This switches everything to use the memory attribute proposed in
https://discourse.llvm.org/t/rfc-unify-memory-effect-attributes/65579.
The old argmemonly, inaccessiblememonly and inaccessiblemem_or_argmemonly
attributes are dropped. The readnone, readonly and writeonly attributes
are restricted to parameters only.
The old attributes are auto-upgraded both in bitcode and IR.
The bitcode upgrade is a policy requirement that has to be retained
indefinitely. The IR upgrade is mainly there so it's not necessary
to update all tests using memory attributes in this patch, which
is already large enough. We could drop that part after migrating
tests, or retain it longer term, to make it easier to import IR
from older LLVM versions.
High-level Function/CallBase APIs like doesNotAccessMemory() or
setDoesNotAccessMemory() are mapped transparently to the memory
attribute. Code that directly manipulates attributes (e.g. via
AttributeList) on the other hand needs to switch to working with
the memory attribute instead.
Differential Revision: https://reviews.llvm.org/D135780
The pointsToConstantMemory() method returns true only if the memory pointed to
by the memory location is globally invariant. However, the LLVM memory model
also has the semantic notion of *locally-invariant*: memory that is known to be
invariant for the life of the SSA value representing that pointer. The most
common example of this is a pointer argument that is marked readonly noalias,
which the Rust compiler frequently emits.
It'd be desirable for LLVM to treat locally-invariant memory the same way as
globally-invariant memory when it's safe to do so. This patch implements that,
by introducing the concept of a *ModRefInfo mask*. A ModRefInfo mask is a bound
on the Mod/Ref behavior of an instruction that writes to a memory location,
based on the knowledge that the memory is globally-constant memory (in which
case the mask is NoModRef) or locally-constant memory (in which case the mask
is Ref). ModRefInfo values for an instruction can be combined with the
ModRefInfo mask by simply using the & operator. Where appropriate, this patch
has modified uses of pointsToConstantMemory() to instead examine the mask.
The most notable optimization change I noticed with this patch is that now
redundant loads from readonly noalias pointers can be eliminated across calls,
even when the pointer is captured. Internally, before this patch,
AliasAnalysis was assigning Ref to reads from constant memory; now AA can
assign NoModRef, which is a tighter bound.
Differential Revision: https://reviews.llvm.org/D136659
Reapplying after the fix for volatile modelling in D135863.
-----
Don't add argmem if the pointer is clearly not an argument (e.g.
a global). I don't think this makes a difference right now, but
gives more obvious results with D135780.
Per LangRef, volatile operations are allowed to access the location
of their pointer argument, plus inaccessible memory:
> Any volatile operation can have side effects, and any volatile
> operation can read and/or modify state which is not accessible
> via a regular load or store in this module.
> [...]
> The allowed side-effects for volatile accesses are limited. If
> a non-volatile store to a given address would be legal, a volatile
> operation may modify the memory at that address. A volatile
> operation may not modify any other memory accessible by the
> module being compiled. A volatile operation may not call any
> code in the current module.
FunctionAttrs currently does not model this and ends up marking
functions with volatile accesses on arguments as argmemonly,
even though they should be inaccessiblemem_or_argmemonly.
Differential Revision: https://reviews.llvm.org/D135863
Followup to D135962 to rename remaining uses of
FunctionModRefBehavior to MemoryEffects. Does not touch API names
yet, but also updates variables names FMRB/MRB to ME, to match the
new type name.
This reverts commit b05f5b90a1.
There are thread sanitizer buildbot failures in simple_stack.c.
I think that's because this ended up affecting the handling of
volatile accesses to allocas. Reverting for now.
Don't add argmem if the pointer is clearly not an argument (e.g.
a global). I don't think this makes a difference right now, but
gives more obvious results with D135780.
We have to account for accesses to argument memory via captures.
I don't think there's any way to make this produce incorrect
results right now (because as soon as "other" is set, we lose the
ability to infer argmemonly), but this avoids incorrect results
once we have more precise representation.
