In this patch I have added support for a new loop hint called
vectorize.scalable.enable that says whether we should enable scalable
vectorization or not. If a user wants to instruct the compiler to
vectorize a loop with scalable vectors they can now do this as
follows:
br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !2
...
!2 = !{!2, !3, !4}
!3 = !{!"llvm.loop.vectorize.width", i32 8}
!4 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
Setting the hint to false simply reverts the behaviour back to the
default, using fixed width vectors.
Differential Revision: https://reviews.llvm.org/D88962
This is similar to the existing alloca and program address spaces (D37052)
and should be used when creating/accessing global variables.
We need this in our CHERI fork of LLVM to place all globals in address space 200.
This ensures that values are accessed using CHERI load/store instructions
instead of the normal MIPS/RISC-V ones.
The problem this is trying to fix is that most of the time the type of
globals is created using a simple PointerType::getUnqual() (or ::get() with
the default address-space value of 0). This does not work for us and we get
assertion/compilation/instruction selection failures whenever a new call
is added that uses the default value of zero.
In our fork we have removed the default parameter value of zero for most
address space arguments and use DL.getProgramAddressSpace() or
DL.getGlobalsAddressSpace() whenever possible. If this change is accepted,
I will upstream follow-up patches to use DL.getGlobalsAddressSpace() instead
of relying on the default value of 0 for PointerType::get(), etc.
This patch and the follow-up changes will not have any functional changes
for existing backends with the default globals address space of zero.
A follow-up commit will change the default globals address space for
AMDGPU to 1.
Reviewed By: dylanmckay
Differential Revision: https://reviews.llvm.org/D70947
The `dso_local_equivalent` constant is a wrapper for functions that represents a
value which is functionally equivalent to the global passed to this. That is, if
this accepts a function, calling this constant should have the same effects as
calling the function directly. This could be a direct reference to the function,
the `@plt` modifier on X86/AArch64, a thunk, or anything that's equivalent to the
resolved function as a call target.
When lowered, the returned address must have a constant offset at link time from
some other symbol defined within the same binary. The address of this value is
also insignificant. The name is leveraged from `dso_local` where use of a function
or variable is resolved to a symbol in the same linkage unit.
In this patch:
- Addition of `dso_local_equivalent` and handling it
- Update Constant::needsRelocation() to strip constant inbound GEPs and take
advantage of `dso_local_equivalent` for relative references
This is useful for the [Relative VTables C++ ABI](https://reviews.llvm.org/D72959)
which makes vtables readonly. This works by replacing the dynamic relocations for
function pointers in them with static relocations that represent the offset between
the vtable and virtual functions. If a function is externally defined,
`dso_local_equivalent` can be used as a generic wrapper for the function to still
allow for this static offset calculation to be done.
See [RFC](http://lists.llvm.org/pipermail/llvm-dev/2020-August/144469.html) for more details.
Differential Revision: https://reviews.llvm.org/D77248
Clarify the semantics of GEP inbounds, in particular with regard
to what it means for wrapping. This cleans up some confusion on
when it is legal to apply nuw/nsw flags to various parts of the
GEP calculation.
Differential Revision: https://reviews.llvm.org/D90708
This patch adds a new !annotation metadata kind which can be used to
attach annotation strings to instructions.
It also adds a new pass that emits summary remarks per function with the
counts for each annotation kind.
The intended uses cases for this new metadata is annotating
'interesting' instructions and the remarks should provide additional
insight into transformations applied to a program.
To motivate this, consider these specific questions we would like to get answered:
* How many stores added for automatic variable initialization remain after optimizations? Where are they?
* How many runtime checks inserted by a frontend could be eliminated? Where are the ones that did not get eliminated?
Discussed on llvm-dev as part of 'RFC: Combining Annotation Metadata and Remarks'
(http://lists.llvm.org/pipermail/llvm-dev/2020-November/146393.html)
Reviewed By: thegameg, jdoerfert
Differential Revision: https://reviews.llvm.org/D91188
This changes the definition of t2DoLoopStart from
t2DoLoopStart rGPR
to
GPRlr = t2DoLoopStart rGPR
This will hopefully mean that low overhead loops are more tied together,
and we can more reliably generate loops without reverting or being at
the whims of the register allocator.
