('int_x86_rdtsc' and
'int_x86_rdtscp') as GCCBuiltin intrinsics;
- Teaches the backend how to lower the two new builtins;
- Introduces a common function to lower READCYCLECOUNTER dag nodes
and the two new rdtsc/rdtscp intrinsics;
- Improves (and extends) the existing x86 test 'rdtsc.ll'; now test 'rdtsc.ll'
correctly verifies that both READCYCLECOUNTER and the two new intrinsics
work fine for both 64bit and 32bit Subtargets.
llvm-svn: 207127
This allows us to compile
return (mask & 0x8 ? a : b);
into
testb $8, %dil
cmovnel %edx, %esi
instead of
andl $8, %edi
shrl $3, %edi
cmovnel %edx, %esi
which we formed previously because dag combiner canonicalizes setcc of and into shift.
llvm-svn: 207088
Generating BZHI in the variable mask case, i.e. (and X, (sub (shl 1, N), 1)),
was already supported, but we were missing the constant-mask case. This patch
fixes that.
<rdar://problem/15480077>
llvm-svn: 206738
For a 256-bit BUILD_VECTOR consisting mostly of shuffles of 256-bit vectors,
both the BUILD_VECTOR and its operands may need to be legalized in multiple
steps. Consider:
(v8f32 (BUILD_VECTOR (extract_vector_elt (v8f32 %vreg0,) Constant<1>),
(extract_vector_elt %vreg0, Constant<2>),
(extract_vector_elt %vreg0, Constant<3>),
(extract_vector_elt %vreg0, Constant<4>),
(extract_vector_elt %vreg0, Constant<5>),
(extract_vector_elt %vreg0, Constant<6>),
(extract_vector_elt %vreg0, Constant<7>),
%vreg1))
a. We can't build a 256-bit vector efficiently so, we need to split it into
two 128-bit vecs and combine them with VINSERTX128.
b. Operands like (extract_vector_elt (v8f32 %vreg0), Constant<7>) needs to be
split into a VEXTRACTX128 and a further extract_vector_elt from the
resulting 128-bit vector.
c. The extract_vector_elt from b. is lowered into a shuffle to the first
element and a movss.
Depending on the order in which we legalize the BUILD_VECTOR and its
operands[1], buildFromShuffleMostly may be faced with:
(v4f32 (BUILD_VECTOR (extract_vector_elt
(vector_shuffle<1,u,u,u> (extract_subvector %vreg0, Constant<4>), undef),
Constant<0>),
(extract_vector_elt
(vector_shuffle<2,u,u,u> (extract_subvector %vreg0, Constant<4>), undef),
Constant<0>),
(extract_vector_elt
(vector_shuffle<3,u,u,u> (extract_subvector %vreg0, Constant<4>), undef),
Constant<0>),
%vreg1))
In order to figure out the underlying vector and their identity we need to see
through the shuffles.
[1] Note that the order in which operations and their operands are legalized is
only guaranteed in the first iteration of LegalizeDAG.
Fixes <rdar://problem/16296956>
llvm-svn: 206634
This patch teaches the backend how to efficiently lower logical and
arithmetic packed shifts on both SSE and AVX/AVX2 machines.
When possible, instead of scalarizing a vector shift, the backend should try
to expand the shift into a sequence of two packed shifts by immedate count
followed by a MOVSS/MOVSD.
Example
(v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
Can be rewritten as:
(v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
[with X and Y ConstantInt]
The advantage is that the two new shifts from the example would be lowered into
X86ISD::VSRLI nodes. This is always cheaper than scalarizing the vector into
four scalar shifts plus four pairs of vector insert/extract.
llvm-svn: 206316
I found this from a particular GDB test suite case of inlining
(something similar is provided as a test case) but came across a few
other related cases (other callers of the same functions, and one other
instance of the same coding mistake in a separate function).
I'm not sure what the best way to test this is (let alone to cover the
other cases I discovered), so hopefully this sufficies - open to ideas.
llvm-svn: 206130
This removes the -segmented-stacks command line flag in favor of a
per-function "split-stack" attribute.
Patch by Luqman Aden and Alex Crichton!
llvm-svn: 205997
and isTargetCygwin() to isTargetWindowsCygwin() to be consistent with the
four Windows environments in Triple.h.
