This introduces the `captures` attribute as described in:
https://discourse.llvm.org/t/rfc-improvements-to-capture-tracking/81420
This initial patch only introduces the IR/bitcode support for the
attribute and its in-memory representation as `CaptureInfo`. This will
be followed by a patch to upgrade and remove the `nocapture` attribute,
and then by actual inference/analysis support.
Based on the RFC feedback, I've used a syntax similar to the `memory`
attribute, though the only "location" that can be specified is `ret`.
I've added some pretty extensive documentation to LangRef on the
semantics. One non-obvious bit here is that using ptrtoint will not
result in a "return-only" capture, even if the ptrtoint result is only
used in the return value. Without this requirement we wouldn't be able
to continue ordinary capture analysis on the return value.
Add support for taking the intersection of two AttributeLists s.t the
result list contains attributes that are valid in the context of both
inputs.
i.e if we have `nonnull align(32) noundef` intersected with `nonnull
align(16) dereferenceable(10)`, the result is `nonnull align(16)`.
Further it handles attributes that are not-droppable. For example
dropping `byval` can change the nature of a callsite/function so its
impossible to correct a correct intersection if its dropped from the
result. i.e `nonnull byval(i64)` intersected with `nonnull` is
invalid.
The motivation for the infrastructure is to enable sinking/hoisting
callsites with differing attributes.
This carries a bitmask indicating forbidden floating-point value kinds
in the argument or return value. This will enable interprocedural
-ffinite-math-only optimizations. This is primarily to cover the
no-nans and no-infinities cases, but also covers the other floating
point classes for free. Textually, this provides a number of names
corresponding to bits in FPClassTest, e.g.
call nofpclass(nan inf) @must_be_finite()
call nofpclass(snan) @cannot_be_snan()
This is more expressive than the existing nnan and ninf fast math
flags. As an added bonus, you can represent fun things like nanf:
declare nofpclass(inf zero sub norm) float @only_nans()
Compared to nnan/ninf:
- Can be applied to individual call operands as well as the return value
- Can distinguish signaling and quiet nans
- Distinguishes the sign of infinities
- Can be safely propagated since it doesn't imply anything about
other operands.
- Does not apply to FP instructions; it's not a flag
This is one step closer to being able to retire "no-nans-fp-math" and
"no-infs-fp-math". The one remaining situation where we have no way to
represent no-nans/infs is for loads (if we wanted to solve this we
could introduce !nofpclass metadata, following along with
noundef/!noundef).
This is to help simplify the GPU builtin math library
distribution. Currently the library code has explicit finite math only
checks, read from global constants the compiler driver needs to set
based on the compiler flags during linking. We end up having to
internalize the library into each translation unit in case different
linked modules have different math flags. By propagating known-not-nan
and known-not-infinity information, we can automatically prune the
edge case handling in most functions if the function is only reached
from fast math uses.
Use deduction guides instead of helper functions.
The only non-automatic changes have been:
1. ArrayRef(some_uint8_pointer, 0) needs to be changed into ArrayRef(some_uint8_pointer, (size_t)0) to avoid an ambiguous call with ArrayRef((uint8_t*), (uint8_t*))
2. CVSymbol sym(makeArrayRef(symStorage)); needed to be rewritten as CVSymbol sym{ArrayRef(symStorage)}; otherwise the compiler is confused and thinks we have a (bad) function prototype. There was a few similar situation across the codebase.
3. ADL doesn't seem to work the same for deduction-guides and functions, so at some point the llvm namespace must be explicitly stated.
4. The "reference mode" of makeArrayRef(ArrayRef<T> &) that acts as no-op is not supported (a constructor cannot achieve that).
Per reviewers' comment, some useless makeArrayRef have been removed in the process.
This is a follow-up to https://reviews.llvm.org/D140896 that introduced
the deduction guides.
Differential Revision: https://reviews.llvm.org/D140955
As suggested on D135572, return Optional<> from getAllocSizeArgs()
rather than the peculiar pair(0, 0) sentinel.
The method on Attribute itself does not return Optional, because
the attribute must exist in that case.
This regularly comes up as a stumbling stone when adding int
attributes: They currently need to be encoded in a way to avoids
the zero value.
This adds support for zero-value int attributes by a) making the
ctor determine int/enum attribute based on attribute kind, not
whether the value is non-zero and b) switching getRawIntAttr()
to return an Optional, so that it's possible to distinguish a zero
value from non-existence.
Differential Revision: https://reviews.llvm.org/D135572
I chose to encode the allockind information in a string constant because
otherwise we would get a bit of an explosion of keywords to deal with
the possible permutations of allocation function types.
