The pass works in two modes:
Mode 1: Just set attributes starting from kernels. This can work at
the very beginning of opt and llc pipeline, but cannot clone functions
because it must be a function pass.
Mode 2: Actually clone functions for new attributes. This can only work
after all function passes in the opt pipeline because it has to be a
module pass.
Differential Revision: https://reviews.llvm.org/D63208
llvm-svn: 363586
If a YMM/ZMM non-temporal store has less than natural alignment, split the vector - either they will be satisfactorily aligned or will continue to be split until they are XMMs - at which point the legalizer will scalarize it.
llvm-svn: 363582
When considering a loop containing nontemporal stores or loads for
vectorization, suppress the vectorization if the corresponding
vectorized store or load with the aligment of the original scaler
memory op is not supported with the nontemporal hint on the target.
This adds two new functions:
bool isLegalNTStore(Type *DataType, unsigned Alignment) const;
bool isLegalNTLoad(Type *DataType, unsigned Alignment) const;
to TTI, leaving the target independent default implementation as
returning true, but with overriding implementations for X86 that
check the legality based on available Subtarget features.
This fixes https://llvm.org/PR40759
Differential Revision: https://reviews.llvm.org/D61764
llvm-svn: 363581
I keep using the wrong instruction when manually writing tests. This
really needs to check the number of operands, but I don't see an easy
way to do that right now.
llvm-svn: 363579
This is currently only used for ymm->xmm splitting but we shouldn't hardcode the offsets/alignment.
This is necessary for an upcoming patch to split under-aligned non-temporal vector loads.
llvm-svn: 363570
For loads, pre-SSE41 we can't perform NT loads at all, and after that we can only perform vector aligned loads, so if the alignment is less than for a xmm we'll just end up using the regular unaligned vector loads anyway.
First step towards fixing PR42026 - the next step for stores will be to use SSE4A movntsd where possible and to avoid the stack spill on SSE2 targets.
Differential Revision: https://reviews.llvm.org/D63246
llvm-svn: 363564
The HardwareLoops pass finds exit blocks with a scevable exit count.
If the target specifies to update the loop counter in a register,
through a phi, we need to ensure that the exit block is a latch so
that we can insert the phi with the correct value for the incoming
edge.
Differential Revision: https://reviews.llvm.org/D63336
llvm-svn: 363556
Some GEPs were not being split, presumably because that split would just be
undone by the DAGCombiner. Not performing those splits can prevent important
optimizations, such as preventing the element indices / member offsets from
being (partially) folded into load/store instruction immediates. This patch:
- Makes the splits also occur in the cases where the base address and the GEP
are in the same BB.
- Ensures that the DAGCombiner doesn't reassociate them back again.
Differential Revision: https://reviews.llvm.org/D60294
llvm-svn: 363544
This patch changes MIR stack-id from an integer to an enum,
and adds printing/parsing support for this in MIR files. The default
stack-id '0' is now renamed to 'default'.
This should make MIR tests that have stack objects with different stack-ids
more descriptive. It also clarifies code operating on StackID.
Reviewers: arsenm, thegameg, qcolombet
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D60137
llvm-svn: 363533
Create the ARMBasicBlockUtils class for tracking and querying basic
blocks sizes so we can use them when generating low-overhead loops.
Differential Revision: https://reviews.llvm.org/D63265
llvm-svn: 363530
Summary:
SPE passes doubles the same as soft-float, in register pairs as i32
types. This is all handled by the target-independent layer. However,
this is not optimal when splitting or reforming the doubles, as it
pushes to the stack and loads from, on either side.
For instance, to pass a double argument to a function, assuming the
double value is in r5, the sequence currently looks like this:
evstdd 5, X(1)
lwz 3, X(1)
lwz 4, X+4(1)
Likewise, to form a double into r5 from args in r3 and r4:
stw 3, X(1)
stw 4, X+4(1)
evldd 5, X(1)
This optimizes the fence to use SPE instructions. Now, to pass a double
to a function:
mr 4, 5
evmergehi 3, 5, 5
And to form a double into r5 from args in r3 and r4:
evmergelo 5, 3, 4
This is comparable to the way that gcc generates the double splits.
This also fixes a bug with expanding builtins to libcalls, where the
LowerCallTo() code path was generating intermediate illegal type nodes.
Reviewers: nemanjai, hfinkel, joerg
Subscribers: kbarton, jfb, jsji, llvm-commits
Differential Revision: https://reviews.llvm.org/D54583
llvm-svn: 363526
It would not be safe to unfold the memory form the register form
without checking that we are compiling for 64-bit mode.
This probaby isn't a real functional issue since we are unlikely
to unfold any of these instructions since they don't have any
tied registers, aren't commutable, and don't have any inputs
other than the address.
llvm-svn: 363523
Summary:
Instead of encoding a high-word of 0 using a fake TargetGlobalAddress,
just use a literal target constant. This simplifies some subsequent changes.
The generated assembly is now more explicit about the kind of relocation
that is to be used.
Change-Id: I066835202d23b5941fa7a358eb4b89e9b71ab6f8
Reviewers: arsenm, rampitec
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61491
llvm-svn: 363516
This is similar logic/motivation to the select splitting in D62969.
