Move isImplied{True,False}ByMatchingCmp from CmpInst to ICmpInst, so
that it can operate on CmpPredicate instead of CmpInst::Predicate, and
teach it about samesign. There are two callers of this function, and we
choose to migrate the one in ValueTracking, namely
isImpliedCondMatchingOperands to CmpPredicate, hence teaching it about
samesign, with visible test impact.
Strip hash_value() for CmpPredicate, as different callers have different
hashing use-cases. In this case, there is just one caller, namely
EarlyCSE, which calls hash_combine() on a CmpPredicate, which used to
call hash_combine() on a CmpInst::Predicate prior to 4a0d53a
(PatternMatch: migrate to CmpPredicate). This has uncovered a bug where
two icmp instructions differing in just the fact that one of them has
the samesign flag on it are hashed differently, leading to divergent
hashing, and a crash. Fix this crash by dropping samesign information on
icmp instructions before hashing them, preserving the former behavior.
Fixes#119893.
With the introduction of CmpPredicate in 51a895a (IR: introduce struct
with CmpInst::Predicate and samesign), PatternMatch is one of the first
key pieces of infrastructure that must be updated to match a CmpInst
respecting samesign information. Implement this change to Cmp-matchers.
This is a preparatory step in migrating the codebase over to
CmpPredicate. Since we no functional changes are desired at this stage,
we have chosen not to migrate CmpPredicate::operator==(CmpPredicate)
calls to use CmpPredicate::getMatching(), as that would have visible
impact on tests that are not yet written: instead, we call
CmpPredicate::operator==(Predicate), preserving the old behavior, while
also inserting a few FIXME comments for follow-ups.
Introduce llvm::CmpPredicate, an abstraction over a floating-point
predicate, and a pack of an integer predicate with samesign information,
in order to ease extending large portions of the codebase that take a
CmpInst::Predicate to respect the samesign flag.
We have chosen to demonstrate the utility of this new abstraction by
migrating parts of ValueTracking, InstructionSimplify, and InstCombine
from CmpInst::Predicate to llvm::CmpPredicate. There should be no
functional changes, as we don't perform any extra optimizations with
samesign in this patch, or use CmpPredicate::getMatching.
The design approach taken by this patch allows for unaudited callers of
APIs that take a llvm::CmpPredicate to silently drop the samesign
information; it does not pose a correctness issue, and allows us to
migrate the codebase piece-wise.
De-duplicate the functions getSignedPredicate and getUnsignedPredicate,
nearly identical versions of which were present in CmpInst and ICmpInst,
creating less confusion.
Add `ICmpInst::compare()` overload accepting `KnownBits`, similar to the
existing one accepting `APInt`. This is not directly part of KnownBits
(or APInt) for layering reasons.
Steal impliesEquivalanceIf{True,False} (sic) from GVN, and extend it for
floating-point constant vectors, and accounting for denormal values.
Since InstCombine also performs GVN-like replacements, introduce
CmpInst::isEquivalence, and remove the corresponding code in GVN, with
the intent of using it in more places.
The code in GVN also has a bad FIXME saying that the optimization may be
valid in the nsz case, but this is not the case.
Alive2 proof: https://alive2.llvm.org/ce/z/vEaK8M
This replaces some of the most frequent offenders of using a DenseMap that
cause a malloc, where the typical element-count is small enough to fit in
an initial stack allocation.
Most of these are fairly obvious, one to highlight is the collectOffset
method of GEP instructions: if there's a GEP, of course it's going to have
at least one offset, but every time we've called collectOffset we end up
calling malloc as well for the DenseMap in the MapVector.
In `User::operator new` a single allocation is created to store the
`User` object itself, "intrusive" operands or a pointer for "hung off"
operands, and the descriptor. After allocation, details about the layout
(number of operands, how the operands are stored, if there is a
descriptor) are stored in the `User` object by settings its fields. The
`Value` and `User` constructors are then very careful not to initialize
these fields so that the values set during allocation can be
subsequently read. However, when the `User` object is returned from
`operator new` [its value is technically "indeterminate" and so reading
a field without first initializing it is undefined behavior (and will be
erroneous in
C++26)](https://en.cppreference.com/w/cpp/language/default_initialization#Indeterminate_and_erroneous_values).
We discovered this issue when trying to build LLVM using MSVC's [`/sdl`
flag](https://learn.microsoft.com/en-us/cpp/build/reference/sdl-enable-additional-security-checks?view=msvc-170)
which clears class fields after allocation (the docs say that this
feature shouldn't be turned on for custom allocators and should only
clear pointers, but that doesn't seem to match the implementation).
