When an extend more than doubles the size of the elements (e.g., a zext
from v16i8 to v16i32), the normal legalization method of splitting the
vectors will run into problems as by the time the destination vector is
legal, the source vector is illegal. The end result is the operation
often becoming scalarized, with the typical horrible performance. For
example, on x86_64, the simple input of:
define void @bar(<16 x i8> %a, <16 x i32>* %p) nounwind {
%tmp = zext <16 x i8> %a to <16 x i32>
store <16 x i32> %tmp, <16 x i32>*%p
ret void
}
Generates:
.section __TEXT,__text,regular,pure_instructions
.section __TEXT,__const
.align 5
LCPI0_0:
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.long 255 ## 0xff
.section __TEXT,__text,regular,pure_instructions
.globl _bar
.align 4, 0x90
_bar:
vpunpckhbw %xmm0, %xmm0, %xmm1
vpunpckhwd %xmm0, %xmm1, %xmm2
vpmovzxwd %xmm1, %xmm1
vinsertf128 $1, %xmm2, %ymm1, %ymm1
vmovaps LCPI0_0(%rip), %ymm2
vandps %ymm2, %ymm1, %ymm1
vpmovzxbw %xmm0, %xmm3
vpunpckhwd %xmm0, %xmm3, %xmm3
vpmovzxbd %xmm0, %xmm0
vinsertf128 $1, %xmm3, %ymm0, %ymm0
vandps %ymm2, %ymm0, %ymm0
vmovaps %ymm0, (%rdi)
vmovaps %ymm1, 32(%rdi)
vzeroupper
ret
So instead we can check if there are legal types that enable us to split
more cleverly when the input vector is already legal such that we don't
turn it into an illegal type. If the extend is such that it's more than
doubling the size of the input we check if
- the number of vector elements is even,
- the source type is legal,
- the type of a split source is illegal,
- the type of an extended (by doubling element size) source is legal, and
- the type of that extended source when split is legal.
If the conditions are met, instead of just splitting both the
destination and the source types, we create an extend that only goes up
one "step" (doubling the element width), and the continue legalizing the
rest of the operation normally. The result is that this operates as a
new, more effecient, termination condition for the loop of "split the
operation until the destination type is legal."
With this change, the above example now compiles to:
_bar:
vpxor %xmm1, %xmm1, %xmm1
vpunpcklbw %xmm1, %xmm0, %xmm2
vpunpckhwd %xmm1, %xmm2, %xmm3
vpunpcklwd %xmm1, %xmm2, %xmm2
vinsertf128 $1, %xmm3, %ymm2, %ymm2
vpunpckhbw %xmm1, %xmm0, %xmm0
vpunpckhwd %xmm1, %xmm0, %xmm3
vpunpcklwd %xmm1, %xmm0, %xmm0
vinsertf128 $1, %xmm3, %ymm0, %ymm0
vmovaps %ymm0, 32(%rdi)
vmovaps %ymm2, (%rdi)
vzeroupper
ret
This generalizes a custom lowering that was added a while back to the
ARM backend. That lowering is no longer necessary, and is removed. The
testcases for it, however, provide excellent ARM tests for this change
and so remain.
rdar://14735100
llvm-svn: 193727
The function verifyFunction() in lib/IR/Verifier.cpp misses some
calls. It creates a temporary FunctionPassManager that will run a
single Verifier pass. Unfortunately, FunctionPassManager is no
PassManager and does not call doInitialization() and doFinalization()
by itself. Verifier does important tasks in doInitialization() such as
collecting type information used to check DebugInfo metadata and
doFinalization() does some additional checks. Therefore these checks
were missed and debug info couldn't be verified at all, it just
crashed if the function had some.
verifyFunction() is currently not used in llvm unless -debug option is
enabled, and in unittests/IR/VerifierTest.cpp
VerifierTest had to be changed to create the function in a module from
which the type debug info can be collected.
Patch by Michael Kruse.
llvm-svn: 193719
With this patch llvm produces a weak_def_can_be_hidden for linkonce_odr
if they are also unnamed_addr or don't have their address taken.
