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
Based on discussions with Lang Hames and Jakob Stoklund Olesen at the hacker's lab, and in the light of upcoming work on the PBQP register allocator, it was though that CalcSpillWeights does not need to be a pass. This change will enable to customize / tune the spill weight computation depending on the allocator.
Update the documentation style while there.
No functionnal change.
llvm-svn: 194356
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 new graph structure replaces the node and edge linked lists with vectors.
Free lists (well, free vectors) are used for fast insertion/deletion.
The ultimate aim is to make PBQP graphs cheap to clone. The motivation is that
the PBQP solver destructively consumes input graphs while computing a solution,
forcing the graph to be fully reconstructed for each round of PBQP. This
imposes a high cost on large functions, which often require several rounds of
solving/spilling to find a final register allocation. If we can cheaply clone
the PBQP graph and incrementally update it between rounds then hopefully we can
reduce this cost. Further, once we begin pooling matrix/vector values (future
work), we can cache some PBQP solver metadata and share it between cloned
graphs, allowing the PBQP solver to re-use some of the computation done in
earlier rounds.
For now this is just a data structure update. The allocator and solver still
use the graph the same way as before, fully reconstructing it between each
round. I expect no material change from this update, although it may change
the iteration order of the nodes, causing ties in the solver to break in
different directions, and this could perturb the generated allocations
(hopefully in a completely benign way).
Thanks very much to Arnaud Allard de Grandmaison for encouraging me to get back
to work on this, and for a lot of discussion and many useful PBQP test cases.
llvm-svn: 194300
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
Based on discussions with Lang Hames and Jakob Stoklund Olesen at the hacker's lab, and in the light of upcoming work on the PBQP register allocator, it was though that CalcSpillWeights does not need to be a pass. This change will enable to customize / tune the spill weight computation depending on the allocator.
Update the documentation style while there.
No functionnal change.
llvm-svn: 194269
MorphNodeTo is not safe to call during DAG building. It eagerly
deletes dependent DAG nodes which invalidates the NodeMap. We could
expose a safe interface for morphing nodes, but I don't think it's
worth it. Just create a new MachineNode and replaceAllUsesWith.
My understaning of the SD design has been that we want to support
early target opcode selection. That isn't very well supported, but
generally works. It seems reasonable to rely on this feature even if
it isn't widely used.
llvm-svn: 194102
linkonce_odr_auto_hide was in incomplete attempt to implement a way
for the linker to hide symbols that are known to be available in every
TU and whose addresses are not relevant for a particular DSO.
It was redundant in that it all its uses are equivalent to
linkonce_odr+unnamed_addr. Unlike those, it has never been connected
to clang or llvm's optimizers, so it was effectively dead.
Given that nothing produces it, this patch just nukes it
(other than the llvm-c enum value).
llvm-svn: 193865
We add a map in DwarfDebug to map MDNodes that are shareable across CUs to the
corresponding DIEs: MDTypeNodeToDieMap. These DIEs can be shared across CUs,
that is why we keep the maps in DwarfDebug instead of CompileUnit.
We make the assumption that if a DIE is not added to an owner yet, we assume
it belongs to the current CU. Since DIEs for the type system are added to
their owners immediately after creation, and other DIEs belong to the current
CU, the assumption should be true.
A testing case is added to show that we only create a single DIE for a type
MDNode and we use ref_addr to refer to the type DIE.
We also add a testing case to show ref_addr relocations for non-darwin
platforms.
llvm-svn: 193779
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
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
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
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