Suggestion by Saleem Abdulrasool!
llvm-svn: 205393
This adds back r204781.
Original message:
Aliases are just another name for a position in a file. As such, the
regular symbol resolutions are not applied. For example, given
define void @my_func() {
ret void
}
@my_alias = alias weak void ()* @my_func
@my_alias2 = alias void ()* @my_alias
We produce without this patch:
.weak my_alias
my_alias = my_func
.globl my_alias2
my_alias2 = my_alias
That is, in the resulting ELF file my_alias, my_func and my_alias are
just 3 names pointing to offset 0 of .text. That is *not* the
semantics of IR linking. For example, linking in a
@my_alias = alias void ()* @other_func
would require the strong my_alias to override the weak one and
my_alias2 would end up pointing to other_func.
There is no way to represent that with aliases being just another
name, so the best solution seems to be to just disallow it, converting
a miscompile into an error.
llvm-svn: 204934
This reverts commit r204781.
I will follow up to with msan folks to see what is what they
were trying to do with aliases to weak aliases.
llvm-svn: 204784
Aliases are just another name for a position in a file. As such, the
regular symbol resolutions are not applied. For example, given
define void @my_func() {
ret void
}
@my_alias = alias weak void ()* @my_func
@my_alias2 = alias void ()* @my_alias
We produce without this patch:
.weak my_alias
my_alias = my_func
.globl my_alias2
my_alias2 = my_alias
That is, in the resulting ELF file my_alias, my_func and my_alias are
just 3 names pointing to offset 0 of .text. That is *not* the
semantics of IR linking. For example, linking in a
@my_alias = alias void ()* @other_func
would require the strong my_alias to override the weak one and
my_alias2 would end up pointing to other_func.
There is no way to represent that with aliases being just another
name, so the best solution seems to be to just disallow it, converting
a miscompile into an error.
llvm-svn: 204781
This used to resort to splitting the 256-bit operation into two 128-bit
shuffles and then recombining the results.
Fixes <rdar://problem/16167303>
llvm-svn: 204735
I found three implementations of this. This splits it out into a new function
and uses it from the three places.
My plan is to add a fourth use when lowering a vector_shuffle:v16i16.
Compared the assembly output of test/CodeGen/X86 before and after.
The only change is due to how the first PSHUFB was generated in
LowerVECTOR_SHUFFLEv8i16. If the shuffle mask specified undef (i.e. -1), the
old implementation would write -1 * 2 and -1 * 2 + 1 (254 and 255) in the
control mask. Now we write 0x80. These are of course interchangeable since
bit 7 decides if a constant zero is written in the result byte. The other
instances of this code use 0x80 consistently.
Related to <rdar://problem/16167303>
llvm-svn: 204734
This can be observed with the old testcase of CodeGen/X86/pr12312.ll:
47c47
< vorps %ymm0, %ymm1, %ymm0
---
> vorps %ymm1, %ymm0, %ymm0
97c97
< vorps %ymm1, %ymm0, %ymm0
---
> vorps %ymm0, %ymm1, %ymm0
The vector VecIns is populated with all the values from VecInMap. This is done
while iterating VecInMap. VecInMap uses a hash of pointer values so the
resulting order can vary depending on the memory layout.
The fix is to populate the vector VecIns earlier as VecInMap is populated.
This is done in DAG traversal order.
Fixes <rdar://problem/16398806>
llvm-svn: 204623
Rather than LegalizeAction::Expand, this needs LegalizeAction::Promote to get
promoted to fp_to_sint v8f32->v8i32. This is a legal operation on AVX.
For that to work properly, we also need to teach the legalizer about the
specific promotion required here. The default vector promotion uses
bitcasting to a vector type of the same total size. We want to promote the
vector element type, effectively widening the operation and then truncating
the result. This is analogous to the current logic of how int_to_fp is
promoted.
The change also factors out some code from the int_to_fp promotion code to
ValueType::widenIntegerVectorElementType. This is now shared between
int_to_fp and fp_to_int.
There is no longer need for the custom lowering of fp_to_sint f32->v8i16 in
X86. It can now go through the new target-independent fp_to_*int promotion
logic.
I also checked that no other target uses Promote for these ops yet, so there
shouldn't be any unexpected change in behavior.