MSVC's behavior both with and without the `/sdl` flag is standards
conforming since a program is supposed to initialize storage before
reading from it, thus the compiler implementation changing any values
will never be observed in a well-formed program. The standard also
provides no provisions for making storage bytes not indeterminate by
setting them during allocation or `operator new`.
The fix for this is to create a set of types that encode the layout and
provide these to both `operator new` and the constructor:
* The `AllocMarker` types are used to select which `operator new` to
use.
* `AllocMarker` can then be implicitly converted to a `AllocInfo` which
tells the constructor how the type was laid out.
These are the final few places in LLVM where we use instruction pointers
to identify the position that we're inserting something. We're trying to
get away from that with a view to deprecating those methods, thus use
iterators in all these places. I believe they're all debug-info safe.
The sketchiest part is the ExtractValueInst copy constructor, where we
cast nullptr to a BasicBlock pointer, so that we take the non-default
insert-into-no-block path for instruction insertion, instead of the
default nullptr-instruction path for UnaryInstruction. Such a hack is
necessary until we get rid of the instruction constructor entirely.
It is now translated to `<1 x i64>`, which allows the removal of a bunch
of special casing.
This _incompatibly_ changes the ABI of any LLVM IR function with
`x86_mmx` arguments or returns: instead of passing in mmx registers,
they will now be passed via integer registers. However, the real-world
incompatibility caused by this is expected to be minimal, because Clang
never uses the x86_mmx type -- it lowers `__m64` to either `<1 x i64>`
or `double`, depending on ABI.
This change does _not_ eliminate the SelectionDAG `MVT::x86mmx` type.
That type simply no longer corresponds to an IR type, and is used only
by MMX intrinsics and inline-asm operands.
Because SelectionDAGBuilder only knows how to generate the
operands/results of intrinsics based on the IR type, it thus now
generates the intrinsics with the type MVT::v1i64, instead of
MVT::x86mmx. We need to fix this before the DAG LegalizeTypes, and thus
have the X86 backend fix them up in DAGCombine. (This may be a
short-lived hack, if all the MMX intrinsics can be removed in upcoming
changes.)
Works towards issue #98272.
This is a helper to avoid writing `getModule()->getDataLayout()`. I
regularly try to use this method only to remember it doesn't exist...
`getModule()->getDataLayout()` is also a common (the most common?)
reason why code has to include the Module.h header.
This patch simplifies instruction creation by replacing all overloads of
instruction constructors/Create methods that are identical other than
the Instruction *InsertBefore/BasicBlock *InsertAtEnd/BasicBlock::iterator
InsertBefore argument with a single version that takes an InsertPosition
argument. The InsertPosition class can be implicitly constructed from
any of the above, internally converting them to the appropriate
BasicBlock::iterator value which can then be used to insert the
instruction (or to not insert it if an invalid iterator is passed).
The upshot of this is that code will be deduplicated, and all callsites
will switch to calling the new unified version without any changes
needed to make the compiler happy. There is at least one exception to
this; the construction of InsertPosition is a user-defined conversion,
so any caller that was already relying on a different user-defined
conversion won't work. In all of LLVM and Clang this happens exactly
once: at clang/lib/CodeGen/CGExpr.cpp:123 we try to construct an alloca
with an AssertingVH<Instruction> argument, which must now be cast to an
Instruction* by using `&*`. If this is more common elsewhere, it could
be fixed by adding an appropriate constructor to InsertPosition.
…f weights" #95136
Reverts #95060, and relands #86609, with the unintended code generation
changes addressed.
This patch implements the changes to LLVM IR discussed in
https://discourse.llvm.org/t/rfc-update-branch-weights-metadata-to-allow-tracking-branch-weight-origins/75032
In this patch, we add an optional field to MD_prof meatdata nodes for
branch weights, which can be used to distinguish weights added from
llvm.expect* intrinsics from those added via other methods, e.g. from
profiles or inserted by the compiler.
One of the major motivations, is for use with MisExpect diagnostics,
which need to know if branch_weight metadata originates from an
llvm.expect intrinsic. Without that information, we end up checking
branch weights multiple times in the case if ThinLTO + SampleProfiling,
leading to some inaccuracy in how we report MisExpect related
diagnostics to users.
Since we change the format of MD_prof metadata in a fundamental way, we
need to update code handling branch weights in a number of places.
We also update the lang ref for branch weights to reflect the change.
This patch implements the changes to LLVM IR discussed in
https://discourse.llvm.org/t/rfc-update-branch-weights-metadata-to-allow-tracking-branch-weight-origins/75032
In this patch, we add an optional field to MD_prof metadata nodes for
branch weights, which can be used to distinguish weights added from
`llvm.expect*` intrinsics from those added via other methods, e.g.
from profiles or inserted by the compiler.
