Make getPointersDiff() and sortPtrAccesses() compatible with opaque
pointers by explicitly passing in the element type instead of
determining it from the pointer element type.
The SLPVectorizer result is slightly non-optimal in that unnecessary
pointer bitcasts are added.
Differential Revision: https://reviews.llvm.org/D104784
This borrows as much as possible from the SDAG version of the code
(originally added with D27129 and since updated with big endian support).
In IR, we can test more easily for correctness than we did in the
original patch. I'm using the simplest cases that I could find for
InstSimplify: we computeKnownBits on variable shift amounts to see if
they are zero or in range. So shuffle constant elements into a vector,
cast it, and shift it.
The motivating x86 example from https://llvm.org/PR50123 is also here.
We computeKnownBits in the caller code, but we only check if the shift
amount is in range. That could be enhanced to catch the 2nd x86 test -
if the shift amount is known too big, the result is 0.
Alive2 understands the datalayout and agrees that the tests here are
correct - example:
https://alive2.llvm.org/ce/z/KZJFMZ
Differential Revision: https://reviews.llvm.org/D104472
v6m cores only have a limited number of registers available. Unrolling
can mean we spend more on stack spills and reloads than we save from the
unrolling. This patch adds an extra heuristic to put a limit on the
unroll count for loops with multiple live out values, as measured from
the LCSSA phi nodes.
Differential Revision: https://reviews.llvm.org/D104659
If a ctlz operation is performed on higher datatype and then
downcasted, then this can be optimized by doing a ctlz operation
on a lower datatype and adding the difference bitsize to the result
of ctlz to provide the same output:
https://alive2.llvm.org/ce/z/8uup9M
The original problem is shown in
https://llvm.org/PR50173
Differential Revision: https://reviews.llvm.org/D103788
This is part of improving floating-point patterns seen in:
https://llvm.org/PR39480
We don't require any FMF because the 2 potential corner cases
(-0.0 and NaN) are correctly handled without FMF:
1. -0.0 is treated as strictly less than +0.0 with
maximum/minimum, so fabs/fneg work as expected.
2. +/- 0.0 with maxnum/minnum is indeterminate, so
transforming to fabs/fneg is more defined.
3. The sign of a NaN may be altered by this transform,
but that is allowed in the default FP environment.
If there are FMF, they are propagated from the min/max call to
one or both new operands which seems to agree with Alive2:
https://alive2.llvm.org/ce/z/bem_xC
This changes the approach taken to tail-merge the blocks
to always create a new block instead of trying to reuse some block,
and generalizes it to support dealing not with just the `ret` in the future.
This effectively lifts the CallBr restriction, although this isn't really intentional.
That is the only non-NFC change here, i'm not sure if it's reasonable/feasible to temporarily retain it.
Other restrictions of the transform remain.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D104598
This adds more poison folding optimizations to InstSimplify.
Since all binary operators propagate poison, these are fine.
Also, the precondition of `select cond, undef, x` -> `x` is relaxed to allow the case when `x` is undef.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D104661
With regards to overrunning, the langref (llvm/docs/LangRef.rst)
specifies:
(llvm.experimental.vector.insert)
Elements ``idx`` through (``idx`` + num_elements(``subvec``) - 1)
must be valid ``vec`` indices. If this condition cannot be determined
statically but is false at runtime, then the result vector is
undefined.
(llvm.experimental.vector.extract)
Elements ``idx`` through (``idx`` + num_elements(result_type) - 1)
must be valid vector indices. If this condition cannot be determined
statically but is false at runtime, then the result vector is
undefined.
For the non-mixed cases (e.g. inserting/extracting a scalable into/from
another scalable, or inserting/extracting a fixed into/from another
fixed), it is possible to statically check whether or not the above
conditions are met. This was previously missing from the verifier, and
if the conditions were found to be false, the result of the
insertion/extraction would be replaced with an undef.
With regards to invalid indices, the langref (llvm/docs/LangRef.rst)
specifies:
(llvm.experimental.vector.insert)
``idx`` represents the starting element number at which ``subvec``
will be inserted. ``idx`` must be a constant multiple of
``subvec``'s known minimum vector length.
(llvm.experimental.vector.extract)
The ``idx`` specifies the starting element number within ``vec``
from which a subvector is extracted. ``idx`` must be a constant
multiple of the known-minimum vector length of the result type.
Similarly, these conditions were not previously enforced in the
verifier. In some circumstances, invalid indices were permitted
silently, and in other circumstances, an undef was spawned where a
verifier error would have been preferred.
This commit adds verifier checks to enforce the constraints above.
Differential Revision: https://reviews.llvm.org/D104468
Follow-up on Roman's idea expressed in D103959.
- If a Phi has undefined inputs from live blocks:
- and no other inputs, assume it is undef itself;
- and exactly one non-undef input, we can assume that all undefs are equal to this input.
Differential Revision: https://reviews.llvm.org/D104618
Reviewed By: lebedev.ri, nikic
Zero factor leads to division by zero and failure of corresponding
assert as shown in PR50765. We should filter out such factors.
Differential Revision: https://reviews.llvm.org/D104702
Reviewed By: huihuiz, reames
After landing the globalization optimizations, the precense of globalization on
the device that was not put in shared or stack memory is a failed optimization
with performance consequences so it should indicate a missed remark.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104735
When the load type is changed to ptr, we need the load pointer type
to also be ptr, because it's not allowed to create a pointer to an
opaque pointer. This is achieved by adjusting the getPointerTo() API
to return an opaque pointer for an opaque pointer base type.
Differential Revision: https://reviews.llvm.org/D104718
Summary:
This patch adds support for the Attributor to emit remarks on behalf of some
other pass. The attributor can now optionally take a callback function that
returns an OptimizationRemarkEmitter object when given a Function pointer. If
this is availible then a remark will be emitted for the corresponding pass
name.
Depends on D102197
Reviewed By: sstefan1 thegameg
Differential Revision: https://reviews.llvm.org/D102444
Summary:
The changes to globalization introduced in D97680 introduce a large amount of overhead by default. The old globalization method would always ignore globalization code if executing in SPMD mode. This wasn't strictly correct as data sharing is still possible in SPMD mode. The new interface is correct but introduces globalization code even when unnecessary. This optimization will use the existing HeapToStack transformation in the attributor to allow for unneeded globalization to be replaced with thread-private stack memory. This is done using the newly introduced library instances for the RTL functions added in D102087.