Fixes <rdar://problem/16202247>
llvm-svn: 204058
operator* on the by-operand iterators to return a MachineOperand& rather than
a MachineInstr&. At this point they almost behave like normal iterators!
Again, this requires making some existing loops more verbose, but should pave
the way for the big range-based for-loop cleanups in the future.
llvm-svn: 203865
This fixes the bug where we would bitcast the 64-bit floating point result
of cmpneqsd to a 64-bit integer even on 32-bit targets.
Differential Revision: http://llvm-reviews.chandlerc.com/D3009
llvm-svn: 203581
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 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
llvm-svn: 203524
The current approach to lower a vsetult is to flip the sign bit of the
operands, swap the operands and then use a (signed) pcmpgt. psubus (unsigned
saturating subtract) can be used to emulate a vsetult more efficiently:
+ case ISD::SETULT: {
+ // If the comparison is against a constant we can turn this into a
+ // setule. With psubus, setule does not require a swap. This is
+ // beneficial because the constant in the register is no longer
+ // destructed as the destination so it can be hoisted out of a loop.
I also enable lowering via psubus in a few other cases where it's clearly
beneficial: setule and setuge if minu/maxu cannot be used.
rdar://problem/14338765
Patch by Adam Nemet <anemet@apple.com>.
llvm-svn: 202301
On x86, shifting a vector by a scalar is significantly cheaper than shifting a
vector by another fully general vector. Unfortunately, because SelectionDAG
operates on just one basic block at a time, the shufflevector instruction that
reveals whether the right-hand side of a shift *is* really a scalar is often
not visible to CodeGen when it's needed.
This adds another handler to CodeGenPrepare, to sink any useful shufflevector
instructions down to the basic block where they're used, predicated on a target
hook (since on other architectures, doing so will often just introduce extra
real work).
rdar://problem/16063505
llvm-svn: 201655
Instead of expanding a packed shift into a sequence of scalar shifts,
the backend now tries (when possible) to convert the vector shift into a
vector multiply.
Before this change, a shift of a MVT::v8i16 vector by a
build_vector of constants was always scalarized into a long sequence of "vector
extracts + scalar shifts + vector insert".
With this change, if there is SSE2 support, we emit a single vector multiply.
This change also affects SSE4.1, AVX, AVX2 shifts:
- A shift of a MVT::v4i32 vector by a build_vector of non uniform constants
is now lowered when possible into a single SSE4.1 vector multiply.
- Packed v16i16 shift left by constant build_vector are now expanded when
possible into a single AVX2 vpmullw.
This change also improves the lowering of AVX512f vector shifts.
Added test CodeGen/X86/vec_shift6.ll with some code examples that are affected
by this change.
llvm-svn: 201271
I believe VZEXT_MOVL means "zero all vector elements except the first" (and
should have identical input & output types) whereas VZEXT means "zero extend
each element of a vector (discarding higher elements if necessary)".
For example:
(v4i32 (vzext (v16i8 ...)))
should zero extend the low 4 bytes of the incoming vector to 32-bits,
discarding higher bytes.
However, somewhere in the past, these two concepts had become confused, even
leading to a nonsensical VSEXT_MOVL.
This re-merges the nodes where appropriate (all VSEXT_MOVL -> VSEXT, VZEXT_MOVL
-> VZEXT when it's an actual extension).
rdar://problem/15981990
llvm-svn: 200918
Calls with inalloca are lowered by skipping all stores for arguments
passed in memory and the initial stack adjustment to allocate argument
memory.
Now the frontend is responsible for the memory layout, and the backend
doesn't have to do any work. As a result these changes are pretty
minimal.
Reviewers: echristo
Differential Revision: http://llvm-reviews.chandlerc.com/D2637
llvm-svn: 200596
Before this patch we used getIntImmCost from TargetTransformInfo to determine if
a load of a constant should be converted to just a constant, but the threshold
for this was set to an arbitrary value. This value works well for the two
targets (X86 and ARM) that implement this target-hook, but it isn't
target-independent at all.
Now targets have the possibility to decide directly if this optimization should
be performed. The default value is set to false to preserve the current
behavior. The target hook has been moved to TargetLowering, which removed the
last use and need of TargetTransformInfo in SelectionDAG.
llvm-svn: 200271
This commit teaches the X86 backend to create the same X86 instructions when it
lowers an sadd/ssub with overflow intrinsic and a conditional branch that uses
that overflow result. This allows SelectionDAG to recognize and remove one of
the redundant operations.
