When translating offset info from the callee at a call site, first check if the
offset is Unknown. Any offset in the caller should be added only if the callee
offset is valid.
Differential Revision: https://reviews.llvm.org/D137011
Loop Fusion (Function Pass) requires loops in simplified form. With
legacy-pm, loop-simplify pass is added as a dependency for loop-fusion.
But the new pass manager does not always ensure this format. This patch
tries to invoke simplifyLoop() on loops that are not in simplified form
only for new PM.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D136781
The existing way of creating the predicate in the guard blocks uses
a boolean value per outgoing block. This increases the number of live
booleans as the number of outgoing blocks increases. The new way added
in this change is to store one integer to represent the outgoing block
we want to branch to, then at each guard block, an integer equality
check is performed to decide which a specific outgoing block is taken.
Using an integer reduces the number of live values and decreases
register pressure especially in cases where there are a large number
of outgoing blocks. The integer based approach is used when the
number of outgoing blocks crosses a threshold, which is currently set
to 32.
Patch by Ruiling Song.
Differential review: https://reviews.llvm.org/D127831
This is a sibling to:
6064e92b0a84
...but we canonicalize the shl+add to shl+xor,
so the pattern is different than I expected:
https://alive2.llvm.org/ce/z/8CX16e
I have not found any patterns that are safe
to propagate no-wrap, so that is not included
here.
This adds the default attributes (nocallback, nosync, nofree,
willreturn) to some X86 intrinsics. This will be needed to avoid
optimization regressions in the future (once we remove the
readonly -> willreturn implication for intrinsics).
Due to the number of intrinsics, this patch focuses just on the
IntrNoMem intrinsics up to the AVX2 section.
Differential Revision: https://reviews.llvm.org/D136939
Currently, InstCombine can elide a memcpy from a constant to a local alloca if
that alloca is passed as a nocapture parameter to a *function* that's readnone
or readonly, but it can't forward the memcpy if the *argument* is marked
readonly nocapture, even though readonly guarantees that the callee won't
mutate the pointee through that pointer. This patch adds support for detecting
and handling such situations, which arise relatively frequently in Rust, a
frontend that liberally emits readonly.
A more general version of this optimization would use alias analysis to check
the call's ModRef info for the pointee, but I was concerned about blowing up
compile time, so for now I'm just checking for one of readnone on the function,
readonly on the function, or readonly on the parameter.
Differential Revision: https://reviews.llvm.org/D136822
This extends the computeKnownBits analysis to support scalable vectors. The critical detail is in deciding how to represent the demanded elements of a vector whose length is unknown at compile time.
For this patch, I adopt the convention that we track one bit which corresponds to all lanes. That is, that bit is implicitly broadcast to all lanes of the scalable vector resulting in all lanes being demanded. This is the same convention we use in getSplatValue in SelectionDAG.
Note that this convention doesn't actually impact much. Most of the code is agnostic to the interpretation of the demanded elements, and the few cases which actually care need case by case handling anyways. In this patch, I just bail out of those cases.
A prior patch (D128159) proposed using a different convention in SDAG. I don't see any strong reason to prefer one scheme over the other, so I propose we go with this one as it's conceptually the simplest. Getting known and demanded bit optimizations unblocked at all is a significant win.
I've locally implemented this scheme in reasonable large parts of ValueTracking.cpp and SelectionDAG equivalents, and have not hit any blockers. If this is approved, I plan to post a series of patches plumbing this through all the relevant parts.
In the discussion on that patch, a preference was expressed for introducing some form of abstraction around the demanded elements. I'll note that I've played with several variations on that idea locally, and have yet to find anything which results in more readable code. If anyone has concrete ideas in this area, I'm happy to explore in follow up patches. I'd strongly prefer to be making API changes in NFC manner with tests in place.
Differential Revision: https://reviews.llvm.org/D136470
This commit adds tests to Transforms/InstCombine/memcpy-from-global.ll that
test various situations involving memcpy from a constant to an alloca that is
then passed to function parameters with various attributes. The forthcoming
D136822 allows InstCombine to remove these memcpys if they're passed to a
single readonly nocapture parameter.
Differential Revision: https://reviews.llvm.org/D137033
InstCombine can replace memcpy to an alloca with a pointer directly to the
source in certain cases. Unfortunately, it also did so for volatile memcpys.
This patch makes it stop doing that.
This was discovered in D136822.
Differential Revision: https://reviews.llvm.org/D137031
D136822 demonstrated that we currently delete volatile memcpys in InstCombine,
which we shouldn't do. This commit adds a test for this. A forthcoming commit
will fix it.
