This revision introduces a `subtensor` op, which is the counterpart of `subview` for a tensor operand. This also refactors the relevant pieces to allow reusing the `subview` implementation where appropriate.
This operation will be used to implement tiling for Linalg on tensors.
Marks constants of an ICmp instruction as free if it's only user is a select
instruction that is part of a min(max()) pattern. Ensures that in loops, in
particular when loop unrolling is turned on, SSAT will still be correctly generated.
Differential Revision: https://reviews.llvm.org/D88662
Implement vmsge{u}.vx pseudo instruction.
According to RISC-V V specification, there are different scenarios for this
pseudo instruction. I list them below.
unmasked va >= x
pseudoinstruction: vmsge{u}.vx vd, va, x
expansion: vmslt{u}.vx vd, va, x; vmnand.mm vd, vd, vd
masked va >= x, vd != v0
pseudoinstruction: vmsge{u}.vx vd, va, x, v0.t
expansion: vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0
masked va >= x, vd == v0
pseudoinstruction: vmsge{u}.vx vd, va, x, v0.t, vt
expansion: vmslt{u}.vx vt, va, x; vmandnot.mm vd, vd, vt
Use pseudo instruction to model vmsge{u}.vx. The pseudo instruction will convert
to different expansion according to the condition.
Differential Revision: https://reviews.llvm.org/D84732
Intent was a nice idea but it ends up being a bit awkward/heavyweight
without adding much.
In particular, it makes it hard to implement `CodeActionParams.only` properly
(there's an inheritance hierarchy for kinds).
Differential Revision: https://reviews.llvm.org/D88427
Specification for SHT_HASH table says (https://refspecs.linuxbase.org/elf/gabi4+/ch5.dynamic.html#hash)
that it contains Elf32_Word entries for both 32/64 bit objects.
Currently both GNU linkers and LLD sets the `sh_entsize` field to `4`.
At the same time, `yaml2obj` ignores the `EntSize` field for SHT_HASH sections.
This patch fixes this and also adds a support for obj2yaml: it will not
dump this field when the `sh_entsize` contains the default value (`4`).
Differential revision: https://reviews.llvm.org/D88652
v128.const was recently implemented in V8, but until it rolls into Chrome
stable, we can't enable it in the WebAssembly backend without breaking origin
trial users. So far we have been lowering build_vectors that would otherwise
have been lowered to v128.const to splats followed by sequences of replace_lane
instructions to initialize each lane individually. That produces large and
inefficient code, so this patch introduces new logic to lower integer vector
constants to a single i64x2.splat where possible, with at most a single
i64x2.replace_lane following it if necessary.
Adapted from a patch authored by @omnisip.
Differential Revision: https://reviews.llvm.org/D88591
The TypePromotion pass only operates on scalar types so I've fixed up
all places where we were relying upon the implicit cast from
TypeSize->uint64_t.
Differential Revision: https://reviews.llvm.org/D88575
When we know that a particular type is always going to be fixed
width we have so far been writing code like this:
getSizeInBits().getFixedSize()
Since we are doing this in quite a few places now it seems to make
sense to add a new helper function that allows us to replace
these calls with a single getFixedSizeInBits() call.
Differential Revision: https://reviews.llvm.org/D88649
For these cases, we already omit the prologue directives, if
(!AFI->hasStackFrame() && !windowsRequiresStackProbe && !NumBytes).
When writing the epilogue (after the prolog has been written), if
the function doesn't have the WinCFI flag set (i.e. if no prologue
was generated), assume that no epilogue will be needed either,
and don't emit any epilog start pseudo instruction. After completing
the epilogue, make sure that it actually matched the prologue.
Previously, when epilogue start/end was generated, but no prologue,
the unwind info for such functions actually was huge; 12 bytes xdata
(4 bytes header, 4 bytes for one non-folded epilogue header, 4 bytes
for padded opcodes) and 8 bytes pdata. Because the epilog consisted of
one opcode (end) but the prolog was empty (no .seh_endprologue), the
epilogue couldn't be folded into the prologue, and thus couldn't be
considered for packed form either.