This fixes <rdar://problem/15874016> and <rdar://problem/15661073>.
Reviewed by Nadav
llvm-svn: 199976
Add target specific rules for combining vselect dag nodes into movss/movsd
when possible.
If the vector type of the vselect dag node in input is either MVT::v4i13 or
MVT::v4f32, then try to fold according to rules:
1) fold (vselect (build_vector (0, -1, -1, -1)), A, B) -> (movss A, B)
2) fold (vselect (build_vector (-1, 0, 0, 0)), A, B) -> (movss B, A)
If the vector type of the vselect dag node in input is either MVT::v2i64 or
MVT::v2f64 (and we have SSE2), then try to fold according to rules:
3) fold (vselect (build_vector (0, -1)), A, B) -> (movsd A, B)
4) fold (vselect (build_vector (-1, 0)), A, B) -> (movsd B, A)
llvm-svn: 199683
MSVC on x64 requires that we create image relative symbol
references to refer to RTTI data. Seeing as how there is no way to
explicitly make reference to a given relocation type in LLVM IR, pattern
match expressions of the form &foo - &__ImageBase.
Differential Revision: http://llvm-reviews.chandlerc.com/D2523
llvm-svn: 199312
promotion code, Tablegen will now select FPExt for floating point promotions
(previously it had returned AExt, which is not valid for floating point types).
Any out-of-tree targets that were relying on AExt being returned for FP
promotions will need to update their code check for FPExt instead.
llvm-svn: 199252
Representing dllexport/dllimport as distinct linkage types prevents using
these attributes on templates and inline functions.
Instead of introducing further mixed linkage types to include linkonce and
weak ODR, the old import/export linkage types are replaced with a new
separate visibility-like specifier:
define available_externally dllimport void @f() {}
@Var = dllexport global i32 1, align 4
Linkage for dllexported globals and functions is now equal to their linkage
without dllexport. Imported globals and functions must be either
declarations with external linkage, or definitions with
AvailableExternallyLinkage.
llvm-svn: 199218
This fixes a regression intruced by r198113.
Revision r198113 introduced an algorithm that tries to fold a vector shift
by immediate count into a build_vector if the input vector is a known vector
of constants.
However the algorithm only worked under the assumption that the input vector
type and the shift type are exactly the same.
This patch disables the folding of vector shift by immediate count if the
input vector type and the shift value type are not the same.
llvm-svn: 199213
Representing dllexport/dllimport as distinct linkage types prevents using
these attributes on templates and inline functions.
Instead of introducing further mixed linkage types to include linkonce and
weak ODR, the old import/export linkage types are replaced with a new
separate visibility-like specifier:
define available_externally dllimport void @f() {}
@Var = dllexport global i32 1, align 4
Linkage for dllexported globals and functions is now equal to their linkage
without dllexport. Imported globals and functions must be either
declarations with external linkage, or definitions with
AvailableExternallyLinkage.
llvm-svn: 199204
This moves the check up into the parent class so that all targets can use it
without having to copy (and keep in sync) the same error message.
llvm-svn: 198579
__builtin_returnaddress requires that the value passed into is be a constant.
However, at -O0 even a constant expression may not be converted to a constant.
Emit an error message intead of crashing.
llvm-svn: 198531
vector shift by immedate count (VSHLI/VSRLI/VSRAI) into a build_vector when
the vector in input to the shift is a build_vector of all constants or UNDEFs.
Target specific nodes for packed shifts by immediate count are in
general introduced by function 'getTargetVShiftByConstNode' (in
X86ISelLowering.cpp) when lowering shift operations, SSE/AVX immediate
shift intrinsics and (only in very few cases) SIGN_EXTEND_INREG dag
nodes.
This patch adds extra rules for simplifying vector shifts inside
function 'getTargetVShiftByConstNode'.
Added file test/CodeGen/X86/vec_shift5.ll to verify that packed
shifts by immediate are correctly folded into a build_vector when the
input vector to the shift dag node is a vector of constants or undefs.
llvm-svn: 198113
This reverts commit r197481, recommiting r197469 with an extra fix.