Depends on D97818
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102197
Summary:
Currently the attributor needs to give up if a function has external linkage.
This means that the optimization introduced in D97818 will only apply to static
functions. This change uses the Attributor to internalize OpenMP device
routines by making a copy of each function with private linkage and replacing
the uses in the module with it. This allows for the optimization to be applied
to any regular function.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102824
Summary:
The changes introduced in D97680 create a simpler interface to code that needs
to be globalized. This interface is used to simplify the globalization calls in
the middle end. We can check any globalization call that is only called by a
single thread in the team and replace it with a static shared memory buffer.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D97818
This adds support for addrspace casts involving opaque pointers to
InstCombine, as well as the isEliminableCastPair() helper
(otherwise the assertion failure would just move there).
Add PointerType::hasSameElementTypeAs() to hide the element type
details.
Differential Revision: https://reviews.llvm.org/D104668
There was a bug from cost calculation for partially invariant unswitch.
The costs of non-duplicated blocks are substracted from the total LoopCost, so
anything that is duplicated should not be counted.
Differential Revision: https://reviews.llvm.org/D103816
Summary:
Memory globalization is required to maintain OpenMP standard semantics for data sharing between
worker and master threads. The GPU cannot share data between its threads so must allocate global or
shared memory to store the data in. Currently this is implemented fully in the frontend using the
`__kmpc_data_sharing_push_stack` and __kmpc_data_sharing_pop_stack` functions to emulate standard
CPU stack sharing. The front-end scans the target region for variables that escape the region and
must be shared between the threads. Each variable then has a field created for it in a global record
type.
This patch replaces this functinality with a single allocation command, effectively mimicing an
alloca instruction for the variables that must be shared between the threads. This will be much
slower than the current solution, but makes it much easier to optimize as we can analyze each
variable independently and determine if it is not captured. In the future, we can replace these
calls with an `alloca` and small allocations can be pushed to shared memory.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D97680
Don't do this while stipping pointer casts, instead fetch it at
the end. This improves compatibility with opaque pointers for the
case where the base object is not opaque.
Currently we drop wrapping flags for expressions like (A + C1)<flags> - C2.
But we can retain flags under certain conditions:
* Adding a smaller constant is NUW if the original AddExpr was NUW.
* Adding a constant with the same sign and small magnitude is NSW, if the
original AddExpr was NSW.
This can improve results after using `SimplifyICmpOperands`, which may
subtract one in order to use stricter predicates, as is the case for
`isKnownPredicate`.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D104319
Patch was reverted due to a bug that existed before it and was exposed
by it. Returning after the underlying bug has been fixed.
Differential Revision: https://reviews.llvm.org/D103959
A backedge-taken count doesn't refer to memory; returning a pointer type
is nonsense. So make sure we always return an integer.
The obvious way to do this would be to just convert the operands of the
icmp to integers, but that doesn't quite work out at the moment:
isLoopEntryGuardedByCond currently gets confused by ptrtoint operations.
So we perform the ptrtoint conversion late for lt/gt operations.
The test changes are mostly innocuous. The most interesting changes are
more complex SCEV expressions of the form "(-1 * (ptrtoint i8* %ptr to
i64)) + %ptr)". This is expected: we can't fold this to zero because we
need to preserve the pointer base.
The call to isLoopEntryGuardedByCond in howFarToZero is less precise
because of ptrtoint operations; this shows up in the function
pr46786_c26_char in ptrtoint.ll. Fixing it here would require more
complex refactoring. It should eventually be fixed by future
improvements to isImpliedCond.
See https://bugs.llvm.org/show_bug.cgi?id=46786 for context.
Differential Revision: https://reviews.llvm.org/D103656
Reapplied without changes -- this was reverted together with an
underlying patch.
-----
Bitcasts having opaque pointer source or result type cannot be
converted into a zero-index GEP, GEP source and result types
always have the same opaque-ness.
Relative to the original patch, an InstCombine test has been
added to show a previously missed pattern, and the Coroutine
test that resulted in the revert has been regenerated.
-----
Move this into a separate function, to make sure that early
returns do not accidentally skip other transforms. This previously
happened for the isSized() check, which skipped folds like
distributing a bitcast over a select.
Perform better analysis when trying to vectorize PHIs.
1. Do not try to vectorize vector PHIs.
2. Do deeper analysis for more profitable nodes for the vectorization.
Before we just tried to vectorize the PHIs of the same type. Patch
improves this and tries to vectorize PHIs with incoming values which
come from the same basic block, have the same and/or alternative
opcodes.
It allows to save the compile time and provides better vectorization
results in general.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103638
Bitcasts having opaque pointer source or result type cannot be
converted into a zero-index GEP, GEP source and result types
always have the same opaque-ness.
They are not conducive to being stored in git. Instead, we autogenerate
mock model artifacts for use in tests. Production models can be
specified with the cmake flag LLVM_INLINER_MODEL_PATH.
LLVM_INLINER_MODEL_PATH has two sentinel values:
- download, which will download the most recent compatible model.
- autogenerate, which will autogenerate a "fake" model for testing the
model uptake infrastructure.
Differential Revision: https://reviews.llvm.org/D104251
This reverts commit bb1dc876eb.
This patch causes an assertion failure when building an arm64 defconfig
Linux kernel.
See https://reviews.llvm.org/D103959 for a link to the original bug
report and a reduced reproducer.
Use poison instead of undef for cases dealing with unreachable
code. This still leaves the more interesting case of "load from
uninitialized memory" as undef.
getSpecializationCost was returning INT_MAX for a case when specialisation
shouldn't happen, but this wasn't properly checked if specialisation was
forced.
Differential Revision: https://reviews.llvm.org/D104461
This patch lifts the requirement to have the only incoming live block
for Phis. There can be multiple live blocks if the same value comes to
phi from all of them.
Differential Revision: https://reviews.llvm.org/D103959
Reviewed By: nikic, lebedev.ri
This patch updates InstCombine to use poison constant to represent the resulting value of (either semantically or syntactically) unreachable instrs, or a don't-care value of an unreachable store instruction.
This allows more aggressive folding of unused results, as shown in llvm/test/Transforms/InstCombine/getelementptr.ll .
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D104602
This is a more general alternative/extension to D102635. Rather than
handling the special case of "header exit with non-exiting latch",
this unrolls against the smallest exact trip count from any exit.
The latch exit is no longer treated as priviledged when it comes to
full unrolling.