There is not a lot of documentation about .weak_def_can_be_hidden, but
from the old discussion about linkonce_odr_auto_hide and the name of
the directive this looks correct: these symbols can be hidden.
Testing this with the ld64 in Xcode 5 linking clang reduces the number of
exported symbols from 21053 to 19049.
llvm-svn: 193718
startswith_lower is ocassionally useful and I think worth adding.
endwith_lower is added for completeness.
Differential Revision: http://llvm-reviews.chandlerc.com/D2041
llvm-svn: 193706
Currently, instead of showing up as link, it is rendered as
...of FunctionPass <writing-an-llvm-pass-FunctionPass>. The...
PR17733. Patch by Tay Ray Chuan!
llvm-svn: 193698
Fixing this Windows build error:
..\lib\Target\Mips\MipsSEISelLowering.cpp(997) : error C2027: use of undefined type 'llvm::raw_ostream'
llvm-svn: 193696
Also corrected the definition of the intrinsics for these instructions (the
result register is also the first operand), and added intrinsics for bsel and
bseli to clang (they already existed in the backend).
These four operations are mostly equivalent to bsel, and bseli (the difference
is which operand is tied to the result). As a result some of the tests changed
as described below.
bitwise.ll:
- bsel.v test adapted so that the mask is unknown at compile-time. This stops
it emitting bmnzi.b instead of the intended bsel.v.
- The bseli.b test now tests the right thing. Namely the case when one of the
values is an uimm8, rather than when the condition is a uimm8 (which is
covered by bmnzi.b)
compare.ll:
- bsel.v tests now (correctly) emits bmnz.v instead of bsel.v because this
is the same operation (see MSA.txt).
i8.ll
- CHECK-DAG-ized test.
- bmzi.b test now (correctly) emits equivalent bmnzi.b with swapped operands
because this is the same operation (see MSA.txt).
- bseli.b still emits bseli.b though because the immediate makes it
distinguishable from bmnzi.b.
vec.ll:
- CHECK-DAG-ized test.
- bmz.v tests now (correctly) emits bmnz.v with swapped operands (see
MSA.txt).
- bsel.v tests now (correctly) emits bmnz.v with swapped operands (see
MSA.txt).
llvm-svn: 193693
This problem was found and fixed by José Fonseca in March 2011 for
SmallPtrSet, committed r128566. But as far as I can tell, all other
llvm hash tables retain the same problem: the bucket count can grow
without bound while size() remains near constant by repeated
insert/erase cycles that tend to fill the container with tombstones.
Here is a demo that has been reduced to a trivial case:
int
main()
{
llvm::DenseSet<unsigned> d;
for (unsigned i = 0; i < 0xFFFFFFF; ++i)
{
d.insert(i);
d.erase(i);
}
}
While the container size() never grows above 1, the bucket count grows
like this:
nb = 64
nb = 128
nb = 256
nb = 512
nb = 1024
nb = 2048
nb = 4096
nb = 8192
nb = 16384
nb = 32768
nb = 65536
nb = 131072
nb = 262144
nb = 524288
nb = 1048576
nb = 2097152
nb = 4194304
nb = 8388608
nb = 16777216
nb = 33554432
nb = 67108864
nb = 134217728
nb = 268435456
The above program currently consumes a few GB ram. This patch brings
the memory consumption down by several orders of magnitude, and keeps
the bucket count at 64 for the above test.
llvm-svn: 193689
This required correcting the definition of the bins[lr]i intrinsics because
the result is also the first operand.
It also required removing the (arbitrary) check for 32-bit immediates in
MipsSEDAGToDAGISel::selectVSplat().
Currently using binsli.d with 2 bits set in the mask doesn't select binsli.d
because the constant is legalized into a ConstantPool. Similar things can
happen with binsri.d with more than 10 bits set in the mask. The resulting
code when this happens is correct but not optimal.
llvm-svn: 193687
(or (and $a, $mask), (and $b, $inverse_mask)) => (vselect $mask, $a, $b).
where $mask is a constant splat. This allows bitwise operations to make use
of bsel.
It's also a stepping stone towards matching bins[lr], and bins[lr]i from
normal IR.