The vastart_save_xmm_regs pseudo-instruction expands to a test and a
branch, so it modifies EFLAGS. Mark it so, or else the scheduler might
place it in the middle of another test+branch.
This fixes a bug exposed by r192750, which changed the initial scheduler
to source-order as part of enabling the MI Scheduler for X86.
This re-commit changes the VASTART_SAVE_XMM_REGS custom inserter not to
try to save %flags, and adds a test that catches the bad behavior of
r197469.
<rdar://problem/15627766>
llvm-svn: 197503
http://llvm.org/bugs/show_bug.cgi?id=18045
Short issue description:
For X86 machines with sse < sse4.1 we got failures for some
particular load/store vector sequences:
$ clang-trunk -m32 -O2 test-case.c
fatal error: error in backend: Cannot select: 0x4200920: v4i32,ch = load 0x41d6ab0, 0x4205850,
0x41dcb10<LD16[getelementptr inbounds ([4 x i32]* @e, i32 0, i32 0)](align=4)> [ORD=82]
[ID=58]
0x4205850: i32 = X86ISD::Wrapper 0x41d5490 [ORD=26] [ID=43]
0x41d5490: i32 = TargetGlobalAddress<[4 x i32]* @e> 0 [ORD=26] [ID=23]
0x41dcb10: i32 = undef [ID=2]
The reason is that EltsFromConsecutiveLoads could emit such load instruction
both before and after legalize stage. Though this instruction is not legal for
machines with SSSE3 and lower.
The fix: In EltsFromConsecutiveLoads, if we have passed legalize stage, we
check whether nodes it emits are legal.
P.S.: If you get failure in time from 12:00 and till 22:00 (UTC-8),
perhaps I'll slow with response, so you better reject this commit. Thanks!
llvm-svn: 197492
Added scalar compare VCMPSS, VCMPSD.
Implemented LowerSELECT for scalar FP operations.
I replaced FSETCCss, FSETCCsd with one node type FSETCCs.
Node extract_vector_elt(v16i1/v8i1, idx) returns an element of type i1.
llvm-svn: 197384
While it's safe for the X86-specific shift nodes, dag combining will
kill generic nodes. Insert an AND to make it safe, isel will nuke it
as x86's shift instructions have an implicit AND.
Fixes PR16108, which contains a contraption to hit this case in between
constant folders.
llvm-svn: 197228
Most users would be surprised if "isCOFF" and "isMachO" were simultaneously
true, unless they'd put the compiler in a box with a gun attached to a photon
detector.
This makes sure precisely one of the three formats is true for any triple and
simplifies some target logic based on that.
llvm-svn: 196934
target independent.
Most of the x86 specific stackmap/patchpoint handling was necessitated by the
use of the native address-mode format for frame index operands. PEI has now
been modified to treat stackmap/patchpoint similarly to DEBUG_INFO, allowing
us to use a simple, platform independent register/offset pair for frame
indexes on stackmap/patchpoints.
Notes:
- Folding is now platform independent and automatically supported.
- Emiting patchpoints with direct memory references now just involves calling
the TargetLoweringBase::emitPatchPoint utility method from the target's
XXXTargetLowering::EmitInstrWithCustomInserter method. (See
X86TargetLowering for an example).
- No more ugly platform-specific operand parsers.
This patch shouldn't change the generated output for X86.
llvm-svn: 195944
- Fix bug in (vsext (vzext x)) -> (vsext x) in SIGN_EXTEND_IN_REG
lowering where we need to check whether x is a vector type (in-reg
type) of i8, i16 or i32; otherwise, that optimization is not valid.
llvm-svn: 195779
A Direct stack map location records the address of frame index. This
address is itself the value that the runtime requested. This differs
from IndirectMemRefOp locations, which refer to a stack locations from
which the requested values must be loaded. Direct locations can
directly communicate the address if an alloca, while IndirectMemRefOp
handle register spills.