The motivating case is in full-unroll-one-unpredictable-exit.ll.
Here the header exit is an IV-based exit, while the latch exit is
a data comparison. This kind of loop does not get rotated, because
the latch is already exiting, and loop rotation doesn't try to
distinguish IV-based/analyzable latches.
Differential Revision: https://reviews.llvm.org/D102982
DSE will currently only remove stores in the same block unless they can
be guaranteed to be loop invariant. This expands that to any stores that
are in the same Loop, at the same loop level. This should still account
for where AA/MSSA will not handle aliasing between loops, but allow the
dead stores to be removed where they overlap in the same loop iteration.
It requires adding loop info to DSE, but that looks fairly harmless.
The test case this helps is from code like this, which can come up in
certain matrix operations:
for(i=..)
dst[i] = 0;
for(j=..)
dst[i] += src[i*n+j];
After LICM, this becomes:
for(i=..)
dst[i] = 0;
sum = 0;
for(j=..)
sum += src[i*n+j];
dst[i] = sum;
The first store is dead, and with this patch is now removed.
Differntial Revision: https://reviews.llvm.org/D100464
This problem is exposed by D104598, after it tail-merges `ret` in
`@test_inline_constraint_S_label`, the verifier would start complaining
`invalid operand for inline asm constraint 'S'`.
Essentially, taking address of a block is mismodelled in IR.
It should probably be an explicit instruction, a first one in block,
that isn't identical to any other instruction of the same type,
so that it can't be hoisted.
This patch adds an optional PriorityInlineOrder, which uses the heap to order inlining.
The callsite which size is smaller would have a higher priority.
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D104028
On ELF, the D1003372 optimization can apply to more cases. There are two
prerequisites for making `__profd_` private:
* `__profc_` keeps `__profd_` live under compiler/linker GC
* `__profd_` is not referenced by code
The first is satisfied because all counters/data are in a section group (either
`comdat any` or `comdat noduplicates`). The second requires that the function
does not use value profiling.
Regarding the second point: `__profd_` may be referenced by other text sections
due to inlining. There will be a linker error if a prevailing text section
references the non-prevailing local symbol.
With this change, a stage 2 (`-DLLVM_TARGETS_TO_BUILD=X86 -DLLVM_BUILD_INSTRUMENTED=IR`)
clang is 4.2% smaller (1-169620032/177066968).
`stat -c %s **/*.o | awk '{s+=$1}END{print s}' is 2.5% smaller.
Reviewed By: davidxl, rnk
Differential Revision: https://reviews.llvm.org/D103717
As a follow-up to https://reviews.llvm.org/D104129, I'm cleaning up the danling probe related code in both the compiler and llvm-profgen.
I'm seeing a 5% size win for the pseudo_probe section for SPEC2017 and 10% for Ciner. Certain benchmark such as 602.gcc has a 20% size win. No obvious difference seen on build time for SPEC2017 and Cinder.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D104477
Remove dependence on ULO.TripCount/ULO.TripMultiple from ORE and
debug code. For debug code, print information about all exits.
For optimization remarks, only include the unroll count and the
type of unroll (complete, partial or runtime), but omit detailed
information about exit folding, now that more than one exit may
be folded.
Differential Revision: https://reviews.llvm.org/D104482
This patch handles one particular case of one-iteration loops for which SCEV
cannot straightforwardly prove BECount = 1. The idea of the optimization is to
symbolically execute conditional branches on the 1st iteration, moving in topoligical
order, and only visiting blocks that may be reached on the first iteration. If we find out
that we never reach header via the latch, then the backedge can be broken.
This implementation uses InstSimplify. SCEV version was rejected due to high
compile time impact.
Differential Revision: https://reviews.llvm.org/D102615
Reviewed By: nikic
InstCombine didn't perform (sext bool X) * (sext bool X) --> zext (and X, X) which can result in just (zext X). The patch adds regression tests to check this transformation and adds a check for equality of mul's operands for that case.
Differential Revision: https://reviews.llvm.org/D104193
This patch adds an optional PriorityInlineOrder, which uses the heap to order inlining.
The callsite which size is smaller would have a higher priority.
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D104028
This looks like not a practical pattern in our codebase (it could fail
in some sandbox environement).
Instead we print it via standard output, and it is controled by the
-attributor-print-call-graph, this follows a similiar pattern of attributor-print-dep.
InstCombine didn't perform (sext bool X) * (sext bool X) --> zext (and X, X) which can result in just (zext X). The patch adds regression tests to check this transformation and adds a check for equality of mul's operands for that case.
Differential Revision: https://reviews.llvm.org/D104193
This reverts commit 76d0747e08.
If a group has `__llvm_prf_vals` due to static value profiler counters
(`NS!=0`), we cannot make `__llvm_prf_data` private, because a prevailing text
section may reference `__llvm_prf_data` and will cause a `relocation refers to a
discarded section` linker error.
Note: while a `__profc_` group is non-prevailing, it may be referenced by a
prevailing text section due to inlining.
```
group section [ 66] `.group' [__profc__ZN5clang20EmitClangDeclContextERN4llvm12RecordKeeperERNS0_11raw_ostreamE] contains 4 sections:
[Index] Name
[ 67] __llvm_prf_cnts
[ 68] __llvm_prf_vals
[ 69] __llvm_prf_data
[ 70] .rela__llvm_prf_data
```
This should fix PR50683. The wrong assumption was that we
could always know what the callee is when we replace a call site
argument with undef. We wanted to know that to remove the `noundef`
that might be attached to the argument. Since no callee means we
did the propagation on the caller site, there is no need to remove
an attribute. It is only needed if we replace all uses and therefore
pass `undef` instead of the value that was passed in otherwise.
To allow outside AAs that simplify values we need to ensure all value
simplification goes through the Attributor, not AAValueSimplify (or any
of the other AAs we have already like AAPotentialValues). This patch
also introduces an interface for the outside AAs to register
simplification callbacks for an IRPosition. To make this work as
expected we have to pass IRPositions instead of Values in
AAValueSimplify, which makes sense by itself.
If the target stack is not accessible between different running
"threads" we have to make sure not to create allocas for mallocs
that might be used by multiple "threads". The "use check" is
sufficient to prevent this but if we apply the "free check" we have
to make sure the pointer is not communicated to others before
the free is reached.
Differential Revision: https://reviews.llvm.org/D98608
This patch is to address https://bugs.llvm.org/show_bug.cgi?id=48857.