In D63233, the pattern changes so that we no longer have an extract_subvector of vselect,
but the operands of the select are still being concatenated.
The closest case is represented in either the first or last test diffs here - we have an
extra instruction, but we converted 3-4 ymm instructions into 4-5 xmm instructions.
I think that's the right trade-off for most AVX1 targets.
In the example based on PR37428:
https://bugs.llvm.org/show_bug.cgi?id=37428
...this makes the loop about 30% faster (tested on Haswell by compiling with -mavx).
Differential Revision: https://reviews.llvm.org/D63364
llvm-svn: 363508
Summary:
If the nested loop is an innermost loop, prefer to a 32-byte alignment, so that
we can decrease cache misses and branch-prediction misses. Actual alignment of
the loop will depend on the hotness check and other logic in alignBlocks.
The old code will only align hot loop to 32 bytes when the LoopSize larger than
16 bytes and smaller than 32 bytes, this patch will align the innermost hot loop
to 32 bytes not only for the hot loop whose size is 16~32 bytes.
Reviewed By: steven.zhang, jsji
Differential Revision: https://reviews.llvm.org/D61228
llvm-svn: 363495
Currently you get extra waits, because waits are inserted for the
register dependencies of the call, and the function prolog waits on
everything.
Currently waits are still inserted on returns. It may make sense to
not do this, and wait in the caller instead.
llvm-svn: 363465
The way SelectionDAG treats memory operands is very frustrating, and
by default drops them unless a property is set on the pattern. There
is no pattern for manually selected instructions, so this requires
manually setting them.
llvm-svn: 363455
I recently discovered a bug on the x86 platform: The fp80 type was not handled well by x86 for constrained floating point nodes, as their regular counterparts are replaced by extending loads and truncating stores during the preprocess phase. Normally, platforms don't have this issue, as they don't typically attempt to perform such legalizations during instruction selection preprocessing. Before this change, strict_fp nodes survived until they were mutated to normal nodes, which happened shortly after preprocessing on other platforms. This modification lowers these nodes at the same phase while properly utilizing the chain.5
Submitted by: Drew Wock <drew.wock@sas.com>
Reviewed by: Craig Topper, Kevin P. Neal
Approved by: Craig Topper
Differential Revision: https://reviews.llvm.org/D63271
llvm-svn: 363417
Earlier commit has added AMDGPUOperand::isBoolReg(). Turns out
gcc issues warning about unused function since D63204 is not
yet submitted.
Added NFC part of D63204 to have a use of that function and
mute the warning.
llvm-svn: 363416
Avoid producing illegal register bank copies for reg_sequence and
phi. The default implementation assumes it is possible to pick any
operand's bank and use that for the result, introducing a copy for
operands with a different bank. This does not check for illegal
copies. It is not legal to introduce a VGPR->SGPR copy, so any VGPR
operand requires the result to be a VGPR.
The changes in getInstrMappingImpl aren't strictly necessary, since
AMDGPU now just bypasses this for reg_sequence/phi. This could be
replaced with an assert in case other targets run into this. It is
currently responsible for producing the error for unsatisfiable
copies, but this will be better served with a verifier check.
For phis, for now assume any undetermined operands must be
VGPRs. Eventually, this needs to be able to defer mapping these
operations. This also does not yet have a way to check for whether the
block is in a divergent region.
llvm-svn: 363410
This is the family of vector instructions that combine all the lanes
in their input vector(s), and output a value in one or two GPRs.
Differential Revision: https://reviews.llvm.org/D62670
llvm-svn: 363403
Initial commit of a new pass to create vector predication blocks, called VPT
blocks, that are supported by the Armv8.1-M MVE architecture.
This is a first naive implementation. I.e., for 2 consecutive predicated
instructions I1 and I2, for example, it will generate 2 VPT blocks:
VPST
I1
VPST
I2
A more optimal implementation would obviously put instructions in the same VPT
block when they are predicated on the same condition and when it is allowed to
do this:
VPTT
I1
I2
We will address this optimisation with follow up patches when the groundwork is
in. Creating VPT Blocks is very similar to IT Blocks, which is the reason I
added this to Thumb2ITBlocks.cpp. This allows reuse of the def use analysis
that we need for the more optimal implementation.
VPT blocks cannot be nested in IT blocks, and vice versa, and so these 2 passes
cannot interact with each other. Instructions allowed in VPT blocks must
be MVE instructions that are marked as VPT compatible.
Differential Revision: https://reviews.llvm.org/D63247
llvm-svn: 363370
Merging the two bits shrinks the context table from 16384 bytes to 8192 bytes.
Remove the ATTRIBUTE_BITS macro and just create an enum directly. Then fix the ATTR_max define to be 8192 to reflect the table size so we stop hardcoding it separately.
llvm-svn: 363330
Previously it copied over MachineMemOperands verbatim which caused MOV32rm to have store flags set, and MOV32mr to have load flags set. This fixes some assertions being thrown with EXPENSIVE_CHECKS on.