This is a fairly simple change in itself, but leads to a number of other
required alterations.
- The hardware loop pass, if UsePhi is set, now generates loops of the
form:
%start = llvm.start.loop.iterations(%N)
loop:
%p = phi [%start], [%dec]
%dec = llvm.loop.decrement.reg(%p, 1)
%c = icmp ne %dec, 0
br %c, loop, exit
- For this a new llvm.start.loop.iterations intrinsic was added, identical
to llvm.set.loop.iterations but produces a value as seen above, gluing
the loop together more through def-use chains.
- This new instrinsic conceptually produces the same output as input,
which is taught to SCEV so that the checks in MVETailPredication are not
affected.
- Some minor changes are needed to the ARMLowOverheadLoop pass, but it has
been left mostly as before. We should now more reliably be able to tell
that the t2DoLoopStart is correct without having to prove it, but
t2WhileLoopStart and tail-predicated loops will remain the same.
- And all the tests have been updated. There are a lot of them!
This patch on it's own might cause more trouble that it helps, with more
tail-predicated loops being reverted, but some additional patches can
hopefully improve upon that to get to something that is better overall.
Differential Revision: https://reviews.llvm.org/D89881
This patch adds the llvm.loop.mustprogress loop metadata. This is to be
added to loops where the frontend language requires that the loop makes
observable interactions with the environment. This is the loop-level
equivalent to the function attribute `mustprogress` defined in D86233.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D88464
The neutral value is -0.0, not 0.0. This doesn't matter for "fast"
reductions due to nsz, but does matter for reassoc-only and seq
reductions.
Change tests to mostly use -0.0 where the neutral value was intended,
and add some additional test coverage in some places. Also update
LangRef to use the right value.
If `null_pointer_is_valid` is present, `dereferenceable` does not imply
`nonnull`, make it clear.
Came up in D17993.
Reviewed By: aqjune
Differential Revision: https://reviews.llvm.org/D89417
This change introduces a GC parseable lowering for element atomic
memcpy/memmove intrinsics. This way runtime can provide an
implementation which can take a safepoint during copy operation.
See "GC-parseable element atomic memcpy/memmove" thread on llvm-dev
for the background and details:
https://groups.google.com/g/llvm-dev/c/NnENHzmX-b8/m/3PyN8Y2pCAAJ
Differential Revision: https://reviews.llvm.org/D88861
It's currently ambiguous in IR whether the source language explicitly
did not want a stack a stack protector (in C, via function attribute
no_stack_protector) or doesn't care for any given function.
It's common for code that manipulates the stack via inline assembly or
that has to set up its own stack canary (such as the Linux kernel) would
like to avoid stack protectors in certain functions. In this case, we've
been bitten by numerous bugs where a callee with a stack protector is
inlined into an __attribute__((__no_stack_protector__)) caller, which
generally breaks the caller's assumptions about not having a stack
protector. LTO exacerbates the issue.
While developers can avoid this by putting all no_stack_protector
functions in one translation unit together and compiling those with
-fno-stack-protector, it's generally not very ergonomic or as
ergonomic as a function attribute, and still doesn't work for LTO. See also:
https://lore.kernel.org/linux-pm/20200915172658.1432732-1-rkir@google.com/https://lore.kernel.org/lkml/20200918201436.2932360-30-samitolvanen@google.com/T/#u
Typically, when inlining a callee into a caller, the caller will be
upgraded in its level of stack protection (see adjustCallerSSPLevel()).
By adding an explicit attribute in the IR when the function attribute is
used in the source language, we can now identify such cases and prevent
inlining. Block inlining when the callee and caller differ in the case that one
contains `nossp` when the other has `ssp`, `sspstrong`, or `sspreq`.
Fixes pr/47479.
Reviewed By: void
Differential Revision: https://reviews.llvm.org/D87956
LLVM IR currently assumes some form of forward progress. This form is
not explicitly defined anywhere, and is the cause of miscompilations
in most languages that are not C++11 or later. This implicit forward progress
guarantee can not be opted out of on a function level nor on a loop
level. Languages such as C (C11 and later), C++ (pre-C++11), and Rust
have different forward progress requirements and this needs to be
evident in the IR.