Previous attempts can be found in D104007 and D101980.
A lot of discussions can be found in those two patches.
To summarize the bug:
When Clang emits IR for coroutines, the first thing it does is to make a copy of every argument to the local stack, so that uses of the arguments in the function will all refer to the local copies instead of the arguments directly.
However, in some cases we find that arguments are still directly used:
When Clang emits IR for a function that has pass-by-value arguments, sometimes it emits an argument with byval attribute. A byval attribute is considered to be local to the function (just like alloca) and hence it can be easily determined that it does not alias other values. If in the IR there exists a memcpy from a byval argument to a local alloca, and then from that local alloca to another alloca, MemCpyOpt will optimize out the first memcpy because byval argument's content will not change. This causes issues because after a coroutine suspension, the byval argument may die outside of the function, and latter uses will lead to memory use-after-free.
This is only a problem for arguments with either byval attribute or noalias attribute, because only these two kinds are considered local. Arguments without these two attributes will be considered to alias coro_suspend and hence we won't have this problem. So we need to be able to deal with these two attributes in coroutines properly.
For noalias arguments, since coro_suspend may potentially change the value of any argument outside of the function, we simply shouldn't mark any argument in a coroutiune as noalias. This can be taken care of in CoroEarly pass.
For byval arguments, if such an argument needs to live across suspensions, we will have to copy their value content to the frame, not just the pointer.
Differential Revision: https://reviews.llvm.org/D104184
This attribute computes the optimistic live call edges using the attributor
liveness information. This attribute will be used for deriving a
inter-procedural function reachability attribute.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104059
For ELF, since all counters/data are in a section group (either `comdat any` or
`comdat noduplicates`), and the signature for `comdat any` is `__profc_`, the
D1003372 optimization prerequisite (linker GC cannot discard data variables
while the text section is retained) is always satisified, we can make __profd_
unconditionally private.
Reviewed By: davidxl, rnk
Differential Revision: https://reviews.llvm.org/D103717
This pass emits a floating point compare and a conditional branch,
but if strictfp is enabled we don't emit a constrained compare
intrinsic.
The backend also won't expand the readonly sqrt call this pass inserts
to a sqrt instruction under strictfp. So we end up with 2 libcalls as
seen here. https://godbolt.org/z/oax5zMEWd
Fix these things by disabling the pass.
Differential Revision: https://reviews.llvm.org/D104479
The old version of this code would blindly perform arithmetic without
paying attention to whether the types involved were pointers or
integers. This could lead to weird expressions like negating a pointer.
Explicitly handle simple cases involving pointers, like "x < y ? x : y".
In all other cases, coerce the operands of the comparison to integer
types. This avoids the weird cases, while handling most of the
interesting cases.
Differential Revision: https://reviews.llvm.org/D103660
Fold all exits based on known trip count/multiple information from
SCEV. Previously only the latch exit or the single exit were folded.
This doesn't yet eliminate ULO.TripCount and ULO.TripMultiple
entirely: They're still used to a) decide whether runtime unrolling
should be performed and b) for ORE remarks. However, the core
unrolling logic is independent of them now.
Differential Revision: https://reviews.llvm.org/D104203
In D103169 I'm adding to InstSimplify support for NaN to constrained
intrinsics that have a regular FP IR instruction counterpart. Precommit
the tests for clarity when that ticket lands.
As noted in PR45210: https://bugs.llvm.org/show_bug.cgi?id=45210
...the bug is triggered as Eli say when sext(idx) * ElementSize overflows.
```
// assume that GV is an array of 4-byte elements
GEP = gep GV, 0, Idx // this is accessing Idx * 4
L = load GEP
ICI = icmp eq L, value
=>
ICI = icmp eq Idx, NewIdx
```
The foldCmpLoadFromIndexedGlobal function simplifies GEP+load operation to icmp.
And there is a problem because Idx * ElementSize can overflow.
Let's assume that the wanted value is at offset 0.
Then, there are actually four possible values for Idx to match offset 0: 0x00..00, 0x40..00, 0x80..00, 0xC0..00.
We should return true for all these values, but currently, the new icmp only returns true for 0x00..00.
This problem can be solved by masking off (trailing zeros of ElementSize) bits from Idx.
```
...
=>
Idx' = and Idx, 0x3F..FF
ICI = icmp eq Idx', NewIdx
```
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D99481
The Interleave Access pass will convert shuffle(binop(load, load)) to
binop(shuffle(load), shuffle(load)), in order to create more
interleaving load patterns (VLD2/3/4) that might have been messed up by
instcombine. As shown in D104247 we were missing copying IR flags to the
new instruction though, which should just be kept the same as the
original instruction.
Differential Revision: https://reviews.llvm.org/D104255
This can be seen as a follow up to commit 0ee439b705,
that changed the second argument of __powidf2, __powisf2 and
__powitf2 in compiler-rt from si_int to int. That was to align with
how those runtimes are defined in libgcc.
One thing that seem to have been missing in that patch was to make
sure that the rest of LLVM also handle that the argument now depends
on the size of int (not using the si_int machine mode for 32-bit).
When using __builtin_powi for a target with 16-bit int clang crashed.
And when emitting libcalls to those rtlib functions, typically when
lowering @llvm.powi), the backend would always prepare the exponent
argument as an i32 which caused miscompiles when the rtlib was
compiled with 16-bit int.
The solution used here is to use an overloaded type for the second
argument in @llvm.powi. This way clang can use the "correct" type
when lowering __builtin_powi, and then later when emitting the libcall
it is assumed that the type used in @llvm.powi matches the rtlib
function.
One thing that needed some extra attention was that when vectorizing
calls several passes did not support that several arguments could
be overloaded in the intrinsics. This patch allows overload of a
scalar operand by adding hasVectorInstrinsicOverloadedScalarOpd, with
an entry for powi.
Differential Revision: https://reviews.llvm.org/D99439
Previously dangling samples were represented by INT64_MAX in sample profile while probes never executed were not reported. This was based on an observation that dangling probes were only at a smaller portion than zero-count probes. However, with compiler optimizations, dangling probes end up becoming at large portion of all probes in general and reporting them does not make sense from profile size point of view. This change flips sample reporting by reporting zero-count probes instead. This enabled dangling probe to be represented by none (missing entry in profile). This has a couple benefits:
1. Reducing sample profile size in optimize mode, even when the number of non-executed probes outperform the number of dangling probes, since INT64_MAX takes more space over 0 to encode.