Differential Revision: https://reviews.llvm.org/D137029
This commit adds some tests in preparation for D136659, which allows alias
analysis to treat locally-invariant memory pointed to by readonly noalias
pointers the same as globally-invariant memory in some cases. The existing
behavior for these tests is marked as expected and will be changed when that
diff lands.
Differential Revision: https://reviews.llvm.org/D136993
In D136659 I found a few tests that write through readonly parameters:
* Analysis/BasicAA/pr18573.ll: @foo1 writes through %arr.ptr, but declares it
readonly. I removed the readonly annotation.
* CodeGen/ARM/ParallelDSP/aliasing.ll: @restrict writes through the readonly
%arg3, @store_alias_arg3_illegal_1 writes through the readonly %arg3, and
@store_alias_arg3_illegal_2 writes through the readonly %arg3. I removed
readonly from all three. Also, I added some CHECK-LABEL directives to make it
harder for FileCheck output to be mixed up.
* Transforms/LoopVectorize/AArch64/sve-gather-scatter.ll:
@gather_nxv4i32_ind64_stride2 writes through the readonly %a. I removed the
readonly attribute.
* Transforms/LoopVectorize/interleaved-accesses.ll: @load_gap_reverse writes
through the readonly %P1 and %P2. Also, the corresponding C code in the comment
didn't match the test. I removed the readonly attribute from both parameters
and corrected the C code.
Differential Revision: https://reviews.llvm.org/D136880
Replace custom code to check if only the first lane is used by generic
helper `onlyFirstLaneUsed`. This enables VPlan-based sinking in a few
additional cases and was suggested in D133760.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D136368
X * ((1 << Z) + 1) --> (X << Z) + X
https://alive2.llvm.org/ce/z/P-7WK9
It's possible that we could do better with propagating
no-wrap, but this carries over the existing logic and
appears to be correct.
The naming differences on the existing folds are a result
of using getName() to set the final value via Builder.
That makes it easier to transfer no-wrap rather than the
gymnastics required from the raw create instruction APIs.
In InstCombine we treat i8/i16 as desirable, even if they are not legal.
The current logic in shouldChangeType will decide to convert from an
illegal but desirable type (such as an i8) to an illegal and undesirable
type (such as i3). This patch prevents changing the switch conditions to
an irregular type on like Arm/AArch64 where i8/i16 are not legal.
This is the same issue as https://reviews.llvm.org/D54115. In the case I
was looking it is was converting an i32 switch to an i8 switch, which
then became a i3 switch.
Differential Revision: https://reviews.llvm.org/D136763
Currently pseudo probe encoding for a function is like:
- For the first probe, a relocation from it to its physical position in the code body
- For subsequent probes, an incremental offset from the current probe to the previous probe
The relocation could potentially cause relocation overflow during link time. I'm now replacing it with an offset from the first probe to the function start address.
A source function could be lowered into multiple binary functions due to outlining (e.g, coro-split). Since those binary function have independent link-time layout, to really avoid relocations from .pseudo_probe sections to .text sections, the offset to replace with should really be the offset from the probe's enclosing binary function, rather than from the entry of the source function. This requires some changes to previous section-based emission scheme which now switches to be function-based. The assembly form of pseudo probe directive is also changed correspondingly, i.e, reflecting the binary function name.
Most of the source functions end up with only one binary function. For those don't, a sentinel probe is emitted for each of the binary functions with a different name from the source. The sentinel probe indicates the binary function name to differentiate subsequent probes from the ones from a different binary function. For examples, given source function
```
Foo() {
…
Probe 1
…
Probe 2
}
```
If it is transformed into two binary functions:
```
Foo:
…
Foo.outlined:
…
```
The encoding for the two binary functions will be separate:
```
GUID of Foo
Probe 1
GUID of Foo
Sentinel probe of Foo.outlined
Probe 2
```
Then probe1 will be decoded against binary `Foo`'s address, and Probe 2 will be decoded against `Foo.outlined`. The sentinel probe of `Foo.outlined` makes sure there's not accidental relocation from `Foo.outlined`'s probes to `Foo`'s entry address.
On the BOLT side, to be minimal intrusive, the pseudo probe re-encoding sticks with the old encoding format. This is fine since unlike linker, Bolt processes the pseudo probe section as a whole and it is free from relocation overflow issues.
The change is downwards compatible as long as there's no mixed use of the old encoding and the new encoding.
Reviewed By: wenlei, maksfb
Differential Revision: https://reviews.llvm.org/D135912
Differential Revision: https://reviews.llvm.org/D135914
Differential Revision: https://reviews.llvm.org/D136394
The LSR may suggest less profitable transformation to the loop. This
patch adds check to prevent LSR from generating worse code than what
we already have.