I'm not sure that CodeGen.h is the correct place for this enum, but it
seemed to kind of match the UWTableKind enum so I put it in the same
place. Constructive suggestions on a better location most certainly
encouraged.
Differential Revision: https://reviews.llvm.org/D123088
We have the `clang -cc1` command-line option `-funwind-tables=1|2` and
the codegen option `VALUE_CODEGENOPT(UnwindTables, 2, 0) ///< Unwind
tables (1) or asynchronous unwind tables (2)`. However, this is
encoded in LLVM IR by the presence or the absence of the `uwtable`
attribute, i.e. we lose the information whether to generate want just
some unwind tables or asynchronous unwind tables.
Asynchronous unwind tables take more space in the runtime image, I'd
estimate something like 80-90% more, as the difference is adding
roughly the same number of CFI directives as for prologues, only a bit
simpler (e.g. `.cfi_offset reg, off` vs. `.cfi_restore reg`). Or even
more, if you consider tail duplication of epilogue blocks.
Asynchronous unwind tables could also restrict code generation to
having only a finite number of frame pointer adjustments (an example
of *not* having a finite number of `SP` adjustments is on AArch64 when
untagging the stack (MTE) in some cases the compiler can modify `SP`
in a loop).
Having the CFI precise up to an instruction generally also means one
cannot bundle together CFI instructions once the prologue is done,
they need to be interspersed with ordinary instructions, which means
extra `DW_CFA_advance_loc` commands, further increasing the unwind
tables size.
That is to say, async unwind tables impose a non-negligible overhead,
yet for the most common use cases (like C++ exceptions), they are not
even needed.
This patch extends the `uwtable` attribute with an optional
value:
- `uwtable` (default to `async`)
- `uwtable(sync)`, synchronous unwind tables
- `uwtable(async)`, asynchronous (instruction precise) unwind tables
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D114543
While it is nice to have separate methods in the public AttributeSet
API, we can fetch the type from the internal AttributeSetNode
using a generic API for all type attribute kinds.
This is now the same as isIntAttrKind(), so use that instead, as
it does not require manual maintenance. The naming is also more
accurate in that both int and type attributes have an argument,
but this method was only targeting int attributes.
I initially wanted to tighten the AttrBuilder assertion, but we
have some in-tree uses that would violate it.
Such attributes can either be unset, or set to "true" or "false" (as string).
throughout the codebase, this led to inelegant checks ranging from
if (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true")
to
if (Fn->hasAttribute("no-jump-tables") && Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true")
Introduce a getValueAsBool that normalize the check, with the following
behavior:
no attributes or attribute set to "false" => return false
attribute set to "true" => return true
Differential Revision: https://reviews.llvm.org/D99299
I think byval/sret and the others are close to being able to rip out
the code to support the missing type case. A lot of this code is
shared with inalloca, so catch this up to the others so that can
happen.
This attribute represents the minimum and maximum values vscale can
take. For now this attribute is not hooked up to anything during
codegen, this will be added in the future when such codegen is
considered stable.
Additionally hook up the -msve-vector-bits=<x> clang option to emit this
attribute.
Differential Revision: https://reviews.llvm.org/D98030
This reverts commit eb9f7c28e5fe6d75fed3587023e17f2997c8024b.
Previously this was incorrectly handling linking of the contained
type, so this merges the fixes from D88973.
This reverts commit 55c4ff91bd820d72014f63dcf7f3d5a0d3397986.
Issues were introduced as discussed in https://reviews.llvm.org/D88241
where this change made previous bugs in the linker and BitCodeWriter
visible.
Make the corresponding change that was made for byval in
b7141207a483d39b99c2b4da4eb3bb591eca9e1a. Like byval, this requires a
bulk update of the test IR tests to include the type before this can
be mandatory.
There's a special case in hasAttribute for None when pImpl is null. If pImpl is not null we dispatch to pImpl->hasAttribute which will always return false for Attribute::None.
So if we just want to check for None its sufficient to just check that pImpl is null. Which can even be done inline.
This patch adds a helper for that case which I hope will speed up our getSubtargetImpl implementations.
Differential Revision: https://reviews.llvm.org/D86744
This allows tracking the in-memory type of a pointer argument to a
function for ABI purposes. This is essentially a stripped down version
of byval to remove some of the stack-copy implications in its
definition.
This includes the base IR changes, and some tests for places where it
should be treated similarly to byval. Codegen support will be in a
future patch.
My original attempt at solving some of these problems was to repurpose
byval with a different address space from the stack. However, it is
technically permitted for the callee to introduce a write to the
argument, although nothing does this in reality. There is also talk of
removing and replacing the byval attribute, so a new attribute would
need to take its place anyway.