Committed on behalf of @luke (Luke Lau)
Differential Revision: https://reviews.llvm.org/D62726
llvm-svn: 363268
This commit prepares the way to start adding the main collection of
MVE instructions, which operate on the 128-bit vector registers.
The most obvious thing that's needed, and the simplest, is to add the
MQPR register class, which is like the existing QPR except that it has
fewer registers in it.
The more complicated part: MVE defines a system of vector predication,
in which instructions operating on 128-bit vector registers can be
constrained to operate on only a subset of the lanes, using a system
of prefix instructions similar to the existing Thumb IT, in that you
have one prefix instruction which designates up to 4 following
instructions as subject to predication, and within that sequence, the
predicate can be inverted by means of T/E suffixes ('Then' / 'Else').
To support instructions of this type, we've added two new Tablegen
classes `vpred_n` and `vpred_r` for standard clusters of MC operands
to add to a predicated instruction. Both include a flag indicating how
the instruction is predicated at all (options are T, E and 'not
predicated'), and an input register field for the register controlling
the set of active lanes. They differ from each other in that `vpred_r`
also includes an input operand for the previous value of the output
register, for instructions that leave inactive lanes unchanged.
`vpred_n` lacks that extra operand; it will be used for instructions
that don't preserve inactive lanes in their output register (either
because inactive lanes are zeroed, as the MVE load instructions do, or
because the output register isn't a vector at all).
This commit also adds the family of prefix instructions themselves
(VPT / VPST), and all the machinery needed to work with them in
assembly and disassembly (e.g. generating the 't' and 'e' mnemonic
suffixes on disassembled instructions within a predicated block)
I've added a couple of demo instructions that derive from the new
Tablegen base classes and use those two operand clusters. The bulk of
the vector instructions will come in followup commits small enough to
be manageable. (One exception is that I've added the full version of
`isMnemonicVPTPredicable` in the AsmParser, because it seemed
pointless to carefully split it up.)
Reviewers: dmgreen, samparker, SjoerdMeijer, t.p.northover
Subscribers: javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62669
llvm-svn: 363258
During assembly, the mask operand to an IT instruction (storing the
sequence of T/E for 'Then' and 'Else') is parsed out of the mnemonic
into a representation that encodes 'Then' and 'Else' in the same way
regardless of the condition code. At some point during encoding it has
to be converted into the instruction encoding used in the
architecture, in which the mask encodes a sequence of replacement
low-order bits for the condition code, so that which bit value means
'then' and which 'else' depends on whether the original condition code
had its low bit set.
Previously, that transformation was done by processInstruction(), half
way through assembly. So an MCOperand storing an IT mask would
sometimes store it in one format, and sometimes in the other,
depending on where in the assembly pipeline you were. You can see this
in diagnostics from `llvm-mc -debug -triple=thumbv8a -show-inst`, for
example: if you give it an instruction such as `itete eq`, you'd see
an `<MCOperand Imm:5>` in a diagnostic become `<MCOperand Imm:11>` in
the final output.
Having the same data structure store values with time-dependent
semantics is confusing already, and it will get more confusing when we
introduce the MVE VPT instruction which reuses the Then/Else bitmask
idea in a different context. So I'm refactoring: now, all `ARMOperand`
and `MCOperand` representations of an IT mask work exactly the same
way, namely, 0 means 'Then' and 1 means 'Else', regardless of what
original predicate is being referred to. The architectural encoding of
IT that depends on the original condition is now constructed at the
point when we turn the `MCOperand` into the final instruction bit
pattern, and decoded similarly in the disassembler.
The previous condition-independent parse-time format used 0 for Else
and 1 for Then. I've taken the opportunity to flip the sense of it
while I'm changing all of this anyway, because it seems to me more
natural to use 0 for 'leave the starting condition unchanged' and 1
for 'invert it', as if those bits were an XOR mask.
Reviewers: ostannard
Subscribers: javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63219
llvm-svn: 363244
TTI should report that it's not profitable to generate a hardware loop
if it, or one of its child loops, has already been converted.
Differential Revision: https://reviews.llvm.org/D63212
llvm-svn: 363234
Summary:
- Remove redundant initializations from pass constructors that were
already being initialized by LLVMInitializeX86Target().
- Add initialization function for the FPS pass.
Reviewers: craig.topper
Reviewed By: craig.topper
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63218
llvm-svn: 363221
Add support for s.d instruction for Mips1 which expands into two swc1
instructions.
Patch by Mirko Brkusanin.
Differential Revision: https://reviews.llvm.org/D63199
llvm-svn: 363184
As discussed on D62910, we need to check whether particular types of memory access are allowed, not just their alignment/address-space.
This NFC patch adds a MachineMemOperand::Flags argument to allowsMemoryAccess and allowsMisalignedMemoryAccesses, and wires up calls to pass the relevant flags to them.
If people are happy with this approach I can then update X86TargetLowering::allowsMisalignedMemoryAccesses to handle misaligned NT load/stores.