One of the major motivations, is for use with MisExpect diagnostics,
which need to know if branch_weight metadata originates from an
llvm.expect intrinsic. Without that information, we end up checking
branch weights multiple times in the case if ThinLTO + SampleProfiling,
leading to some inaccuracy in how we report MisExpect related
diagnostics to users.
Since we change the format of MD_prof metadata in a fundamental way, we
need to update code handling branch weights in a number of places.
We also update the lang ref for branch weights to reflect the change.
This implements the `nusw` and `nuw` flags for `getelementptr` as
proposed at
https://discourse.llvm.org/t/rfc-add-nusw-and-nuw-flags-for-getelementptr/78672.
The three possible flags are encapsulated in the new `GEPNoWrapFlags`
class. Currently this class has a ctor from bool, interpreted as the
InBounds flag. This ctor should be removed in the future, as code gets
migrated to handle all flags.
There are a few places annotated with `TODO(gep_nowrap)`, where I've had
to touch code but opted to not infer or precisely preserve the new
flags, so as to keep this as NFC as possible and make sure any changes
of that kind get test coverage when they are made.
A related change is https://reviews.llvm.org/D133121, which correctly
preserves both branch weights and value profiles for invoke instruction.
* If the branch weight of the `invokeinst` specifies taken / not-taken branches, there is no scale.
This patch is moving out following intrinsics:
* vector.interleave2/deinterleave2
* vector.reverse
* vector.splice
from the experimental namespace.
All these intrinsics exist in LLVM for more than a year now, and are
widely used, so should not be considered as experimental.
Similar to #87934, this adds costs to the shuffles in a canonical LD3/LD4
pattern, which are represented in LLVM as deinterleaving-shuffle(load). This
likely has less effect at the moment than the ST3/ST4 costs as instcombine will
perform certain transforms without considering the cost.
- Put the helper function in `ProfDataUtil.h/cpp`, which is already a
dependency of `Instructions.cpp`
- The helper function could be re-used to update profiles of
`InvokeInst` (in a follow-up pull request)
I'd reverted this in 6c7805d5d1 after a bad stage. Original commit
messsage follows:
[NFC][RemoveDIs] Bulk update utilities to insert with iterators
As part of the RemoveDIs project we need LLVM to insert instructions using
iterators wherever possible, so that the iterators can carry a bit of
debug-info. This commit implements some of that by updating the contents of
llvm/lib/Transforms/Utils to always use iterator-versions of instruction
constructors.
There are two general flavours of update:
* Almost all call-sites just call getIterator on an instruction
* Several make use of an existing iterator (scenarios where the code is
actually significant for debug-info)
The underlying logic is that any call to getFirstInsertionPt or similar
APIs that identify the start of a block need to have that iterator passed
directly to the insertion function, without being converted to a bare
Instruction pointer along the way.
I've also switched DemotePHIToStack to take an optional iterator: it needs
to take an iterator, and having a no-insert-location behaviour appears to
be important. The constructors for ICmpInst and FCmpInst have been updated
too. They're the only instructions that take block _references_ rather than
pointers for certain calls, and a future patch is going to make use of
default-null block insertion locations.
All of this should be NFC.
As part of the RemoveDIs project we need LLVM to insert instructions using
iterators wherever possible, so that the iterators can carry a bit of
debug-info. This commit implements some of that by updating the contents of
llvm/lib/Transforms/Utils to always use iterator-versions of instruction
constructors.
There are two general flavours of update:
* Almost all call-sites just call getIterator on an instruction
* Several make use of an existing iterator (scenarios where the code is
actually significant for debug-info)
The underlying logic is that any call to getFirstInsertionPt or similar
APIs that identify the start of a block need to have that iterator passed
directly to the insertion function, without being converted to a bare
Instruction pointer along the way.
I've also switched DemotePHIToStack to take an optional iterator: it needs
to take an iterator, and having a no-insert-location behaviour appears to
be important. The constructors for ICmpInst and FCmpInst have been updated
too. They're the only instructions that take block _references_ rather than
pointers for certain calls, and a future patch is going to make use of
default-null block insertion locations.
All of this should be NFC.
In a previous commit I added declarations for all these functions, but
forgot to add bodies for them (as nothing uses them yet). These
iterator-taking constructors are necessary for the future where we only
use iterators for insertion, preserving some debug-info properties.
Also adds two extra declarations I missed in 76dd4bc036f
Removing debug-intrinsics requires that we always insert with an
iterator, not with an instruction position. To enforce that, we need to
eliminate the `Instruction *` taking functions. It's safe to leave the
insert-at-end-of-block functions as the intention is clear for debug
info purposes (i.e., insert after both instructions and debug-info at
the end of the function).