2. Binary size savings. No need to encode dangling probe anymore, since missing probes are treated as dangling in the profile reader.
3. Reducing compiler work to track dangling probes. However, for probes that are real dead and removed, we still need the compiler to identify them so that they can be reported as zero-count, instead of mistreated as dangling probes.
4. Improving counts quality by respecting the counts already collected on the non-dangling copy of a probe. A probe, when duplicated, gets two copies at runtime. If one of them is dangling while the other is not, merging the two probes at profile generation time will cause the real samples collected on the non-dangling one to be discarded. Not reporting the dangling counterpart will keep the real samples.
5. Better readability.
6. Be consistent with non-CS dwarf line number based profile. Zero counts are trusted by the compiler counts inferencer while missing counts will be inferred by the compiler.
Note that the current patch does include any work for #3. There will be follow-up changes.
For #1, I've seen for a large Facebook service, the text profile is reduced by 7%. For extbinary profile, the size of LBRProfileSection is reduced by 35%.
For #4, I have seen general counts quality for SPEC2017 is improved by 10%.
Reviewed By: wenlei, wlei, wmi
Differential Revision: https://reviews.llvm.org/D104129
We already have this fold:
fadd float poison, 1.0 --> poison
...via ConstantFolding, so this makes the behavior consistent
if the other operand(s) are non-constant.
The fold for undef was added before poison existed as a
value/type in IR.
This came up in D102673 / D103169
because we're trying to sort out the more complicated handling
for constrained math ops.
We should have the handling for the regular instructions done
first, so we can build on that (or diverge as needed).
Differential Revision: https://reviews.llvm.org/D104383
The original implementation calculating UserBonus uses operator ^, which means XOR in C++
language.
At the first glance of reviewing, I thought it should be power, my bad.
It doesn't make sense to use XOR here. So I believe it should be a
carelessness as I made.
Test Plan: check-all
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D104282
We can look through invariant group intrinsics for the purposes of
simplifying the result of a load.
Since intrinsics can't be constants, but we also don't want to
completely rewrite load constant folding, we convert the load operand to
a constant. For GEPs and bitcasts we just treat them as constants. For
invariant group intrinsics, we treat them as a bitcast.
Relanding with a check for self-referential values.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D101103
Addition of this pass has been botched.
There is no particular reason why it had to be sold as an inseparable part
of new-pm transition. It was added when old-pm was still the default,
and very *very* few users were actually tracking new-pm,
so it's effects weren't measured.
Which means, some of the turnoil of the new-pm transition
are actually likely regressions due to this pass.
Likewise, there has been a number of post-commit feedback
(post new-pm switch), namely
* https://reviews.llvm.org/D37467#2787157 (regresses HW-loops)
* https://reviews.llvm.org/D37467#2787259 (should not be in middle-end, should run after LSR, not before)
* https://reviews.llvm.org/D95789 (an attempt to fix bad loop backedge metadata)
and in the half year past, the pass authors (google) still haven't found time to respond to any of that.
Hereby it is proposed to backout the pass from the pipeline,
until someone who cares about it can address the issues reported,
and properly start the process of adding a new pass into the pipeline,
with proper performance evaluation.
Furthermore, neither google nor facebook reports any perf changes
from this change, so i'm dropping the pass completely.
It can always be re-reverted should/if anyone want to pick it up again.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D104099
This adds support for functions outlined by the IR Outliner to be
recognized by the debugger. The expected behavior is that it will
skip over the instructions included in that section. This is due to the
fact that we can not say which of the original locations the
instructions originated from.
These functions will show up in the call stack, but you cannot step
through them.
Reviewers: paquette, vsk, djtodoro
Differential Revision: https://reviews.llvm.org/D87302
Loops with irreducible cycles may loop infinitely. Those cannot be
removed, unless the loop/function is marked as mustprogress.
Also discussed in D103382.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D104238
We canonicalized to these select patterns (poison-safe logic)
with D101191, so we need to reduce 'not' ops when possible
as we would with 'and'/'or' instructions.
This is shown in a secondary example in:
https://llvm.org/PR50389https://alive2.llvm.org/ce/z/BvsESh
Currently the irreducible cycles in the loops are ignored. The
irreducible cycle may loop infinitely in
irreducible_subloop_no_mustprogress, which is allowed and the loop
should not be removed.
Discussed in D103382.
As Eli mentioned post-commit in D103378, the result of the freeze may
still be out-of-range according to Alive2. So for now, just limit the
transform to indices that are non-poison.
This patch is to address https://bugs.llvm.org/show_bug.cgi?id=50610.
In computed goto pattern, there are usually a list of basic blocks that are all targets of indirectbr instruction, and each basic block also has address taken and stored in a variable.
CHR pass could potentially clone these basic blocks, which would generate a cloned version of the indirectbr and clonved version of all basic blocks in the list.
However these basic blocks will not have their addresses taken and stored anywhere. So latter SimplifyCFG pass will simply remove all tehse cloned basic blocks, resulting in incorrect code.
To fix this, when searching for scopes, we skip scopes that contains BBs with addresses taken.
Added a few test cases.
Reviewed By: aeubanks, wenlei, hoy
Differential Revision: https://reviews.llvm.org/D103867
The current implementation for computing relative block frequencies does
not handle correctly control-flow graphs containing irreducible loops. This
results in suboptimally generated binaries, whose perf can be up to 5%
worse than optimal.
To resolve the problem, we apply a post-processing step, which iteratively
updates block frequencies based on the frequencies of their predesessors.
This corresponds to finding the stationary point of the Markov chain by
an iterative method aka "PageRank computation". The algorithm takes at
most O(|E| * IterativeBFIMaxIterations) steps but typically converges faster.
It is turned on by passing option `use-iterative-bfi-inference`
and applied only for functions containing profile data and irreducible loops.
Tested on SPEC06/17, where it is helping to get correct profile counts for one of
the binaries (403.gcc). In prod binaries, we've seen a speedup of up to 2%-5%
for binaries containing functions with hot irreducible loops.
Reviewed By: hoy, wenlei, davidxl
Differential Revision: https://reviews.llvm.org/D103289
I don't like landing this change, but it's an acknowledgement of a practical reality. Despite not having well specified semantics for inttoptr and ptrtoint involving non-integral pointer types, they are used in practice. Here's a quick summary of the current pragmatic reality:
* I happen to know that the main external user of non-integral pointers has effectively disabled the verifier rules.