Differential Revision: https://reviews.llvm.org/D63075
llvm-svn: 363179
Without this fix clang 3.6 complains with:
../lib/Target/ARM/ARMAsmPrinter.cpp:1473:18: error: variable 'BranchTarget' is used uninitialized whenever 'if' condition is false [-Werror,-Wsometimes-uninitialized]
} else if (MI->getOperand(1).isSymbol()) {
^~~~~~~~~~~~~~~~~~~~~~~~~~~~
../lib/Target/ARM/ARMAsmPrinter.cpp:1479:22: note: uninitialized use occurs here
MCInst.addExpr(BranchTarget);
^~~~~~~~~~~~
../lib/Target/ARM/ARMAsmPrinter.cpp:1473:14: note: remove the 'if' if its condition is always true
} else if (MI->getOperand(1).isSymbol()) {
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
../lib/Target/ARM/ARMAsmPrinter.cpp:1465:33: note: initialize the variable 'BranchTarget' to silence this warning
const MCExpr *BranchTarget;
^
= nullptr
1 error generated.
Discussed here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20190610/661417.html
llvm-svn: 363166
Implement the backend target hook to drive the HardwareLoops pass.
The low-overhead branch extension for Arm M-class cores is flexible
enough that we don't have to ensure correctness at this point, except
checking that the loop counter variable can be stored in LR - a
32-bit register. For it to be profitable, we want to avoid loops that
contain function calls, or any other instruction that alters the PC.
This implementation uses TargetLoweringInfo, to query type and
operation actions, looks at intrinsic calls and also performs some
manual checks for remainder/division and FP operations.
I think this should be a good base to start and extra details can be
filled out later.
Differential Revision: https://reviews.llvm.org/D62907
llvm-svn: 363149
Noticed in D63075 - there was a allowsMisalignedMemoryAccesses call to check for unaligned loads and a check for aligned legal type loads - which is exactly what allowsMemoryAccess does.
llvm-svn: 363141
Noticed in D63075 - there was a allowsMisalignedMemoryAccesses call to check for unaligned loads and a check for aligned legal type loads - which is exactly what allowsMemoryAccess does.
llvm-svn: 363137
Extern global merging is good for code-size. There's definitely potential for
performance too, but there's one regression in a benchmark that needs
investigating, so that's why we enable it only when we optimise for size for
now.
Patch by Ramakota Reddy and Sjoerd Meijer.
Differential Revision: https://reviews.llvm.org/D61947
llvm-svn: 363130
The non-masked versions are already in there. I'm having some
trouble coming up with a way to test this right now. Most load
folding should happen during isel so I'm not sure how to get
peephole pass to do it.
llvm-svn: 363125
In order to generate correct debug frame information, it needs to
generate CFI information in prologue and epilog.
Differential Revision: https://reviews.llvm.org/D61773
llvm-svn: 363120
Implement necessary target hooks to enable MachinePipeliner for P9 only.
The pass is off by default, can be enabled with -ppc-enable-pipeliner for P9.
Differential Revision: https://reviews.llvm.org/D62164
llvm-svn: 363085
This patch allows lowering of PIC addresses by using PC-relative
addressing for DSO-local symbols and accessing the address through the
global offset table for non-DSO-local symbols.
Differential Revision: https://reviews.llvm.org/D55303
llvm-svn: 363058
This validates and lowers arguments to inline asm nodes which have the
constraints I, J & K, with the following semantics (equivalent to GCC):
I: Any 12-bit signed immediate.
J: Immediate integer zero only.
K: Any 5-bit unsigned immediate.
Differential Revision: https://reviews.llvm.org/D54093
llvm-svn: 363054
This introduces a new decoding table for MVE instructions, and starts
by adding the family of scalar shift instructions that are part of the
MVE architecture extension: saturating shifts within a single GPR, and
long shifts across a pair of GPRs (both saturating and normal).
Some of these shift instructions have only 3-bit register fields in
the encoding, with the low bit fixed. So they can only address an odd
or even numbered GPR (depending on the operand), and therefore I add
two new register classes, GPREven and GPROdd.
Differential Revision: https://reviews.llvm.org/D62668
Change-Id: Iad95d5f83d26aef70c674027a184a6b1e0098d33
llvm-svn: 363051
As suggested by @arsenm on D63075 - this adds a TargetLowering::allowsMemoryAccess wrapper that takes a Load/Store node's MachineMemOperand to handle the AddressSpace/Alignment arguments and will also implicitly handle the MachineMemOperand::Flags change in D63075.
llvm-svn: 363048
The variable `OffsetMask` is currently only used in an assertion, so
if assertions are compiled out and -Werror is enabled, it becomes a
build failure.
llvm-svn: 363043
This adds support for the new family of conditional selection /
increment / negation instructions; the low-overhead branch
instructions (e.g. BF, WLS, DLS); the CLRM instruction to zero a whole
list of registers at once; the new VMRS/VMSR and VLDR/VSTR
instructions to get data in and out of 8.1-M system registers,
particularly including the new VPR register used by MVE vector
predication.
To support this, we also add a register name 'zr' (used by the CSEL
family to force one of the inputs to the constant 0), and operand
types for lists of registers that are also allowed to include APSR or
VPR (used by CLRM). The VLDR/VSTR instructions also need a new
addressing mode.