The code for inferring memory attributes on arguments claims that
inalloca/preallocated arguments are always clobbered:
d71ad41080/llvm/lib/Transforms/IPO/FunctionAttrs.cpp (L640-L642)
However, we would still infer memory attributes for the whole
function without taking this into account, so we could still end
up inferring readnone for the function. This adds an argument
clobber if there are any inalloca/preallocated arguments.
Differential Revision: https://reviews.llvm.org/D135783
The previous version of the patch would incorrect convert an
existing argmemonly attribute into an inaccessiblemem_or_argmemonly
attribute.
-----
This updates checkFunctionMemoryAccess() to infer a precise
FunctionModRefBehavior, rather than an approximation split into
read/write and argmemonly.
Afterwards, we still map this back to imprecise function attributes.
This still allows us to infer some cases that we previously did not
handle, namely inaccessiblememonly and inaccessiblemem_or_argmemonly.
In practice, this means we get better memory attributes in the
presence of intrinsics like @llvm.assume.
Differential Revision: https://reviews.llvm.org/D134527
This updates checkFunctionMemoryAccess() to infer a precise
FunctionModRefBehavior, rather than an approximation split into
read/write and argmemonly.
Afterwards, we still map this back to imprecise function attributes.
This still allows us to infer some cases that we previously did not
handle, namely inaccessiblememonly and inaccessiblemem_or_argmemonly.
In practice, this means we get better memory attributes in the
presence of intrinsics like @llvm.assume.
Differential Revision: https://reviews.llvm.org/D134527
MemoryLocation::getOrNone() already has the necessary logic to
handle different instruction types. Use it, rather than repeating
a subset of the logic. This adds support for previously unhandled
instructions like atomicrmw.
Currently, FunctionModRefBehavior tracks whether the function reads
or writes memory (ModRefInfo) and which locations it can access
(argmem, inaccessiblemem and other). This patch changes it to track
ModRef information per-location instead.
To give two examples of why this is useful:
* D117095 highlights a weakness of ModRef modelling in the presence
of operand bundles. For a memcpy call with deopt operand bundle,
we want to say that it can read any memory, but only write argument
memory. This would allow them to be treated like any other calls.
However, we currently can't express this and have to say that it
can read or write any memory.
* D127383 would ideally be modelled as a separate threadid location,
where threadid Refs outside pre-split coroutines can be ignored
(like other accesses to constant memory). The current representation
does not allow modelling this precisely.
The patch as implemented is intended to be NFC, but there are some
obvious opportunities for improvements and simplification. To fully
capitalize on this we would also want to change the way we represent
memory attributes on functions, but that's a larger change, and I
think it makes sense to separate out the FunctionModRefBehavior
refactoring.
Differential Revision: https://reviews.llvm.org/D130896
We wanted to check if all uses of the function are direct calls, but the
code didn't account for passing the function as a parameter.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D128104
This patch adds initial argmemonly inference, by checking the underlying
objects of locations returned by MemoryLocation.
I think this should cover most cases, except function calls to other
argmemonly functions.
I'm not sure if there's a reason why we don't infer those yet.
Additional argmemonly can improve codegen in some cases. It also makes
it easier to come up with a C reproducer for 7662d1687b (already fixed,
but I'm trying to see if C/C++ fuzzing could help to uncover similar
issues.)
Compile-time impact:
NewPM-O3: +0.01%
NewPM-ReleaseThinLTO: +0.03%
NewPM-ReleaseLTO+g: +0.05%
https://llvm-compile-time-tracker.com/compare.php?from=067c035012fc061ad6378458774ac2df117283c6&to=fe209d4aab5b593bd62d18c0876732ddcca1614d&stat=instructions
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D121415
Update FunctionAttrs to use FunctionModRefBehavior instead
MemoryAccessKind.
This allows for adding support for inferring argmemonly and others,
see D121415.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D121460
There was a fixme in the code pertaining to attributing functions as
noreturn. By using reachability, if none of the blocks that are
reachable from the entry return, then the function is noreturn.
Previously, the code only checked if any blocks returned. If they're
unreachable, then they don't matter.
This improves codegen for the Linux kernel.