Specifically, C11 and onwards (6.8.5, Paragraph 6) states that "An
iteration statement whose controlling expression is not a constant
expression, that performs no input/output operations, does not access
volatile objects, and performs no synchronization or atomic operations
in its body, controlling expression, or (in the case of for statement)
its expression-3, may be assumed by the implementation to terminate."
C++11 and onwards does not have this assumption, and instead assumes
that every thread must make progress as defined in [intro.progress] when
it comes to scheduling.
This was initially brought up in [0] as a bug, a solution was presented
in [1] which is the current workaround, and the predecessor to this
change was [2].
After defining a notion of forward progress for IR, there are two
options to address this:
1) Set the default to assuming Forward Progress and provide an opt-out for functions and an opt-in for loops.
2) Set the default to not assuming Forward Progress and provide an opt-in for functions, and an opt-in for loops.
Option 2) has been selected because only C++11 and onwards have a
forward progress requirement and it makes sense for them to opt-into it
via the defined `mustprogress` function attribute. The `mustprogress`
function attribute indicates that the function is required to make
forward progress as defined. This is sharply in contrast to the status
quo where this is implicitly assumed. In addition, `willreturn` implies `mustprogress`.
The background for why this definition was chosen is in [3] and for why
the option was chosen is in [4] and the corresponding thread(s). The implementation is in D85393, the
clang patch is in D86841, the LoopDeletion patch is in D86844, the
Inliner patches are in D87180 and D87262, and there will be more
incoming.
[0] https://bugs.llvm.org/show_bug.cgi?id=965#c25
[1] https://lists.llvm.org/pipermail/llvm-dev/2017-October/118558.html
[2] https://reviews.llvm.org/D65718
[3] https://lists.llvm.org/pipermail/llvm-dev/2020-September/144919.html
[4] https://lists.llvm.org/pipermail/llvm-dev/2020-September/145023.html
Reviewed By: jdoerfert, efriedma, nikic
Differential Revision: https://reviews.llvm.org/D86233
The langref description for llvm.test.set.loop.iterations.* were
missing the i1 return type.
Differential Revision: https://reviews.llvm.org/D89564
Patch by: Janek van Oirschot
This patch adds metadata !noundef and makes load instructions can optionally have it.
A load with !noundef always return a well-defined value (has no undef bit or isn't poison).
If the loaded value isn't well defined, the behavior is undefined.
This metadata can be used to encode the assumption from C/C++ that certain reads of variables should have well-defined values.
It is helpful for optimizing freeze instructions away, because freeze can be removed when its operand has well-defined value, and showing that a load from arbitrary location is well-defined is usually hard otherwise.
The same information can be encoded with llvm.assume with operand bundle; using metadata is chosen because I wasn't sure whether code motion can be freely done when llvm.assume is inserted from clang instead.
The existing codebase already is stripping unknown metadata when doing code motion, so using metadata is UB-safe as well.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D89050
LLVM rejects DWARF operator DW_OP_over. This DWARF operator is needed
for Flang to support assumed rank array.
Summary:
Currently LLVM rejects DWARF operator DW_OP_over. Below error is
produced when llvm finds this operator.
[..]
invalid expression
!DIExpression(151, 20, 16, 48, 30, 35, 80, 34, 6)
warning: ignoring invalid debug info in over.ll
[..]
There were some parts missing in support of this operator, which are
now completed.
Testing
-added a unit testcase
-check-debuginfo
-check-llvm
Reviewed By: aprantl
Differential Revision: https://reviews.llvm.org/D89208
Add some minimal documentation for DILabel, originally introduced in
D45024. Update the name and semantics of the `variables:` field in the
documentation for `DISubprogram`; the field is now called
`retainedNodes:` and is a heterogeneous list of `DILocalVariable` and
`DILabel`.
Reviewed By: aprantl
Differential Revision: https://reviews.llvm.org/D89082
This patch adds support for DWARF attribute DW_AT_rank.
Summary:
Fortran assumed rank arrays have dynamic rank. DWARF attribute
DW_AT_rank is needed to support that.
Testing:
unit test cases added (hand-written)
check llvm
check debug-info
Reviewed By: aprantl
Differential Revision: https://reviews.llvm.org/D89141
Motivated by D88183, this seeks to clarify the current loop nomenclature with added illustrations, examples for possibly unexpected situations (infinite loops not part of the "parent" loop, logical loops sharing the same header, ...), and clarification on what other sources may consider a loop. The current document also has multiple errors that are fixed here.