The low-overhead branch instructions exist in their own separate
architecture extension, which we treat as enabled by default, but you
can say -mattr=-lob or equivalent to turn it off.
Reviewers: dmgreen, samparker, SjoerdMeijer, t.p.northover
Reviewed By: samparker
Subscribers: miyuki, javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62667
llvm-svn: 363039
This reverts r362990 (git commit 374571301d)
This was causing linker warnings on Darwin:
ld: warning: direct access in function 'llvm::initializeEvexToVexInstPassPass(llvm::PassRegistry&)'
from file '../../lib/libLLVMX86CodeGen.a(X86EvexToVex.cpp.o)' to global weak symbol
'void std::__1::__call_once_proxy<std::__1::tuple<void* (&)(llvm::PassRegistry&),
std::__1::reference_wrapper<llvm::PassRegistry>&&> >(void*)' from file '../../lib/libLLVMCore.a(Verifier.cpp.o)'
means the weak symbol cannot be overridden at runtime. This was likely caused by different translation
units being compiled with different visibility settings.
llvm-svn: 363028
Summary:
For builds with LLVM_BUILD_LLVM_DYLIB=ON and BUILD_SHARED_LIBS=OFF
this change makes all symbols in the target specific libraries hidden
by default.
A new macro called LLVM_EXTERNAL_VISIBILITY has been added to mark symbols in these
libraries public, which is mainly needed for the definitions of the
LLVMInitialize* functions.
This patch reduces the number of public symbols in libLLVM.so by about
25%. This should improve load times for the dynamic library and also
make abi checker tools, like abidiff require less memory when analyzing
libLLVM.so
One side-effect of this change is that for builds with
LLVM_BUILD_LLVM_DYLIB=ON and LLVM_LINK_LLVM_DYLIB=ON some unittests that
access symbols that are no longer public will need to be statically linked.
Before and after public symbol counts (using gcc 8.2.1, ld.bfd 2.31.1):
nm before/libLLVM-9svn.so | grep ' [A-Zuvw] ' | wc -l
36221
nm after/libLLVM-9svn.so | grep ' [A-Zuvw] ' | wc -l
26278
Reviewers: chandlerc, beanz, mgorny, rnk, hans
Reviewed By: rnk, hans
Subscribers: Jim, hiraditya, michaelplatings, chapuni, jholewinski, arsenm, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, sbc100, jgravelle-google, aheejin, kbarton, fedor.sergeev, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, zzheng, edward-jones, mgrang, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, PkmX, jocewei, kristina, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D54439
llvm-svn: 362990
This was found during HTM cleanup.
Adding a test for builtin_ttest would expose following issue.
*** Bad machine code: Illegal physical register for instruction ***
- function: test10
- basic block: %bb.0 entry (0xf0e57497b58)
- instruction: %5:crrc0 = TABORTWCI 0, $zero, 0
- operand 2: $zero
$zero is not a GPRC register.
LLVM ERROR: Found 1 machine code errors.
Differential Revision: https://reviews.llvm.org/D63079
llvm-svn: 362974
These caused a build failure because I managed not to notice they
depended on a later unpushed commit in my current stack. Sorry about
that.
llvm-svn: 362956
This should have been part of r362953, but I had a finger-trouble
incident and committed the old rather than new version of the patch.
Sorry.
llvm-svn: 362955
This adds support for the new family of conditional selection /
increment / negation instructions; the low-overhead branch
instructions (e.g. BF, WLS, DLS); the CLRM instruction to zero a whole
list of registers at once; the new VMRS/VMSR and VLDR/VSTR
instructions to get data in and out of 8.1-M system registers,
particularly including the new VPR register used by MVE vector
predication.
To support this, we also add a register name 'zr' (used by the CSEL
family to force one of the inputs to the constant 0), and operand
types for lists of registers that are also allowed to include APSR or
VPR (used by CLRM). The VLDR/VSTR instructions also need some new
addressing modes.
The low-overhead branch instructions exist in their own separate
architecture extension, which we treat as enabled by default, but you
can say -mattr=-lob or equivalent to turn it off.
Reviewers: dmgreen, samparker, SjoerdMeijer, t.p.northover
Reviewed By: samparker
Subscribers: miyuki, javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62667
llvm-svn: 362953
Arm v8.1-M supports the VMOV instructions that move a half-precision
value to and from a GPR, but not if the GPR is SP or PC.
To fix this, I've changed those instructions to use the rGPR register
class instead of GPR. rGPR always excludes PC, and it excludes SP
except in the presence of the HasV8Ops target feature (i.e. Arm v8-A).
So the effect is that VMOV.F16 to and from PC is now illegal
everywhere, but VMOV.F16 to and from SP is illegal only on non-v8-A
cores (which I believe is all as it should be).
Reviewers: dmgreen, samparker, SjoerdMeijer, ostannard
Reviewed By: ostannard
Subscribers: ostannard, javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60704
llvm-svn: 362942
This option allows loops with small max trip counts to be fully unrolled. This
can help with code like the remainder loops from manually unrolled loops like
those that appear in the cmsis dsp library. We would apparently previously
runtime unroll them with the default unroll count (4).