This is intended avoid some optimization issues with the current
handling of aggregate arguments, as well as fixes inflexibilty in how
frontends can specify the kernel ABI. The most honest representation
of the amdgpu_kernel convention is to expose all kernel arguments as
loads from constant memory. Today, these are raw, SSA Argument values
and codegen is responsible for turning these into loads.
Background:
There currently isn't a satisfactory way to represent how arguments
for the amdgpu_kernel calling convention are passed. In reality,
arguments are passed in a single, flat, constant memory buffer
implicitly passed to the function. It is also illegal to call this
function in the IR, and this is only ever invoked by a driver of some
kind.
It does not make sense to have a stack passed parameter in this
context as is implied by byval. It is never valid to write to the
kernel arguments, as this would corrupt the inputs seen by other
dispatches of the kernel. These argumets are also not in the same
address space as the stack, so a copy is needed to an alloca. From a
source C-like language, the kernel parameters are invisible.
Semantically, a copy is always required from the constant argument
memory to a mutable variable.
The current clang calling convention lowering emits raw values,
including aggregates into the function argument list, since using
byval would not make sense. This has some unfortunate consequences for
the optimizer. In the aggregate case, we end up with an aggregate
store to alloca, which both SROA and instcombine turn into a store of
each aggregate field. The optimizer never pieces this back together to
see that this is really just a copy from constant memory, so we end up
stuck with expensive stack usage.
This also means the backend dictates the alignment of arguments, and
arbitrarily picks the LLVM IR ABI type alignment. By allowing an
explicit alignment, frontends can make better decisions. For example,
there's real no advantage to an aligment higher than 4, so a frontend
could choose to compact the argument layout. Similarly, there is a
high penalty to using an alignment lower than 4, so a frontend could
opt into more padding for small arguments.
Another design consideration is when it is appropriate to expose the
fact that these arguments are all really passed in adjacent
memory. Currently we have a late IR optimization pass in codegen to
rewrite the kernel argument values into explicit loads to enable
vectorization. In most programs, unrelated argument loads can be
merged together. However, exposing this property directly from the
frontend has some disadvantages. We still need a way to track the
original argument sizes and alignments to report to the driver. I find
using some side-channel, metadata mechanism to track this
unappealing. If the kernel arguments were exposed as a single buffer
to begin with, alias analysis would be unaware that the padding bits
betewen arguments are meaningless. Another family of problems is there
are still some gaps in replacing all of the available parameter
attributes with metadata equivalents once lowered to loads.
The immediate plan is to start using this new attribute to handle all
aggregate argumets for kernels. Long term, it makes sense to migrate
all kernel arguments, including scalars, to be passed indirectly in
the same manner.
Additional context is in D79744.
I noticed that for some benchmarks we spend quite a bit of time
inside AttributeList::hasAttrSomewhere(), mainly when checking
for the "returned" attribute. Most of the time the attribute will
not be present, in which case this function has to walk through
the whole attribute list and check for the attribute at each index.
This patch adds a cache of all "available somewhere" attributes
inside AttributeListImpl. This makes the structure 12 bytes larger,
but I don't think that's problematic, as attribute lists are uniqued.
Compile-time in terms of instructions retired improves by 0.4% on
average, but >1% for sqlite.
Differential Revision: https://reviews.llvm.org/D81867
Add llvm.call.preallocated.{setup,arg} instrinsics.
Add "preallocated" operand bundle which takes a token produced by llvm.call.preallocated.setup.
Add "preallocated" parameter attribute, which is like byval but without the copy.
Verifier changes for these IR constructs.
See https://github.com/rnk/llvm-project/blob/call-setup-docs/llvm/docs/CallSetup.md
Subscribers: hiraditya, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74651
This means AttrBuilder will always create a sorted set of attributes and
we can skip the sorting step. Sorting attributes is surprisingly
expensive, and I recently made it worse by making it use array_pod_sort.
Attributes are currently stored as a simple list. Enum attributes
additionally use a bitset to allow quickly determining whether an
attribute is set. String attributes on the other hand require a
full scan of the list. As functions tend to have a lot of string
attributes (at least when clang is used), this is a noticeable
performance issue.
This patch adds an additional name => attribute map to the
AttributeSetNode, which allows querying string attributes quickly.
This results in a 3% reduction in instructions retired on CTMark.
Changes to memory usage seem to be in the noise (attribute sets are
uniqued, and we don't tend to have more than a few dozen or hundred
unique attribute sets, so adding an extra map does not have a
noticeable cost.)
Differential Revision: https://reviews.llvm.org/D78859
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
TypeAttributeImpl inherits from EnumAttributeImpl which already defines anchor() as a virtual, so we should override this instead of redeclaring it.
llvm-svn: 372877