For example:
entry:
%a = alloca i64...
llvm.experimental.stackmap(i32 <ID>, i32 <shadowBytes>, i64* %a)
Since both the alloca and stackmap intrinsic are in the entry block,
and the intrinsic takes the address of the alloca, the runtime can
assume that LLVM will not substitute alloca with any intervening
value. This must be verified by the runtime by checking that the stack
map's location is a Direct location type. The runtime can then
determine the alloca's relative location on the stack immediately after
compilation, or at any time thereafter. This differs from Register and
Indirect locations, because the runtime can only read the values in
those locations when execution reaches the instruction address of the
stack map.
llvm-svn: 195712
Utilizing the 8 and 16 bit comparison instructions, even when an input can
be folded into the comparison instruction itself, is typically not worth it.
There are too many partial register stalls as a result, leading to significant
slowdowns. By always performing comparisons on at least 32-bit
registers, performance of the calculation chain leading to the
comparison improves. Continue to use the smaller comparisons when
minimizing size, as that allows better folding of loads into the
comparison instructions.
rdar://15386341
llvm-svn: 195496
- When simplifying the mask generation for BLEND, check whether that mask is
also consumed by other non-BLEND insns. If true, skip that simplification.
llvm-svn: 195476
AMD's processors family K7, K8, K10, K12, K15 and K16 are known to have SHLD/SHRD instructions with very poor latency. Optimization guides for these processors recommend using an alternative sequence of instructions. For these AMD's processors, I disabled folding (or (x << c) | (y >> (64 - c))) when we are not optimizing for size.
It might be beneficial to disable this folding for some of the Intel's processors. However, since I couldn't find specific recommendations regarding using SHLD/SHRD instructions on Intel's processors, I haven't disabled this peephole for Intel.
llvm-svn: 195383
clang optimizes tail calls, as in this example:
int foo(void);
int bar(void) {
return foo();
}
where the call is transformed to:
calll .L0$pb
.L0$pb:
popl %eax
.Ltmp0:
addl $_GLOBAL_OFFSET_TABLE_+(.Ltmp0-.L0$pb), %eax
movl foo@GOT(%eax), %eax
popl %ebp
jmpl *%eax # TAILCALL
However, the GOT references must all be resolved at dlopen() time, and so this
approach cannot be used with lazy dynamic linking (e.g. using RTLD_LAZY), which
usually populates the PLT with stubs that perform the actual resolving.
This patch changes X86TargetLowering::LowerCall() to skip tail call
optimization, if the called function is a global or external symbol.
Patch by Dimitry Andric!
PR15086
llvm-svn: 195318
This patch reapplies r193676 with an additional fix for the Hexagon backend. The
SystemZ backend has already been fixed by r194148.
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask type for the given target. Now the type
legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
Reviewed by Nadav
llvm-svn: 194542
This patch moves the jump address materialization inside the noop slide. This
enables patching of the materialization itself or its complete removal. This
patch also adds the ability to define scratch registers that can be used safely
by the code called from the patchpoint intrinsic. At least one scratch register
is required, because that one is used for the materialization of the jump
address. This patch depends on D2009.
Differential Revision: http://llvm-reviews.chandlerc.com/D2074
Reviewed by Andy
llvm-svn: 194306
The idea of the AnyReg Calling Convention is to provide the call arguments in
registers, but not to force them to be placed in a paticular order into a
specified set of registers. Instead it is up tp the register allocator to assign
any register as it sees fit. The same applies to the return value (if
applicable).
Differential Revision: http://llvm-reviews.chandlerc.com/D2009
Reviewed by Andy
llvm-svn: 194293
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask type for the given target. This mask has
usually the same size as the VSELECT return type (except for Intel KNL). Now the
type legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
Reviewed by Nadav
llvm-svn: 193676
This optimization is not SSE specific so I am moving it to DAGco.
The new scalar_to_vector dag node exposed a missing pattern in the AArch64 target that I needed to add.
llvm-svn: 193393
Calling _chkstk is required on ELF as well as COFF on Windows. Without
_chkstk, functions requiring large stack crash in initialization code.
Previous code tested for COFF format but not Mach-O and this patch modifies
the code to test for Windows OS (both Windows target and MingW target)
but not Mach-O object format: Looks like macho environment was used to
build some EFI code.
Credits to Andrew MacPherson.
llvm-svn: 193289
Without _chkstk functions requiring large stack crash in
initialization code. Previous code tested for COFF format but
not Mach-O and this patch modifies the code to test for Windows.