Since LSR affects nearly all targets, the patch is guarded by the
option 'lsr-drop-solution' and default as disable for now.
The next step should be extending an TTI interface to allow target(s)
to enable this enhancememnt.
Debug log is added to remind user of such choice to skip the LSR
solution.
Reviewed By: Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D126043
When calculating the specialization bonus for a given function argument,
we recursively traverse the chain of (certain) users, accumulating the
instruction costs. Then we exponentially increase the bonus to account
for loop nests. This is problematic for two reasons: (a) the users might
not themselves be inside the loop nest, (b) if they are we are accounting
for it multiple times. Instead we should be adjusting the bonus before
traversing the user chain.
This reduces the instruction count for CTMark (newPM-O3) when Function
Specialization is enabled without actually reducing the amount of
specializations performed (geomean: -0.001% non-LTO, -0.406% LTO).
Differential Revision: https://reviews.llvm.org/D136692
This is copying the code that was added for 'add' with D130075.
(That patch removed a fallthrough in the cases, but we can
probably still share at least some code again as a follow-up
cleanup, but I didn't want to risk it here.)
The reasoning is similar to the carry propagation for 'add':
if we don't demand low bits of the subtraction and the
subtrahend (aka RHS or operand 1) is known zero in those low
bits, then there can't be any borrowing required from the
higher bits of operand 0, so the low bits don't matter.
Also, the no-wrap flags can be propagated (and I think that
should be true for add too).
Here's an attempt to prove that in Alive2:
https://alive2.llvm.org/ce/z/xqh7Pa
(can add nsw or nuw to src and tgt, and it should still pass)
Differential Revision: https://reviews.llvm.org/D136788
[fixed test to work with reverse iteration]
The `FunctionSpecialization` pass has support for specialising
functions, which are called with literal arguments. This functionality
is disabled by default and is enabled with the option
`-function-specialization-for-literal-constant` . There are a few
issues with the implementation, though:
* even with the default, the pass will still specialise based on
floating-point literals
* even when it's enabled, the pass will specialise only for the `i1`
type (or `i2` if all of the possible 4 values occur, or `i3` if all
of the possible 8 values occur, etc)
The reason for this is incorrect check of the lattice value of the
function formal parameter. The lattice value is `overdefined` when the
constant range of the possible arguments is the full set, and this is
the reason for the specialisation to trigger. However, if the set of
the possible arguments is not the full set, that must not prevent the
specialisation.
This patch changes the pass to NOT consider a formal parameter when
specialising a function if the lattice value for that parameter is:
* unknown or undef
* a constant
* a constant range with a single element
on the basis that specialisation is pointless for those cases.
Is also changes the criteria for picking up an actual argument to
specialise if the argument is:
* a LLVM IR constant
* has `constant` lattice value
has `constantrange` lattice value with a single element.
Reviewed By: ChuanqiXu
Differential Revision: https://reviews.llvm.org/D135893
Change-Id: Iea273423176082ec51339aa66a5fe9fea83557ee
When a profile is stale and profile mismatch could happen, the mismatched samples are discarded, so we'd like to compute the mismatch metrics to quantify how stale the profile is, which will suggest user to refresh the profile if the number is high.
Two sets of metrics are introduced here:
- (Num_of_mismatched_funchash/Total_profiled_funchash), (Samples_of_mismached_func_hash / Samples_of_profiled_function) : Here it leverages the FunctionSamples's checksums attribute which is a feature of pseudo probe. When the source code CFG changes, the function checksums will be different, later sample loader will discard the whole functions' samples, this metrics can show the percentage of samples are discarded due to this.
- (Num_of_mismatched_callsite/Total_profiled_callsite), (Samples_of_mismached_callsite / Samples_of_profiled_callsite) : This shows how many mismatching for the callsite location as callsite location mismatch will affect the inlining which is highly correlated with the performance. It goes through all the callsite location in the IR and profile, use the call target name to match, report the num of samples in the profile that doesn't match a IR callsite.
This is implemented in a new class(SampleProfileMatcher) and under a switch("--report-profile-staleness"), we plan to extend it with a fuzzy profile matching feature in the future.
Reviewed By: hoy, wenlei, davidxl
Differential Revision: https://reviews.llvm.org/D136627
If we run LTO optimization we migth end up introducing a custom state machine
and later transforming the region into SPMD. This is a problem. While a follow
up will introduce a check for the SPMD conversion, this already prevents the
eager custom state machine generation. Only if the kernel init function is
defined, rather then declared, we will emit a custom state machine. SPMD-zation
can happen eagerly though. Tests are adjusted via a weak definition. The LTO
test was added to verify this works as expected.