Differential Revision: https://reviews.llvm.org/D63064
llvm-svn: 362928
Summary:
Our default behavior is to use sign_extend for signed comparisons and zero_extend for everything else. But for equality we have the freedom to use either extension. If we can prove the input has been truncated from something with enough sign bits, we can use sign_extend instead and let DAG combine optimize it out. A similar rule is used by type legalization in LegalizeIntegerTypes.
This gets rid of the movzx in PR42189. The immediate will still take 4 bytes instead of the 2 bytes plus 0x66 prefix a cmp di, 32767 would get, but it avoids a length changing prefix.
Reviewers: RKSimon, spatel, xbolva00
Reviewed By: xbolva00
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63032
llvm-svn: 362920
Summary:
We can only use the memory form of cvtss2sd under optsize due to a partial register update. So previously we were emitting 2 instructions for extload when optimizing for speed. Also due to a late optimization in preprocessiseldag we had to handle (fpextend (loadf32)) under optsize.
This patch forces extload to expand so that it will always be in the (fpextend (loadf32)) form during isel. And when optimizing for speed we can just let each of those pieces select an instruction independently.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62710
llvm-svn: 362919
Previously we did the equivalent operation in isel patterns with
COPY_TO_REGCLASS operations to transition. By inserting
scalar_to_vetors and extract_vector_elts before isel we can
allow each piece to be selected individually and accomplish the
same final result.
I ideally we'd use vector operations earlier in lowering/combine,
but that looks to be more difficult.
The scalar-fp-to-i64.ll changes are because we have a pattern for
using movlpd for store+extract_vector_elt. While an f64 store
uses movsd. The encoding sizes are the same.
llvm-svn: 362914
Types such as float and i64's do not have legal loads in Thumb1, but will still
be loaded with a LDR (or potentially multiple LDR's). As such we can treat the
cost of addressing mode calculations the same as an i32 and get some optimisation
benefits.
Differential Revision: https://reviews.llvm.org/D62968
llvm-svn: 362874
Now with MVE being added, we can add the vector addressing mode costs for it.
These are generally imm7 multiplied by the size of the type being loaded /
stored.
Differential Revision: https://reviews.llvm.org/D62967
llvm-svn: 362873
The fp16 version of VLDR takes a imm8 multiplied by 2. This updates the costs
to account for those, and adds extra testing. It is dependant upon hasFPRegs16
as this is what the load/store instructions require.
Differential Revision: https://reviews.llvm.org/D62966
llvm-svn: 362872
We are starting to add an entirely separate vector architecture to the ARM
backend. To do that we need at least some separation between the existing NEON
and the new MVE code. This patch just goes through the Neon patterns and
ensures that they are predicated on HasNEON, giving MVE a stable place to start
from.
No tests yet as this is largely an NFC, and we don't have the other target that
will treat any of these intructions as legal.
Differential Revision: https://reviews.llvm.org/D62945
llvm-svn: 362870
This patch aims to reduce spilling and register moves by using the 3-address
versions of instructions per default instead of the 2-address equivalent
ones. It seems that both spilling and register moves are improved noticeably
generally.
Regalloc hints are passed to increase conversions to 2-address instructions
which are done in SystemZShortenInst.cpp (after regalloc).
Since the SystemZ reg/mem instructions are 2-address (dst and lhs regs are
the same), foldMemoryOperandImpl() can no longer trivially fold a spilled
source register since the reg/reg instruction is now 3-address. In order to
remedy this, new 3-address pseudo memory instructions are used to perform the
folding only when the dst and lhs virtual registers are known to be allocated
to the same physreg. In order to not let MachineCopyPropagation run and
change registers on these transformed instructions (making it 3-address), a
new target pass called SystemZPostRewrite.cpp is run just after
VirtRegRewriter, that immediately lowers the pseudo to a target instruction.
If it would have been possibe to insert a COPY instruction and change a
register operand (convert to 2-address) in foldMemoryOperandImpl() while
trusting that the caller (e.g. InlineSpiller) would update/repair the
involved LiveIntervals, the solution involving pseudo instructions would not
have been needed. This is perhaps a potential improvement (see Phabricator
post).
Common code changes:
* A new hook TargetPassConfig::addPostRewrite() is utilized to be able to run a
target pass immediately before MachineCopyPropagation.
* VirtRegMap is passed as an argument to foldMemoryOperand().
Review: Ulrich Weigand, Quentin Colombet
https://reviews.llvm.org/D60888
llvm-svn: 362868
This is a potentially large perf win for AVX1 targets because of the way we
auto-vectorize to 256-bit but then expect the backend to legalize/optimize
for the half-implemented AVX1 ISA.
On the motivating example from PR37428 (even though this patch doesn't solve
the vector shift issue):
https://bugs.llvm.org/show_bug.cgi?id=37428
...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell)
because we eliminate the remaining 256-bit vblendv ops.
I added comments on a couple of tests that require further work. If we have
256-bit logic ops separating the vselect and extract, we should probably narrow
everything to 128-bit, but that requires a larger pattern match.