Some selected errors:
* Loops a defined as strongly-connected components. A component a partition of all nodes, i.e. a subloop can never be a component. That is, the document as it currently is only covers top-level loops, even it also uses the term SCC for subloops.
* "a block can be the header of two separate loops at the same time" (it is considered a single loop by LoopInfo)
* "execute before some interesting event happens" (some interesting event is not well-defined)
Reviewed By: baziotis, Whitney
Differential Revision: https://reviews.llvm.org/D88408
This reverts commit 55c4ff91bd.
Issues were introduced as discussed in https://reviews.llvm.org/D88241
where this change made previous bugs in the linker and BitCodeWriter
visible.
This is a patch to LangRef that clarifies the behavior of load/store/memset/memcpy/memmove when the pointers or sizes are not well-defined
as well.
MSan detects a case when e.g., only lower bits of address are garbage when `-msan-check-access-address` is enabled, and it does not directly conflict with this patch because a C program should not use a pointer with undef bits and reasonable optimizations do not convert a well-defined pointer into a pointer with undef bits.
This patch contains a definition of a well-defined value as well.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D87994
Make the corresponding change that was made for byval in
b7141207a4. Like byval, this requires a
bulk update of the test IR tests to include the type before this can
be mandatory.
As discussed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2020-April/140729.html
This is hopefully the final remaining showstopper before we can remove
the 'experimental' from the reduction intrinsics.
No behavior was specified for the FP min/max reductions, so we have a
mess of different interpretations.
There are a few potential options for the semantics of these max/min ops.
I think this is the simplest based on current behavior/implementation:
make the reductions inherit from the existing llvm.maxnum/minnum intrinsics.
These correspond to libm fmax/fmin, and those are similar to the (now
deprecated?) IEEE-754 maxNum/minNum functions (NaNs are treated as missing
data). So the default expansion creates calls to libm functions.
Another option would be to inherit from llvm.maximum/minimum (NaNs propagate),
but most targets just crash in codegen when given those nodes because no
default expansion was ever implemented AFAICT.
We could also just assume 'nnan' semantics by default (we are already
assuming 'nsz' semantics in the maxnum/minnum intrinsics), but some targets
(AArch64, PowerPC) support the more defined behavior, so it doesn't make much
sense to not allow a tighter spec. Fast-math-flags (nnan) can be used to
loosen the semantics.
(Note that D67507 was proposed to update the LangRef to acknowledge the more
recent IEEE-754 2019 standard, but that patch seems to have stalled. If we do
update based on the new standard, the reduction instructions can seamlessly
inherit from whatever updates are made to the max/min intrinsics.)
x86 sees a regression here on 'nnan' tests because we have underlying,
longstanding bugs in FMF creation/propagation. Those need to be fixed apart
from this change (for example: https://llvm.org/PR35538). The expansion
sequence before this patch may not have been correct.
Differential Revision: https://reviews.llvm.org/D87391
This adjusts the description of `llvm.memcpy` to also allow operands
to be equal. This is in line with what Clang currently expects.
This change is intended to be temporary and followed by re-introduce
a variant with the non-overlapping guarantee for cases where we can
actually ensure that property in the front-end.
See the links below for more details:
http://lists.llvm.org/pipermail/cfe-dev/2020-August/066614.html
and PR11763.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D86815
The wording before this patch applies to llvm.mem.parallel_loop_access, not access groups.
Reviewed By: mppf, hfinkel
Differential Revision: https://reviews.llvm.org/D83781
This patch makes LangRef be explicit about the value of padding when storing an aggregate.
It states that when an aggregate is stored into memory, padding is filled with undef.
Here is a clue that supports this change (edited to reflect the discussion from llvm-dev):
- IPSCCP ignores padding and directly stores a constant aggregate if possible. It loses the data stored in the padding. https://godbolt.org/z/xzenYs Memcpyopt ignores (the preexisting value of) padding when copying an aggregate or storing a constant: https://godbolt.org/z/hY6ndd / https://godbolt.org/z/3WMP5a
The two items below are not relevant with this patch because Clang lowers load/store of individual field of struct into load/stores of the corresponding pointer with a primitive type. Also, when copy is needed, it uses memcpy instead of load/store of an aggregate, as discussed in the llvm-dev. However, this patch is still valid (as discussed) because it is needed to explain the two optimizations above.