Two sets of similar tests have been added in this commit. The bsel_* functions
test the case where binsri cannot be used. The binsr_*_i functions will
start to use the binsri instruction in the next commit.
llvm-svn: 193682
splat.d is implemented but this subtest is currently disabled. This is because
it is difficult to match the appropriate IR on MIPS32. There is a patch under
review that should help with this so I hope to enable the subtest soon.
llvm-svn: 193680
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
Use EmitLabelOffsetDifference for handling on darwin platform when
non-darwin platforms use EmitLabelPlusOffset.
Also fix a bug in EmitLabelOffsetDifference where the size is hard-coded
to 4 even though Size is passed in as an argument.
llvm-svn: 193660
To support ref_addr, we calculate the section offset of a DIE (i.e. offset
of a DIE from beginning of the debug info section). The Offset field in DIE
is currently CU-relative. To calculate the section offset, we add a
DebugInfoOffset field in CompileUnit to store the offset of a CU from beginning
of the debug info section. We set the value in DwarfUnits::computeSizeAndOffset
for each CompileUnit.
A helper function DIE::getCompileUnit is added to return the CU DIE that
the input DIE belongs to. We also add a map CUDieMap in DwarfDebug to help
finding the CU for a given CU DIE.
For a cross-referenced DIE, we first find the CU DIE it belongs to with
getCompileUnit, then we use CUDieMap to get the corresponding CU for the CU DIE.
Adding the section offset of the CU with the CU-relative offset of a DIE gives
us the seciton offset of the DIE.
We correctly emit ref_addr with relocation using EmitLabelPlusOffset when
doesDwarfUseRelocationsAcrossSections is true.
This commit handles the emission of DW_FORM_ref_addr when we have an attribute
with FORM_ref_addr. A follow-on patch will start using ref_addr when adding a
DIEEntry. This commit will be tested and verified in the follow-on patch.
Reviewed off-list by Eric, Thanks.
llvm-svn: 193658
after the DIE creation, we construct the context first.
Ensure that we create the context before we create a type so that we can add
the newly created type to the parent. Remove last use of addToContextOwner
now that it's not needed.
We use createAndAddDIE to wrap around "new DIE(". Now all shareable DIEs
should be added to their parents right after the creation.
Reviewed off-list by Eric, Thanks.
llvm-svn: 193657
Helper functions are added:
emitPostLd: emit a post-increment load operation with given size.
emitPostSt: emit a post-increment store operation with given size.
No functionality change.
llvm-svn: 193656
This modifies the pass to classify every SSP-triggering AllocaInst according to
an SSPLayoutKind (LargeArray, SmallArray, AddrOf). This analysis is collected
by the pass and made available for use, but no other pass uses it yet.
The next patch will make use of this analysis in PEI and StackSlot
passes. The end goal is to support ssp-strong stack layout rules.
WIP.
Differential Revision: http://llvm-reviews.chandlerc.com/D1789
llvm-svn: 193653
Use 32-bit types for the array instead of 64. This should
generally be better anyway.
In optimized + assert builds, I saw a failure when a
cond code / type combination that is never set was loading
a non-zero value and hitting the != Promote assert.
It turns out when loading the 64-bit value to do the shift,
the assembly loads the 2 32-bit halves from non-consecutive
addresses. The address the second half of the loaded uint64_t
doesn't include the offset of the array in the struct. Instead
of being offset + 4, it's just + 4.
I'm not entirely sure why this wasn't observed before.
setCondCodeAction isn't heavily used by the in-tree targets,
and not with the higher valued vector SimpleValueTypes. Only
PPC is using one of the > 32 valued types, and that is probably
never used by anyone on a 32-bit MSVC compiled host.
I ran into this when upgrading LLVM versions, so I guess the
value loaded from the nonsense address happened to work out
before.
No test since I'm not really sure if / how it can be reproduced
with the current in tree targets, and it's not supposed to change
anything.
llvm-svn: 193650
Add a tag before the name attribute for readability. Use CHECK-NEXT
instead of CHECK-NOT followed by a CHECK. Add new lines to separate checking
of different DIEs.
llvm-svn: 193629