* RS4GC (the lowering pass for abstract GC machine model which is the key motivation for non-integral pointers), even supports them. We just have all the tests using an integral pointer space to let the verifier run.
* Certain idioms (such as alignment checks for alignment N, where any relocation is guaranteed to be N byte aligned) are fine in practice.
* As implemented, inttoptr/ptrtoint are CSEd and are not control dependent. This means that any code which is intending to check a particular bit pattern at site of use must be wrapped in an intrinsic or external function call.
This change allows them in the Verifier, and updates the LangRef to specific them as implementation dependent. This allows us to acknowledge current reality while still leaving ourselves room to punt on figuring out "good" semantics until the future.
The problematic code pattern in the test is based on:
https://llvm.org/PR50638
If the IfCond is itself the phi that we are trying to remove,
then the loop around line 2835 can end up with something like:
%cmp = select i1 %cmp, i1 false, i1 true
That can then lead to a use-after-free and assert (although
I'm still not seeing that locally in my release + asserts build).
I think this can only happen with unreachable code.
Differential Revision: https://reviews.llvm.org/D104063
This adds a function specialization pass to LLVM. Constant parameters
like function pointers and constant globals are propagated to the callee by
specializing the function.
This is a first version with a number of limitations:
- The pass is off by default, so needs to be enabled on the command line,
- It does not handle specialization of recursive functions,
- It does not yet handle constants and constant ranges,
- Only 1 argument per function is specialised,
- The cost-model could be further looked into, and perhaps related,
- We are not yet caching analysis results.
This is based on earlier work by Matthew Simpson (D36432) and Vinay Madhusudan.
More recently this was also discussed on the list, see:
https://lists.llvm.org/pipermail/llvm-dev/2021-March/149380.html.
The motivation for this work is that function specialisation often comes up as
a reason for performance differences of generated code between LLVM and GCC,
which has this enabled by default from optimisation level -O3 and up. And while
this certainly helps a few cpu benchmark cases, this also triggers in real
world codes and is thus a generally useful transformation to have in LLVM.
Function specialisation has great potential to increase compile-times and
code-size. The summary from some investigations with this patch is:
- Compile-time increases for short compile jobs is high relatively, but the
increase in absolute numbers still low.
- For longer compile-jobs, the extra compile time is around 1%, and very much
in line with GCC.
- It is difficult to blame one thing for compile-time increases: it looks like
everywhere a little bit more time is spent processing more functions and
instructions.
- But the function specialisation pass itself is not very expensive; it doesn't
show up very high in the profile of the optimisation passes.
The goal of this work is to reach parity with GCC which means that eventually
we would like to get this enabled by default. But first we would like to address
some of the limitations before that.
Differential Revision: https://reviews.llvm.org/D93838
This fixes the concern in single element store scalarization that the
alignment of new store may be larger than it should be. It calculates
the largest alignment if index is constant, and a safe one if not.
Reviewed By: lebedev.ri, spatel
Differential Revision: https://reviews.llvm.org/D103419
First we refactor the code which does no wrapping add sequences
match: we need to allow different operand orders for
the key add instructions involved in the match.
Then we use the refactored code trying 4 variants of matching operands.
Originally the code relied on the fact that the matching operands
of the two last add instructions of memory index calculations
had the same LHS argument. But which operand is the same
in the two instructions is actually not essential, so now we allow
that to be any of LHS or RHS of each of the two instructions.
This increases the chances of vectorization to happen.
Reviewed By: volkan
Differential Revision: https://reviews.llvm.org/D103912
SROA sometimes preserves MD_mem_parallel_loop_access and MD_access_group metadata on loads/stores, and sometimes fails to do so. This change adds copying of the MD after other CreateAlignedLoad/CreateAlignedStores. Also fix a case where the metadata was being copied from a load, rather than the store.
Added a LIT test to catch one case.
Patch by Mark Mendell
Differential Revision: https://reviews.llvm.org/D103254
As noted in https://bugs.llvm.org/show_bug.cgi?id=46666, the current behavior of assuming if-conversion safety if a loop is annotated parallel (`!llvm.loop.parallel_accesses`), is not expectable, the documentation for this behavior was since removed from the LangRef again, and can lead to invalid reads.
This was observed in POCL (https://github.com/pocl/pocl/issues/757) and would require similar workarounds in current work at hipSYCL.
The question remains why this was initially added and what the implications of removing this optimization would be.
Do we need an alternative mechanism to propagate the information about legality of if-conversion?
Or is the idea that conditional loads in `#pragma clang loop vectorize(assume_safety)` can be executed unmasked without additional checks flawed in general?
I think this implication is not part of what a user of that pragma (and corresponding metadata) would expect and thus dangerous.
Only two additional tests failed, which are adapted in this patch. Depending on the further direction force-ifcvt.ll should be removed or further adapted.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D103907
There is no need to schedule insertelement instructions. The compiler
did not schedule them before it started support their vectorization and
it should not do it after. We pre-schedule them manually when finding
a build vector sequence.
Disabling scheduling of insertelement instructions improves compile
time and vectorization of the very large basic blocks by saving
scheduling budget for other instructions.
Differential Revision: https://reviews.llvm.org/D104026
This patch allows folding stepvector + extract to the lane when the lane is
lower than the minimum size of the scalable vector. This fold is possible
because lane X of a stepvector is also X!
For instance, extracting element 3 of a <vscale x 4 x i64>stepvector is 3.
Differential Revision: https://reviews.llvm.org/D103153
This patch allows that scalable vector can fold extractelement and constant splat
only when the lane index is lower than the minimum number of elements of the vector.
Differential Revision: https://reviews.llvm.org/D103180
Previously such folding was enabled for half, float and double values
only. With this change it is allowed for other floating point values
also.
Differential Revision: https://reviews.llvm.org/D103956
Summary:
The current implementation of AANoFreeFloating will incorrectly list floating
point loads and stores as may-free. This prevents other attributor instances
like HeapToStack from pushing some allocations to the stack.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D103975
Upon encountering loads/stores on types whose size is not a multiple of 8 bits the SROA pass would either trip an assertion or use logic that was not meant to work with such irregularly-sized types.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D99435
1. Better sorting of scalars to be gathered. Trying to insert
constants/arguments/instructions-out-of-loop at first and only then
the instructions which are inside the loop. It improves hoisting of
invariant insertelements instructions.