Credits to Andrew MacPherson.
llvm-svn: 193263
On sandy bridge (PR17654) we now get
vpxor %xmm1, %xmm1, %xmm1
vpunpckhbw %xmm1, %xmm0, %xmm2
vpunpcklbw %xmm1, %xmm0, %xmm0
vinsertf128 $1, %xmm2, %ymm0, %ymm0
On haswell it's a simple
vpmovzxbw %xmm0, %ymm0
There is a maze of duplicated and dead transforms and patterns in this
area. Remove the dead custom lowering of zext v8i16 to v8i32, that's
already handled by LowerAVXExtend.
llvm-svn: 193262
the instruction defenitions and ISEL reflect this.
Prior to this patch these instructions took an i32i8imm, and the high bits were
dropped during encoding. This led to incorrect behavior for shifts by
immediates higher than 255. This patch fixes that issue by detecting large
immediate shifts and returning constant zero (for logical shifts) or capping
the shift amount at an encodable value (for arithmetic shifts).
Fixes <rdar://problem/14968098>
llvm-svn: 193096
Consider the following:
typedef unsigned short ushort4U __attribute__((ext_vector_type(4),
aligned(2)));
typedef unsigned short ushort4 __attribute__((ext_vector_type(4)));
typedef unsigned short ushort8 __attribute__((ext_vector_type(8)));
typedef int int4 __attribute__((ext_vector_type(4)));
int4 __bbase_cvt_int(ushort4 v) {
ushort8 a;
a.lo = v;
return _mm_cvtepu16_epi32(a);
}
This generates the, not unreasonable, IR:
define <4 x i32> @foo0(double %v.coerce) nounwind ssp {
%tmp = bitcast double %v.coerce to <4 x i16>
%tmp1 = shufflevector <4 x i16> %tmp, <4 x i16> undef, <8 x i32> <i32
%0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
%tmp2 = tail call <4 x i32> @llvm.x86.sse41.pmovzxwd(<8 x i16> %tmp1)
ret <4 x i32> %tmp2
}
The problem is when type legalization gets hold of the v4i16. It
legalizes that by spilling to the stack, then doing a zero-extending
load. Things go even more silly from there, ending up with something
like:
_foo0:
movsd %xmm0, -8(%rsp) <== Spill to the stack.
movq -8(%rsp), %xmm0 <== Reload it right back out.
pmovzxwd %xmm0, %xmm1 <== Here's what we actually asked for.
pblendw $1, %xmm1, %xmm0 <== We don't need this at all
pmovzxwd %xmm0, %xmm0 <== We already did this
ret
The v8i8 to v8i16 zext intrinsic gives even worse results, with two
table lookups via pshufb instructions(!!).
To avoid all that, we can move the bitcasting until after we've formed
the wider (legal) vector type. Then our normal codegen flows along
nicely and we get the expected:
_foo0:
pmovzxwd %xmm0, %xmm0
ret
rdar://15245794
llvm-svn: 192866
- Type of index used in extract_vector_elt or insert_vector_elt supposes
to be TLI.getVectorIdxTy() which is pointer type on most targets. It'd
better to truncate (or zero-extend in case it's changed later) it to
mask element type to guarantee they are matching instead of asserting
that.
llvm-svn: 192722
- Lower signed division by constant powers-of-2 to target-independent
DAG operators instead of target-dependent ones to support them better
on targets where vector types are legal but shift operators on that
types are illegal. E.g., on AVX, PSRAW is only available on <8 x i16>
though <16 x i16> is a legal type.
llvm-svn: 192721
In AVX 256bit vectors are valid vectors and therefore the Type Legalizer doesn't
split the VSELECT and SETCC nodes. AVX only supports MIN/MAX on 128bit vectors
and this fix enables vector splitting for this special case in the X86 DAG
Combiner.
This fix is related to PR16695, PR17002, and <rdar://problem/14594431>.
llvm-svn: 191131
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask for the given target. This mask has usually
te same size as the VSELECT return type (except for Intel KNL). Now the type
legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
llvm-svn: 191130
If the DAG already has only legal types, then the second round of DAG combines
is skipped. In this case VSELECT+SETCC patterns that match a more efficient
instruction (e.g. min/max) are never recognized.
This fix allows VSELECT+SETCC combines if the types are already legal before DAG
type legalization.