Differential Revision: https://reviews.llvm.org/D62969
llvm-svn: 362797
Summary:
This allows some integer bitwise operations to instead be performed by
hardware fp instructions. This is correct because the RISC-V spec
requires the F and D extensions to use the IEEE-754 standard
representation, and fp register loads and stores to be bit-preserving.
This is tested against the soft-float ABI, but with hardware float
extensions enabled, so that the tests also ensure the optimisation also
fires in this case.
Reviewers: asb, luismarques
Reviewed By: asb
Subscribers: hiraditya, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, kito-cheng, shiva0217, jrtc27, zzheng, edward-jones, rogfer01, MartinMosbeck, brucehoult, the_o, rkruppe, PkmX, jocewei, psnobl, benna, Jim, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62900
llvm-svn: 362790
Summary:
This patch fixes a bug in the assembler that permitted a type suffix on
predicate registers when not expected. For instance, the following was
previously valid:
faddv h0, p0.q, z1.h
This bug was present in all SVE instructions containing predicates with
no type suffix and no predication form qualifier, i.e. /z or /m. The
latter instructions are already caught with an appropiate error message
by the assembler, e.g.:
.text
<stdin>:1:13: error: not expecting size suffix
cmpne p1.s, p0.b/z, z2.s, 0
^
A similar issue for SVE vector registers was fixed in:
https://reviews.llvm.org/D59636
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D62942
llvm-svn: 362780
Patch which introduces a target-independent framework for generating
hardware loops at the IR level. Most of the code has been taken from
PowerPC CTRLoops and PowerPC has been ported over to use this generic
pass. The target dependent parts have been moved into
TargetTransformInfo, via isHardwareLoopProfitable, with
HardwareLoopInfo introduced to transfer information from the backend.
Three generic intrinsics have been introduced:
- void @llvm.set_loop_iterations
Takes as a single operand, the number of iterations to be executed.
- i1 @llvm.loop_decrement(anyint)
Takes the maximum number of elements processed in an iteration of
the loop body and subtracts this from the total count. Returns
false when the loop should exit.
- anyint @llvm.loop_decrement_reg(anyint, anyint)
Takes the number of elements remaining to be processed as well as
the maximum numbe of elements processed in an iteration of the loop
body. Returns the updated number of elements remaining.
llvm-svn: 362774
In r356860, the legalization logic for BSWAP was modified to ISD::ROTL,
rather than the old ISD::{SHL, SRL, OR} nodes.
This works fine on AVR for 8-bit rotations, but 16-bit rotations are
currently unimplemented - they always trigger an assertion error in the
AVRExpandPseudoInsts pass ("RORW unimplemented").
This patch instructions the legalizer to expand 16-bit rotations into
the previous SHL, SRL, OR pattern it did previously.
This fixes the 'issue-cannot-select-bswap.ll' test. Interestingly, this
test failure seems flaky - it passes successfully on the avr-build-01
buildbot, but fails locally on my Arch Linux install.
llvm-svn: 362773
Use the PPC vector min/max instructions for computing the corresponding
operation as these should be faster than the compare/select sequences
we currently emit.
Differential revision: https://reviews.llvm.org/D47332
llvm-svn: 362759
SIInsertSkips really doesn't understand the control flow, and makes
very stupid assumptions about the block layout. This was able to get
away with not skipping return blocks, since usually after
structurization there is only one placed at the end of the
function. Tail duplication can break this assumption.
llvm-svn: 362754
"Divergence driven ISel. Assign register class for cross block values
according to the divergence."
that discovered the design flaw leading to several issues that
required to be solved before.
This change reverts AMDGPU specific changes and keeps common part
unaffected.
llvm-svn: 362749
This primarily affects add/fadd/mul/fmul/and/or/xor/pmuludq/pmuldq/max/min/fmaxc/fminc/pmaddwd/pavg.
We already commuted the unmasked and zero masked versions.
I've added 512-bit stack folding tests for most of the instructions
affected. I've tested needing commuting and not commuting across
unmasked, merged masked, and zero masked. The 128/256 bit instructions
should behave similarly.
llvm-svn: 362746
Summary:
(1) Function descriptor on AIX
On AIX, a called routine may have 2 distinct symbols associated with it:
* A function descriptor (Name)
* A function entry point (.Name)
The descriptor structure on AIX is the same as those in the ELF V1 ABI:
* The address of the entry point of the function.
* The TOC base address for the function.
* The environment pointer.
The descriptor symbol uses the same name as the source level function in C.
The function entry point is analogous to the symbol we would generate for a
function in a non-descriptor-based ABI, except that it is renamed by
prepending a ".".
Which symbol gets referenced depends on the context:
* Taking the address of the function references the descriptor symbol.
* Calling the function references the entry point symbol.
(2) Speaking of implementation on AIX, for direct function call target, we
create proper MCSymbol SDNode(e.g . ".foo") while constructing SDAG to
replace original TargetGlobalAddress SDNode. Then down the path, we can
take advantage of this MCSymbol.
Patch by: Xiangling_L
Reviewed by: sfertile, hubert.reinterpretcast, jasonliu, syzaara
Differential Revision: https://reviews.llvm.org/D62532
llvm-svn: 362735
Summary:
This patch implements SDAG call lowering on AIX for functions
which only have parameters that could fit into GPRs.