2. Better detection of shuffle candidates in gathering function.
3. The cost of insertelement for constants is 0.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103458
Upon encountering loads/stores on types whose size is not a multiple of 8 bits the SROA pass would either trip an assertion or use logic that was not meant to work with such irregularly-sized types.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D99435
As discussed in the post-commit comments for:
3cdd05e519
It seems to be safe to propagate all flags from the final fneg
except for 'nsz' to the new select:
https://alive2.llvm.org/ce/z/J_APDc
nsz has unique FMF semantics: it is not poison, it is only
"insignificant" in the calculation according to the LangRef.
> This reapplies c0f3dfb9, which was reverted following the discovery of
> crashes on linux kernel and chromium builds - these issues have since
> been fixed, allowing this patch to re-land.
This reverts commit 36ec97f76a.
The change caused non-determinism in the compiler, see comments on the code
review at https://reviews.llvm.org/D91722.
Reverting to unbreak people's builds until that can be addressed.
This also reverts the follow-up "[DebugInfo] Limit the number of values
that may be referenced by a dbg.value" in
a0bd6105d8.
This patch changes LoopUnrollAndJamPass from FunctionPass to LoopNest pass.
The next patch will utilize LoopNest to effectively handle loop nests.
Also, a crash problem on legacy pass manager is fixed.
Reviewed By: Whitney
Differential Revision: https://reviews.llvm.org/D99149
Fixes getTypeConversion to return `TypeScalarizeScalableVector` when a scalable vector
type cannot be legalized by widening/splitting. When this is the method of legalization
found, getTypeLegalizationCost will return an Invalid cost.
The getMemoryOpCost, getMaskedMemoryOpCost & getGatherScatterOpCost functions already call
getTypeLegalizationCost and will now also return an Invalid cost for unsupported types.
Reviewed By: sdesmalen, david-arm
Differential Revision: https://reviews.llvm.org/D102515
This patch allows that scalable vector can also use the fold that already
exists for fixed vector, only when the lane index is lower than the minimum
number of elements of the vector.
Differential Revision: https://reviews.llvm.org/D102404
If the `-enable-strict-reductions` flag is set to true, then currently we will
always choose to vectorize the loop with strict in-order reductions. This is
not necessary where we allow the reordering of FP operations, such as
when loop hints are passed via metadata.
This patch moves useOrderedReductions so that we can also check whether
loop hints allow reordering, in which case we should use the default
behaviour of vectorizing with unordered reductions.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D103814
Needs to be discussed more.
This reverts commit 255a5c1baa6020c009934b4fa342f9f6dbbcc46
This reverts commit df2056ff3730316f376f29d9986c9913b95ceb1
This reverts commit faff79b7ca144e505da6bc74aa2b2f7cffbbf23
This reverts commit d2a9020785c6e02afebc876aa2778fa64c5cafd
Noticed via code inspection. We changed the semantics of the IR when we added mustprogress, and we appear to have not updated this location.
Differential Revision: https://reviews.llvm.org/D103834
Fix a bug introduced by f6f6f6375d.
Now for empty PHIs, instead of crashing on assert(hasVal()) in
Optional's internals, we'll return NoAlias, as we did before that patch.
Differential Revision: https://reviews.llvm.org/D103831
Unrolling with more iterations than MaxTripCount is pointless, as
those iterations can never be executed. As such, we clamp ULO.Count
to MaxTripCount if it is known. This means we no longer need to
consider iterations after MaxTripCount for exit folding, and the
CompletelyUnroll flag becomes independent of ULO.TripCount.
Differential Revision: https://reviews.llvm.org/D103748
This is a modified version of a patch by tolziplohu with a style change, and most importantly, a revised commit message.
inttoptr for a non-integral address space is currently ill defined in the LangRef. Figuring out exactly what the dynamic semantics of such a cast would be is hard, and not yet settled. Despite that, we still need to go ahead and implement something in RS4GC for a couple of reasons.
First, as a simple consistency argument. We're apparently added support for constexpr inttoptrs a while back, and even have tests which exercised them. Having a lack of constant folding trigger a crash during lowering is non-ideal.
Second, and more fundementally, the optimizer is allowed to insert undefined constructs in unreachable code. At the same time, we can't assume that dynamically dead code is always pruned before lowering. As a result, we must assume that inttoptrs can occur (even if completely ill defined) along dead paths. We need the lowering to not crash. The stackmaps produced can be garbage (as the assumption is the code is dynamically dead), but the lowering itself can't crash.
Differential Revision: https://reviews.llvm.org/D103492
The non-DOT printing does not include the successors of VPregionBlocks.
This patch use the same style for printing successors as for
VPBasicBlock.
I think the printing of successors could be a bit improved further, as
at the moment it is hard to ensure a check line matches all successors.
But that can be done as follow-up.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D103515
This patch is an extension of D103421. It allows the InstCombiner to
generate the negated form of integer scalable-vector splats. It can
technically handle fixed-length vectors too but those are completely
covered by the preceding logic.
This enables extra combining opportunities for scalable vector types.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103801
This patch extends the various "isXXX" functions of the `Constant` class
to include scalable-vector splats.
In several "isXXX" functions, code that was separately inspecting
`ConstantVector` and `ConstantDataVector` was unified to use
`getSplatValue`, which already includes support for said splats.
In the varous "isNotXXX" functions, code was added to check whether the
scalar splat value -- if any -- satisfies the predicate.
An extra fix for `isNotMinSignedValue` was included, as it previously
crashed when passed a scalable-vector type because it unconditionally
cast to `FixedVectorType`
These changes address numerous missed optimizations, a compiler crash
mentioned above and -- perhaps most egregiously -- an infinite loop in
InstCombine due to the compiler breaking canonical form when it failed
to pick up on a splat in a select instruction.
Test cases have been added to cover as many of these functions as
possible, though existing coverage is slim; it doesn't appear that there
are any in-tree uses of `Constant::isNegativeZeroValue`, for example.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103421
Use llvm.experimental.vector.insert instead of storing into an alloca
when generating code for these intrinsics. This defers the codegen of
the generated vector to instruction selection, allowing existing
shufflevector style optimizations to apply.
Additionally, introduce a new target transform that can recognise fixed
predicate patterns in the svbool variants of these intrinsics.