Reviewer: Nadav
llvm-svn: 190105
This change came about primarily because of two issues in the existing code.
Niether of:
define i64 @test1(i64 %val) {
%in = trunc i64 %val to i32
tail call i32 @ret32(i32 returned %in)
ret i64 %val
}
define i64 @test2(i64 %val) {
tail call i32 @ret32(i32 returned undef)
ret i32 42
}
should be tail calls, and the function sameNoopInput is responsible. The main
problem is that it is completely symmetric in the "tail call" and "ret" value,
but in reality different things are allowed on each side.
For these cases:
1. Any truncation should lead to a larger value being generated by "tail call"
than needed by "ret".
2. Undef should only be allowed as a source for ret, not as a result of the
call.
Along the way I noticed that a mismatch between what this function treats as a
valid truncation and what the backends see can lead to invalid calls as well
(see x86-32 test case).
This patch refactors the code so that instead of being based primarily on
values which it recurses into when necessary, it starts by inspecting the type
and considers each fundamental slot that the backend will see in turn. For
example, given a pathological function that returned {{}, {{}, i32, {}}, i32}
we would consider each "real" i32 in turn, and ask if it passes through
unchanged. This is much closer to what the backend sees as a result of
ComputeValueVTs.
Aside from the bug fixes, this eliminates the recursion that's going on and, I
believe, makes the bulk of the code significantly easier to understand. The
trade-off is the nasty iterators needed to find the real types inside a
returned value.
llvm-svn: 187787
Due to the weird and wondeful usual arithmetic conversions, some
calculations involving negative values were getting performed in
uint32_t and then promoted to int64_t, which is really not a good
idea.
Patch by Katsuhiro Ueno.
llvm-svn: 187703
All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms.
Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors.
Added lowering for EXTRACT/INSERT subvector for 512-bit vectors.
Added a test.
llvm-svn: 187491
CustomLowerNode was not being called during SplitVectorOperand,
meaning custom legalization could not be used by targets.
This also adds a test case for NVPTX that depends on this custom
legalization.
Differential Revision: http://llvm-reviews.chandlerc.com/D1195
Attempt to fix the buildbots by making the X86 test I just added platform independent
llvm-svn: 187202
This reverts commit 187198. It broke the bots.
The soft float test probably needs a -triple because of name differences.
On the hard float test I am getting a "roundss $1, %xmm0, %xmm0", instead of
"vroundss $1, %xmm0, %xmm0, %xmm0".
llvm-svn: 187201
CustomLowerNode was not being called during SplitVectorOperand,
meaning custom legalization could not be used by targets.
This also adds a test case for NVPTX that depends on this custom
legalization.
Differential Revision: http://llvm-reviews.chandlerc.com/D1195
llvm-svn: 187198
Use PMIN/PMAX for UGE/ULE vector comparions to reduce the number of required
instructions. This trick also works for UGT/ULT, but there is no advantage in
doing so. It wouldn't reduce the number of instructions and it would actually
reduce performance.
Reviewer: Ben
radar:5972691
llvm-svn: 186432
Summary:
This patch adds explicit calling convention types for the Win64 and
System V/x86-64 ABIs. This allows code to override the default, and use
the Win64 convention on a target that wants to use SysV (and
vice-versa). This is needed to implement the `ms_abi` and `sysv_abi` GNU
attributes.
Reviewers:
CC:
llvm-svn: 186144
in-tree implementations of TargetLoweringBase::isFMAFasterThanMulAndAdd in
order to resolve the following issues with fmuladd (i.e. optional FMA)
intrinsics:
1. On X86(-64) targets, ISD::FMA nodes are formed when lowering fmuladd
intrinsics even if the subtarget does not support FMA instructions, leading
to laughably bad code generation in some situations.
2. On AArch64 targets, ISD::FMA nodes are formed for operations on fp128,
resulting in a call to a software fp128 FMA implementation.
3. On PowerPC targets, FMAs are not generated from fmuladd intrinsics on types
like v2f32, v8f32, v4f64, etc., even though they promote, split, scalarize,
etc. to types that support hardware FMAs.
The function has also been slightly renamed for consistency and to force a
merge/build conflict for any out-of-tree target implementing it. To resolve,
see comments and fixed in-tree examples.
llvm-svn: 185956
Chandler Carruth