This patch demonstrates how that needs to happen. At a variety of
call-sites to the `CreateNeg` constructor we need to consider:
* Has this instruction been selected because of the operation it
performs? In that case, just call `getIterator` and pass an iterator in.
* Has this instruction been selected because of it's position? If so, we
need to keep the iterator identifying that position (see the 3rd hunk
changing Reassociate.cpp, although it's coincidentally not debug-info
significant).
This also demonstrates what we'll try and do with the constructor
methods going forwards: have one fully explicit set of parameters
including iterator, and another with default-arguments where the
block-to-insert-into argument defaults to nullptr / no-position,
creating an instruction that hasn't been inserted yet.
Part of removing debug-intrinsics from LLVM requires using iterators
whenever we insert an instruction into a block. That means we need all
instruction constructors and factory functions to have an iterator
taking option, which this patch adds.
The whole of this patch should be NFC: it's adding new flavours of
existing constructors, and plumbing those through to the Instruction
constructor that takes iterators. It's almost entirely boilerplate
copy-and-paste too.
This fixes the case where we would shrink an frem to half and then
bitcast to bfloat, producing invalid results. The transformation was
written under the assumption that there is only one type with a given
bit width.
Also add a strategic assert to CastInst::CreateFPCast to turn this
miscompilation into a crash.
Miscompilation arises due to instruction combining of cast pairs of the
type `bitcast bfloat to half` + `<FPOp> bfloat to half` or `bitcast half
to bfloat` + `<FPOp half to bfloat`. For example `bitcast bfloat to
half`+`fpext half to double` or `bitcast bfloat to half`+`fpext bfloat
to double` respectively reduce to `fpext bfloat to double` and `fpext
half to double`. This is an incorrect conversion as it assumes the
representation of `bfloat` and `half` are equivalent due to having the
same width. As a consequence miscompilation arises.
Fixes#61984
It's not enough to just make sure destination type is floating point,
because the following chain may be incorrectly optimized:
```LLVM
%trunc = fptrunc float %src to bfloat
%cast = bitcast bfloat %trunc to half
```
Before the fix, the instruction sequence mentioned above used to be
translated into single fptrunc instruction as follows:
```LLVM
%trunc = fptrunc float %src to half
```
Such transformation was semantically incorrect.
This Op<2> usage was missed in 1ee6ec2bf3, which replaced the third
shuffle operand with a vector of integer mask constants.
I noticed this when attempting to make changes to the layout of
llvm::Value.
Because we're storing some extra debug-info information in the iterator
class, we need to insert new LICM-created stores using such iterators.
Switch LICM to storing iterators instead of pointers when it promotes
variables in loops, add a test for the desired behaviour, and enable
RemoveDIs instrumentation on a variety of other LICM tests for good
measure.
(This would appear to be the only pass in LLVM that needs to store
iterators on the heap).
...behind an experimental CMAKE option that's off by default.
This patch adds a new ilist-iterator-like class that can carry two extra bits
as well as the usual node pointer. This is part of the project to remove
debug-intrinsics from LLVM: see the rationale here [0], they're needed to
signal whether a "position" in a BasicBlock includes any debug-info before or
after the iterator.
This entirely duplicates ilist_iterator, attempting re-use showed it to be a
false economy. It's enable-able through the existing ilist_node options
interface, hence a few sites where the instruction-list type needs to be
updated. The actual main feature, the extra bits in the class, aren't part of
the class unless the cmake flag is given: this is because there's a
compile-time cost associated with it, and I'd like to get everything in-tree
but off-by-default so that we can do proper comparisons.
Nothing actually makes use of this yet, but will do soon, see the Phab patch
stack.
[0] https://discourse.llvm.org/t/rfc-instruction-api-changes-needed-to-eliminate-debug-intrinsics-from-ir/68939
Differential Revision: https://reviews.llvm.org/D153777
Need to add NumSrcElts param to is..Mask functions in
ShuffleVectorInstruction class for better mask analysis. Mask.size() not
always matches the sizes of the permuted vector(s). Allows to better
estimate the cost in SLP and fix uses of the functions in other cases.
Differential Revision: https://reviews.llvm.org/D158449
Need to add NumSrcElts param to is..Mask functions in
ShuffleVectorInstruction class for better mask analysis. Mask.size() not
always matches the sizes of the permuted vector(s). Allows to better
estimate the cost in SLP and fix uses of the functions in other cases.
Differential Revision: https://reviews.llvm.org/D158449
Need to add NumSrcElts param to is..Mask functions in
ShuffleVectorInstruction class for better mask analysis. Mask.size() not
always matches the sizes of the permuted vector(s). Allows to better
estimate the cost in SLP and fix uses of the functions in other cases.
Differential Revision: https://reviews.llvm.org/D158449