Differential Revision: https://reviews.llvm.org/D136740
This was reverted because it was breaking when targeting Darwin which
tried to export these symbols which are now hidden. It should be safe
to just stop attempting to export these symbols in the clang driver,
though Apple folks will need to change their TAPI allow list described
in the commit where these symbols were originally exported
f538018562
Then reverted again because it broke tests on MacOS, they should be
fixed now.
Bug: https://github.com/llvm/llvm-project/issues/58265
Differential Revision: https://reviews.llvm.org/D135340
Increase test coverage - check that functions are not specialised on
constant or unused arguments.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D136184
When rewriting the call sites to call the new specialised functions, a
single call site can be matched by two different specialisations - a
"less specialised" version of the function and a "more specialised"
version of the function, e.g. for a function
void f(int x, int y)
the call like `f(1, 2)` could be matched by either
void f.1(int x /* int y == 2 */);
or
void f.2(/* int x == 1, int y == 2 */);
The `FunctionSpecialisation` pass tries to match specialisation in the
order of decreasing gain, so "more specialised" functions are
preferred to "less specialised" functions. This breaks, however, when
using the flag `-force-function-specialization`, in which case the
cost/benefit analysis is not performed and all the specialisations are
equally preferable.
This patch makes the pass calculate specialisation gain and order the
specialisations accordingly even when `-force-function-specialization`
is used, under the assumption that this flag has purely debugging
purpose and it is reasonable to ignore the extra computing effort it
incurs.
Reviewed By: ChuanqiXu, labrinea
Differential Revision: https://reviews.llvm.org/D136180
The `FunctionSpecialization` pass has support for specialising
functions, which are called with literal arguments. This functionality
is disabled by default and is enabled with the option
`-function-specialization-for-literal-constant` . There are a few
issues with the implementation, though:
* even with the default, the pass will still specialise based on
floating-point literals
* even when it's enabled, the pass will specialise only for the `i1`
type (or `i2` if all of the possible 4 values occur, or `i3` if all
of the possible 8 values occur, etc)
The reason for this is incorrect check of the lattice value of the
function formal parameter. The lattice value is `overdefined` when the
constant range of the possible arguments is the full set, and this is
the reason for the specialisation to trigger. However, if the set of
the possible arguments is not the full set, that must not prevent the
specialisation.
This patch changes the pass to NOT consider a formal parameter when
specialising a function if the lattice value for that parameter is:
* unknown or undef
* a constant
* a constant range with a single element
on the basis that specialisation is pointless for those cases.
Is also changes the criteria for picking up an actual argument to
specialise if the argument is:
* a LLVM IR constant
* has `constant` lattice value
has `constantrange` lattice value with a single element.
Reviewed By: ChuanqiXu
Differential Revision: https://reviews.llvm.org/D135893
When collecting the possible constant arguments to
specialise a function the compiler will abandon the search
on the first argument that is for some reason unsuitable as
a specialisation constant. Thus, depending on the traversal
order of the functions and call sites, the compiler can end
up with a different set of possible constants, hence with
different set of specialisations.
With this patch, the compiler will skip unsuitable
constants, but nevertheless will continue searching for
more.
Reviewed By: ChuanqiXu
Differential Revision: https://reviews.llvm.org/D135867
Epilogue loop vectorization is a feature in the vectorize intended to avoid running fully scalar code when the vector length of the main loop turns out to be either longer than the trip count of the actual loop, or with a huge remainder.
In practice, this feature appears to not have been well tuned. I honestly don't think it should be on by default at all, but it definitely shouldn't be on for RISCV. Note that other targets have also disabled it, but they've done so via disabling interleaving - which is, well, completely unrelated - and we don't want to do that for RISCV.
In the near term, many examples I'm seeing have terrible codegen for epilogue vectorization. We are greatly increasing code size for little value at reasonable VLEN values for small types. In the long term, the cases that epilogue vectorization are intended to handle are likely better handled via tail folding on RISCV.
As an aside, I also don't really trust the correctness of epilogue vectorization. The code structure is such that otherwise straight forward changes sometimes break only epilogue vectorization. The reuse of an existing vplan without careful validation opens significant room for nasty bugs. Given how rarely the code is exercised, that is not a good combination.
As such, this patch introduces a TTI hook, and completely disables epilogue vectorization on RISCV.
Differential Revision: https://reviews.llvm.org/D136695