Fixes: https://github.com/ClangBuiltLinux/linux/issues/1563
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D119571
The naming has come up as a source of confusion in several recent reviews. onlyWritesMemory is consist with onlyReadsMemory which we use for the corresponding readonly case as well.
This builds on the code from D114963, and extends it to handle calls both direct and indirect. With the revised code structure (from series of previously landed NFCs), this is pretty straight forward.
One thing to note is that we can not infer writeonly for arguments which might be captured. If the pointer can be read back by the caller, and then read through, we have no way to track that. This is the same restriction we have for readonly, except that we get no mileage out of the "callee can be readonly" exception since a writeonly param on a readonly function is either a) readnone or b) UB. This means we can't actually infer much unless nocapture has already been inferred.
Differential Revision: https://reviews.llvm.org/D115003
This class is solely used as a lightweight and clean way to build a set of
attributes to be removed from an AttrBuilder. Previously AttrBuilder was used
both for building and removing, which introduced odd situation like creation of
Attribute with dummy value because the only relevant part was the attribute
kind.
Differential Revision: https://reviews.llvm.org/D116110
Arguments to an indirect call is by definition outside the SCC, but there's no reason we can't use locally defined facts on the call site. This also has the nice effect of further simplifying the code.
Differential Revision: https://reviews.llvm.org/D116118
This change allows us to infer access attributes (readnone, readonly) on arguments passed to vararg functions. Since there isn't a formal argument corresponding to the parameter, they'll never be considered part of the speculative SCC, but they can still benefit from attributes on the call site or the callee function.
The main motivation here is just to simplify the code, and remove some special casing. Previously, an indirect vararg call could return more precise results than an direct vararg call which is just weird.
Differential Revision: https://reviews.llvm.org/D115964
This fixes a bug where we would infer readnone/readonly for a function which passed a value to a function which could capture it. With the value captured in memory, the function could reload the value from memory after the call, and write to it. Inferring the argument as readnone or readonly is unsound.
@jdoerfert apparently noticed this about two years ago, and tests were checked in with 76467c4, but the issue appears to have never gotten fixed.
Since this seems like this issue should break everything, let me explain why the case is actually fairly narrow. The main inference loop over the argument SCCs only analyzes nocapture arguments. As such, we can only hit this when construction the partial SCCs. Due to that restriction, we can only hit this when we have either a) a function declaration with a manually annotated argument, or b) an immediately self recursive call.
It's also worth highlighting that we do have cases we can infer readonly/readnone on a capturing argument validly. The easiest example is a function which simply returns its argument without ever accessing it.
Differential Revision: https://reviews.llvm.org/D115961
Instead of having the speculative path be the untaken path in the branch, explicitly have it return. This does require tail duplicating one call, but the resulting code is shorter and easier to understand. Also rewrite the condition using appropriate accessors.
We were being wildly inconsistent about what memory access was implied by an indirect function call. Depending on the call site attributes, you could get anything from a read, to unknown, to none at all. (The last was a miscompile.)
We were also always traversing the uses of a readonly indirect call. This is entirely unneeded as the indirect call does not capture. The callee might capture itself internally, but that has no implications for this caller. (See the nice explanation in the CaptureTracking comments if that case is confusing.)
Note that elsewhere in the same file, we were correctly computing the nocapture attribute for indirect calls. The changed case only resulted in conservatism when computing memory attributes if say the return value was written to.
Differential Revision: https://reviews.llvm.org/D115916
This fixes a bug in 740057d. There's two ways to describe the issue:
* One caller hasn't yet proven nocapture on the argument. Given that, the inference routine is responsible for bailing out on a potential capture.
* Even if we know the argument is nocapture, the access inference needs to traverse the exact set of users the capture tracking would (or exit conservatively). Even if capture tracking can prove a store is non-capturing (e.g. to a local alloc which doesn't escape), we still need to track the copy of the pointer to see if it's later reloaded and accessed again.
Note that all the test changes except the newly added ones appear to be false negatives. That is, cases where we could prove writeonly, but the current code isn't strong enough. That's why I didn't spot this originally.
This change extends the current logic for inferring readonly and readnone argument attributes to also infer writeonly.
This change is deliberately minimal; there's a couple of areas for follow up.