Reviewers: hubert.reinterpretcast, syzaara
Differential Revision: https://reviews.llvm.org/D62823
llvm-svn: 362708
This patch is a follow up for D62018 to add lrint/llrint
support for float16.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D62863
llvm-svn: 362700
This patch is a follow up for D61391 to add lround/llround
support for float16.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D62861
llvm-svn: 362698
We already get support for G_ZEXTLOAD to s32 from the importer, but it can't
deal with the SUBREG_TO_REG in the pattern. Tweaking the existing manual
selection code for G_LOAD to handle an additional SUBREG_TO_REG when dealing
with G_ZEXTLOAD isn't much work.
Also add tests to check the imported pattern selections to s32 work.
llvm-svn: 362681
This is intended to enable the use of an immediate blend or
more optimal instruction. But if the passthru is zero we don't
need any additional instructions.
llvm-svn: 362675
When looking through copies, make sure to not try to find the vreg def of a physreg.
Normally getVRegDef will return nullptr in this case, but if there happens to be
multiple defs then it will assert.
This fixes PR42129.
llvm-svn: 362666
This forced the caller to be aware of this, which is an ugly ABI
feature.
Partially reverts r295877. The original reasons for doing this are
mostly fixed. Alloca is now in a non-0 address space, so it should be
OK to have 0 as a valid pointer. Since we treat the absolute address
as the pointer value, this part only really needed to apply to
kernels.
Since r357093, we avoid the need to increment/decrement the offset
register in more cases, and since r354816 the scavenger can fail
without spilling, so it's less critical that we try to avoid an offset
that fits in the MUBUF offset.
Restrict to callable functions for now to split this into 2 steps to
limit thte number of test updates and in case anything breaks.
llvm-svn: 362665
The ISD::STRICT_ nodes used to implement the constrained floating-point
intrinsics are currently never passed to the target back-end, which makes
it impossible to handle them correctly (e.g. mark instructions are depending
on a floating-point status and control register, or mark instructions as
possibly trapping).
This patch allows the target to use setOperationAction to switch the action
on ISD::STRICT_ nodes to Legal. If this is done, the SelectionDAG common code
will stop converting the STRICT nodes to regular floating-point nodes, but
instead pass the STRICT nodes to the target using normal SelectionDAG
matching rules.
To avoid having the back-end duplicate all the floating-point instruction
patterns to handle both strict and non-strict variants, we make the MI
codegen explicitly aware of the floating-point exceptions by introducing
two new concepts:
- A new MCID flag "mayRaiseFPException" that the target should set on any
instruction that possibly can raise FP exception according to the
architecture definition.
- A new MI flag FPExcept that CodeGen/SelectionDAG will set on any MI
instruction resulting from expansion of any constrained FP intrinsic.
Any MI instruction that is *both* marked as mayRaiseFPException *and*
FPExcept then needs to be considered as raising exceptions by MI-level
codegen (e.g. scheduling).
Setting those two new flags is straightforward. The mayRaiseFPException
flag is simply set via TableGen by marking all relevant instruction
patterns in the .td files.
The FPExcept flag is set in SDNodeFlags when creating the STRICT_ nodes
in the SelectionDAG, and gets inherited in the MachineSDNode nodes created
from it during instruction selection. The flag is then transfered to an
MIFlag when creating the MI from the MachineSDNode. This is handled just
like fast-math flags like no-nans are handled today.
This patch includes both common code changes required to implement the
new features, and the SystemZ implementation.
Reviewed By: andrew.w.kaylor
Differential Revision: https://reviews.llvm.org/D55506
llvm-svn: 362663
Since the beginning, the offset of a frame index has been consistently
interpreted backwards. It was treating it as an offset from the
scratch wave offset register as a frame register. The correct
interpretation is the offset from the SP on entry to the function,
before the prolog. Frame index elimination then should select either
SP or another register as an FP.
Treat the scratch wave offset on kernel entry as the pre-incremented
SP. Rely more heavily on the standard hasFP and frame pointer
elimination logic, and clean up the private reservation code. This
saves a copy in most callee functions.
The kernel prolog emission code is still kind of a mess relying on
checking the uses of physical registers, which I would prefer to
eliminate.
Currently selection directly emits MUBUF instructions, which require
using a reference to some register. Use the register chosen for SP,
and then ignore this later. This should probably be cleaned up to use
pseudos that don't refer to any specific base register until frame
index elimination.
Add a workaround for shaders using large numbers of SGPRs. I'm not
sure these cases were ever working correctly, since as far as I can
tell the logic for figuring out which SGPR is the scratch wave offset
doesn't match up with the shader input initialization in the shader
programming guide.
llvm-svn: 362661
One of the sources controls the pass through value for the upper bits
of the result so we can't really commute it.
In practice this problem isn't a functional issue because we would
only try to commute this instruction in order to fold a load. But
we can't do embedded rounding and fold a load at the same time. So
the load fold would never succeed so I don't think we would ever
commute or at least keep the version after commuting.
llvm-svn: 362647
Stanislav Mekhanoshin