- According to C17, the value of padding bytes when storing values in structures or unions is unspecified.
- I updated Alive2 and it did not find any problematic transformation from LLVM unit tests and while running translation validation of a few C programs.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D86189
We had already specified that second argument `n` of this intrinsic is `n > 0`,
but now add to this that the result is a poison value if this is not the case.
Differential Revision: https://reviews.llvm.org/D86637
According to the current LangRef, Memset/memcpy/memmove can take a
null/dangling pointer if the size is zero.
(Relevant thread: http://lists.llvm.org/pipermail/llvm-dev/2017-July/115665.html )
This patch expands it and allows the functions to take undef/poison pointers
too.
This required the updates in the align attribute since it isn't specified
what is the alignment of undef/poison pointers.
This patch states that their alignment is 1.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D86643
A first version of get.active.lane.mask was committed in rG7fb8a40e5220. One of
the main purposes and uses of this intrinsic is to communicate information from
the middle-end to the back-end, but its current definition and semantics make
this actually very difficult. The intrinsic was defined as:
@llvm.get.active.lane.mask(%IV, %BTC)
where %BTC is the Backedge-Taken Count (variable names are different in the
LangRef spec). This allows to implicitly communicate the loop tripcount, which
can be reconstructed by calculating BTC + 1. But it has been very difficult to
prove that calculating BTC + 1 is safe and doesn't overflow. We need
complicated range and SCEV analysis, and thus the problem is that this
intrinsic isn't really doing what it was supposed to solve. Examples of the
overflow checks that are required in the (ARM) back-end are D79175 and D86074,
which aren't even complete/correct yet.
To solve this problem, we are revising the definitions/semantics for
get.active.lane.mask to avoid all the complicated overflow analysis. This means
that instead of communicating the BTC, we are now using the loop tripcount. Now
using LangRef's variable names, its semantics is changed from:
icmp ule (%base + i), %n
to:
icmp ult (%base + i), %n
with %n > 0 and corresponding to the loop tripcount. The intrinsic signature
remains the same.
Differential Revision: https://reviews.llvm.org/D86147
(Forgot to land this a couple of weeks back.)
In a recent series of changes, I've introduced support for using the respective operand bundle kinds on the statepoint. At the moment, code supports either/or, but there's no need to keep the old support around. For the moment, I am simply changing the specification and verifier to require zero length argument sets in the intrinsic.
The intrinsic itself is experimental. Given that, there's no forward serialization needed. The in tree uses and generation have already been updated to use the new operand bundle based forms, the only folks broken by the change will be those with frontends generating statepoints directly and the updates should be easy.
Why not go ahead and just remove the arguments entirely? Well, I plan to. But while working on this I've found that almost all of the arguments to the statepoint can be expressed via operand bundles or attributes. Given that, I'm planning a radical simplification of the arguments and figured I'd do one update not several small ones.
Differential Revision: https://reviews.llvm.org/D80892
Summary:
This support is needed for the Fortran array variables with pointer/allocatable
attribute. This support enables debugger to identify the status of variable
whether that is currently allocated/associated.
for pointer array (before allocation/association)
without DW_AT_associated
(gdb) pt ptr
type = integer (140737345375288:140737354129776)
(gdb) p ptr
value requires 35017956 bytes, which is more than max-value-size
with DW_AT_associated
(gdb) pt ptr
type = integer (:)
(gdb) p ptr
$1 = <not associated>
for allocatable array (before allocation)
without DW_AT_allocated
(gdb) pt arr
type = integer (140737345375288:140737354129776)
(gdb) p arr
value requires 35017956 bytes, which is more than max-value-size
with DW_AT_allocated
(gdb) pt arr
type = integer, allocatable (:)
(gdb) p arr
$1 = <not allocated>
Testing
- unit test cases added
- check-llvm
- check-debuginfo
Reviewed By: aprantl
Differential Revision: https://reviews.llvm.org/D83544
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.
David Sherwood