* I left out all call handling and thus any benefit from the SCC walk. When examining the test changes, I realized the existing code is imprecise, and am going to fix that in it's own revision before adding in the writeonly handling. (Mostly because updating the tests is hard when I, the human, can't figure out whether the result is correct.)
* I left out handling for storing a value (as opposed to storing to a pointer). This should benefit readonly/readnone as well, and applies to a bunch of other instructions. Seemed worth having as a separate review.
Differential Revision: https://reviews.llvm.org/D114963
Previously, any change in any function in an SCC would cause all
analyses for all functions in the SCC to be invalidated. With this
change, we now manually invalidate analyses for functions we modify,
then let the pass manager know that all function analyses should be
preserved since we've already handled function analysis invalidation.
So far this only touches the inliner, argpromotion, function-attrs, and
updateCGAndAnalysisManager(), since they are the most used.
This is part of an effort to investigate running the function
simplification pipeline less on functions we visit multiple times in the
inliner pipeline.
However, this causes major memory regressions especially on larger IR.
To counteract this, turn on the option to eagerly invalidate function
analyses. This invalidates analyses on functions immediately after
they're processed in a module or scc to function adaptor for specific
parts of the pipeline.
Within an SCC, if a pass only modifies one function, other functions in
the SCC do not have their analyses invalidated, so in later function
passes in the SCC pass manager the analyses may still be cached. It is
only after the function passes that the eager invalidation takes effect.
For the default pipelines this makes sense because the inliner pipeline
runs the function simplification pipeline after all other SCC passes
(except CoroSplit which doesn't request any analyses).
Overall this has mostly positive effects on compile time and positive effects on memory usage.
https://llvm-compile-time-tracker.com/compare.php?from=7f627596977624730f9298a1b69883af1555765e&to=39e824e0d3ca8a517502f13032dfa67304841c90&stat=instructionshttps://llvm-compile-time-tracker.com/compare.php?from=7f627596977624730f9298a1b69883af1555765e&to=39e824e0d3ca8a517502f13032dfa67304841c90&stat=max-rss
D113196 shows that we slightly regressed compile times in exchange for
some memory improvements when turning on eager invalidation. D100917
shows that we slightly improved compile times in exchange for major
memory regressions in some cases when invalidating less in SCC passes.
Turning these on at the same time keeps the memory improvements while
keeping compile times neutral/slightly positive.
Reviewed By: asbirlea, nikic
Differential Revision: https://reviews.llvm.org/D113304
This is in preparation for only invalidating analyses on changed
functions.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D113303
Coroutines have weird semantics that don't quite match normal LLVM functions,
so trying to infer even simple attributes based on thier contents can go wrong.
Thinlink provides an opportunity to propagate function attributes across modules, enabling additional propagation opportunities.
This change propagates (currently default off, turn on with `disable-thinlto-funcattrs=1`) noRecurse and noUnwind based off of function summaries of the prevailing functions in bottom-up call-graph order. Testing on clang self-build:
1. There's a 35-40% increase in noUnwind functions due to the additional propagation opportunities.
2. Throughput is measured at 10-15% increase in thinlink time which itself is 1.5% of E2E link time.
Implementation-wise this adds the following summary function attributes:
1. noUnwind: function is noUnwind
2. mayThrow: function contains a non-call instruction that `Instruction::mayThrow` returns true on (e.g. windows SEH instructions)
3. hasUnknownCall: function contains calls that don't make it into the summary call-graph thus should not be propagated from (e.g. indirect for now, could add no-opt functions as well)
Testing:
Clang self-build passes and 2nd stage build passes check-all
ninja check-all with newly added tests passing
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D36850
getMetadata() currently uses a weird API where it populates a
structure passed to it, and optionally merges into it. Instead,
we can return the AAMDNodes and provide a separate merge() API.
This makes usages more compact.
Differential Revision: https://reviews.llvm.org/D109852
AttributeList::hasAttribute() is confusing, use clearer methods like
hasParamAttr()/hasRetAttr().
Add hasRetAttr() since it was missing from AttributeList.
Krzysztof Parzyszek