On a 6.5 MB DLL with 110 KB pdata and 166 KB xdata, this gets rid of
38 KB pdata and 62 KB xdata.
Differential Revision: https://reviews.llvm.org/D88641
The documentation for the NormalizeMemRefs pass and the associated MemRefsNormalizable
traits was confusing and not on the website. This update clarifies the language
around the difference between a MemRef Type, an operation that accesses the value of
MemRef Type, and better documents the limitations of the current implementation.
This patch also includes some basic debugging information for the pass so people
might have a chance of figuring out why it doesn't work on their code.
Differential Revision: https://reviews.llvm.org/D88532
The logic there only considers `SLT/SGT` predicates. We can use the same logic
for proving `ULT/UGT` predicates if all involved values are non-negative.
Adding full-scale support for unsigned might be challenging because of code amount,
so we can consider this in the future.
Differential Revision: https://reviews.llvm.org/D88087
Reviewed By: reames
If we try to coerce a vector of non-integral pointers to a narrower type (either narrower vector or single pointer), we use inttoptr and violate the semantics of non-integral pointers. In theory, we can handle many of these cases, we just need to use a different code idiom to convert without going through inttoptr and back.
This shows up as wrong code bugs, and in some cases, crashes due to failed asserts. Modeled after a change which has lived downstream for a couple years, though completely rewritten to be more idiomatic.
Convert to use new MachineBasicBlock splitAt function.
Place code in splitBlock function for reuse in future changes.
Should yield no functional change.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D88537
lldb's PlatforDarwinKernel scans the local filesystem (well known
locations, plus user-specified directories) for kernels and kexts
when doing kernel debugging, and loads them automatically. Sometimes
kernel developers want to debug with *only* a dSYM, in which case they
give lldb the DWARF binary + the dSYM as a binary and symbol file.
This patch adds code to lldb to do this automatically if that's the
best thing lldb can find.
A few other bits of cleanup in PlatformDarwinKernel that I undertook
at the same time:
1. Remove the 'platform.plugin.darwin-kernel.search-locally-for-kexts'
setting. When I added the local filesystem index at start of kernel
debugging, I thought people might object to the cost of the search
and want a way to disable it. No one has.
2. Change the behavior of
'plugin.dynamic-loader.darwin-kernel.load-kexts' setting so it does
not disable the local filesystem scan, or use of the local filesystem
binaries.
3. PlatformDarwinKernel::GetSharedModule into GetSharedModuleKext and
GetSharedModuleKernel for easier readability & maintenance.
4. Added accounting of .dSYM.yaa files (an archive format akin to tar)
that I come across during the scan. I'm not using these for now; it
would be very expensive to expand the archives & see if the UUID matches
what I'm searching for.
<rdar://problem/69774993>
Differential Revision: https://reviews.llvm.org/D88632
Summary: How we copying the CRRC to GRC is using a single MFOCRF to copy the contents of CR field n (CR bits 4×n+32:4×n+35) into bits 4×n+32:4×n+35 of register GRC. That’s not correct because we expect the value of destination register equals to source so we have to put the the contents of CR field in the lowest 4 bits. This patch adds a RLWINM after MFOCRF to achieve that.
The problem came up when adding builtins for xvtdivdp, xvtdivsp, xvtsqrtdp, xvtsqrtsp, as posted in D88278. We need to move the outputs (in CR register) to GRC. However outputs of these instructions may not in a fixed CR# register, so we can’t directly add a rotation instruction in the .td patterns, but need to wait until the CR register is determined. Then we confirmed this should be a bug in POST-RA PSEUDO PASS.
Reviewed By: nemanjai, shchenz
Differential Revision: https://reviews.llvm.org/D88274
Support integer and default character stop codes on PAUSE
statements. Add length argument to STOP statement with a
character stop code.
Differential revision: https://reviews.llvm.org/D88692
It turns out that unformatted fixed-size output records
do need to be padded out if short, in order to avoid a
spurious EOF crash on a short record at the end of the file.
While here in AdvanceRecord(), move the unformatted
variable-length record header/footer writing code to here
from EndIoStatement().
Differential revision: https://reviews.llvm.org/D88685
Summary:
Some design decision worth noting about:
I've noticed a recent mailing discussing about why string literal is
not affected by -fdata-sections for ELF target:
http://lists.llvm.org/pipermail/llvm-dev/2020-September/145121.html
But on AIX, our linker could not split the mergeable string like other target.
So I think it would make more sense for us to emit separate csect for
every mergeable string in -fdata-sections mode,
as there might not be other ways for linker to do garbage collection
on unused mergeable string.
Reviewed By: daltenty, hubert.reinterpretcast
Differential Revision: https://reviews.llvm.org/D88339
The buffer needs to be Reset() after a Flush(), since the
Flush() can be a no-op after a read->write transition.
And record numbers are 1-based, not 0-based.
This fixes a bug with rewrites of records that have been
recently read.
Differential revision: https://reviews.llvm.org/D88612
A recent MSVC work-around patch is eliciting unused variable
warnings from clang; package the lambda reference arguments
into a struct to avoid the warning.
Differential revision: https://reviews.llvm.org/D88695
If we allow the non-integral pointers to become memset and memcpy, we loose the ability to reason about pointer propagation. This patch is modeled on changes we've carried downstream for a long time, figured it was worth being equally conservative for other users. There is room to refine the semantics and handling here if anyone is motivated.
Allows the creation of real SOP1 instructions with
assembler mnemonics that differ from their
pseudo-instruction mnemonics. The default behavior
keeps the mnemonics matching.
Corrects a subtarget label typo in a comment.
Authored By: Joe_Nash
Differential Revision: https://reviews.llvm.org/D88708
Tweak binary->decimal conversions to avoid an integer multiplication
in a hot loop to improve readability and get a minor (~5%) speed-up.
Use native integer division by constants for more readability, too,
since current build compilers seem to optimize it correctly now.
Delete the now needless temporary work-around facility in
Common/unsigned-const-division.h.
Differential revision: https://reviews.llvm.org/D88604
Similar to the FP case in `AArch64TargetLowering::LowerBR_CC`.
Instead of emitting the csets + a tbnz, just emit a compare + bcc
(or two bccs, depending on the condition code)
This improves cases like this: https://godbolt.org/z/v8hebx
This is a 0.1% geomean code size improvement for CTMark at -O3.
Differential Revision: https://reviews.llvm.org/D88624
Partially refactoring, partially fixing a bug.
- We shouldn't use TB(N)ZX unless the bit number is >= 32
- We can fold more than xor using emitTestBit
Also remove a check which isn't relevant anymore + update tests.
Rename select-brcond-of-not.mir to select-brcond-of-binop.mir, since it now
tests more than just G_XOR.
Differential Revision: https://reviews.llvm.org/D88702
Before this patch /summary was crashing with some .PCH.OBJ files, because tpiMap[srcIdx++] was reading at the wrong location. When the TpiSource depends on a .PCH.OBJ file, the types should be offset by the previously merged PCH.OBJ set of indices.
Differential Revision: https://reviews.llvm.org/D88678
This seems to fail on ubuntu 18.04.5 with Clang 9 due to:
Error output:
error: Couldn't lookup symbols:
std::__1::default_delete<int>::operator()(int) const
The script's shebang wants Python 3, so we use FindPython3. The
original code didn't work when an unversioned python was not available.
This is explicitly allowed in PEP 394. ("Distributors may choose to set
the behavior of the python command as follows: python2, python3, not
provide python command, allow python to be configurable by an end user
or a system administrator.")
Also I think it's actually required, so let the configuration fail if we
can't find it.
Lastly remove the shebang, since the script is only run via interpreter
and doesn't have the executable bit set anyway.
Reviewed By: jvesely
Differential Revision: https://reviews.llvm.org/D88366
Truncating to v8i8 is a case where we want to split the source but also generate
intermediate truncates to reduce the size of the source vector before truncating
down to v8i8. This implements the same strategy as what SelectionDAG does, but
I'm not certain where if anywhere in generic code it should live.
Use it for legalization of v8s8 = G_ICMP v8s32.
Differential Revision: https://reviews.llvm.org/D88191
Nicolas Vasilache