Differential Revision: https://reviews.llvm.org/D103082
This pass transforms loops that contain a conditional branch with induction
variable. For example, it transforms left code to right code:
newbound = min(n, c)
while (iv < n) { while(iv < newbound) {
A A
if (iv < c) B
B C
C }
} if (iv != n) {
while (iv < n) {
A
C
}
}
Differential Revision: https://reviews.llvm.org/D102234
This patch marks the induction increment of the main induction variable
of the vector loop as NUW when not folding the tail.
If the tail is not folded, we know that End - Start >= Step (either
statically or through the minimum iteration checks). We also know that both
Start % Step == 0 and End % Step == 0. We exit the vector loop if %IV +
%Step == %End. Hence we must exit the loop before %IV + %Step unsigned
overflows and we can mark the induction increment as NUW.
This should make SCEV return more precise bounds for the created vector
loops, used by later optimizations, like late unrolling.
At the moment quite a few tests still need to be updated, but before
doing so I'd like to get initial feedback to make sure I am not missing
anything.
Note that this could probably be further improved by using information
from the original IV.
Attempt of modeling of the assumption in Alive2:
https://alive2.llvm.org/ce/z/H_DL_g
Part of a set of fixes required for PR50412.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D103255
Loop peeling is currently performed as part of UnrollLoop().
Outside test scenarios, it is always performed with an unroll
count of 1. This means that unrolling doesn't actually do anything
apart from performing post-unroll simplification.
When testing, it's currently possible to specify both an explicit
peel count and an explicit unroll count. This doesn't perform any
sensible operation and may result in miscompiles, see
https://bugs.llvm.org/show_bug.cgi?id=45939.
This patch moves peeling from UnrollLoop() into tryToUnrollLoop(),
so that peeling does not also perform a susequent unroll. We only
run the post-unroll simplifications. Specifying both an explicit
peel count and unroll count is forbidden.
In the future, we may want to support both (non-PGO) peeling a
loop and unrolling it, but this needs to be done by first performing
the peel and then recalculating unrolling heuristics on a now
possibly analyzable loop.
Differential Revision: https://reviews.llvm.org/D103362
`__profd_*` variables are referenced by code only when value profiling is
enabled. If disabled (e.g. default -fprofile-instr-generate), the symbols just
waste space on ELF/Mach-O. We change the comdat symbol from `__profd_*` to
`__profc_*` because an internal symbol does not provide deduplication features
on COFF. The choice doesn't matter on ELF.
(In -DLLVM_BUILD_INSTRUMENTED_COVERAGE=on build, there is now no `__profd_*` symbols.)
On Windows this enables further optimization. We are no longer affected by the
link.exe limitation: an external symbol in IMAGE_COMDAT_SELECT_ASSOCIATIVE can
cause duplicate definition error.
https://lists.llvm.org/pipermail/llvm-dev/2021-May/150758.html
We can thus use llvm.compiler.used instead of llvm.used like ELF (D97585).
This avoids many `/INCLUDE:` directives in `.drectve`.
Here is rnk's measurement for Chrome:
```
This reduced object file size of base_unittests.exe, compiled with coverage, optimizations, and gmlt debug info by 10%:
#BEFORE
$ find . -iname '*.obj' | xargs du -b | awk '{ sum += $1 } END { print sum}'
1047758867
$ du -cksh base_unittests.exe
82M base_unittests.exe
82M total
# AFTER
$ find . -iname '*.obj' | xargs du -b | awk '{ sum += $1 } END { print sum}'
937886499
$ du -cksh base_unittests.exe
78M base_unittests.exe
78M total
```
The change is NFC for Mach-O.
Reviewed By: davidxl, rnk
Differential Revision: https://reviews.llvm.org/D103372
This is similar to b865eead76 ( D103617 ) and fixes:
https://llvm.org/PR5057541b71f718b did this and more (noted with TODO
comments in the tests), but it didn't handle the case
where the destination is narrower than the source, so
it got reverted.
This is a simple match-and-replace. If there's evidence
that the TODO cases are useful, we can revisit/extend.
Some floating point lib calls have ABI attributes that need to be set on
the caller. Found via D103412.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D103415
This builds on D103584. The change eliminates the coupling between unroll heuristic and implementation w.r.t. knowing when the passed in trip count is an exact trip count or a max trip count. In theory the new code is slightly less powerful (since it relies on exact computable trip counts), but in practice, it appears to cover all the same cases. It can also be extended if needed.
The test change shows what appears to be a bug in the existing code around the interaction of peeling and unrolling. The original loop only ran 8 iterations. The previous output had the loop peeled by 2, and then an exact unroll of 8. This meant the loop ran a total of 10 iterations which appears to have been a miscompile.
Differential Revision: https://reviews.llvm.org/D103620
`__profd_*` variables are referenced by code only when value profiling is
enabled. If disabled (e.g. default -fprofile-instr-generate), the symbols just
waste space on ELF/Mach-O. We change the comdat symbol from `__profd_*` to
`__profc_*` because an internal symbol does not provide deduplication features
on COFF. The choice doesn't matter on ELF.
(In -DLLVM_BUILD_INSTRUMENTED_COVERAGE=on build, there is now no `__profd_*` symbols.)
On Windows this enables further optimization. We are no longer affected by the
link.exe limitation: an external symbol in IMAGE_COMDAT_SELECT_ASSOCIATIVE can
cause duplicate definition error.
https://lists.llvm.org/pipermail/llvm-dev/2021-May/150758.html
We can thus use llvm.compiler.used instead of llvm.used like ELF (D97585).
This avoids many `/INCLUDE:` directives in `.drectve`.
Here is rnk's measurement for Chrome:
```
This reduced object file size of base_unittests.exe, compiled with coverage, optimizations, and gmlt debug info by 10%:
#BEFORE
$ find . -iname '*.obj' | xargs du -b | awk '{ sum += $1 } END { print sum}'
1047758867
$ du -cksh base_unittests.exe
82M base_unittests.exe
82M total
# AFTER
$ find . -iname '*.obj' | xargs du -b | awk '{ sum += $1 } END { print sum}'
937886499
$ du -cksh base_unittests.exe
78M base_unittests.exe
78M total
```
Reviewed By: davidxl, rnk
Differential Revision: https://reviews.llvm.org/D103372
One exit is unpredictable, the other has a known trip count. For
one function the predictable exit is the latch exit, for the other
the non-latch exit. Currently they are treated differently.
No need to recalculate the cost of extractelements, just no need to
compensate the cost of all extractelements, need to check before if this
is actually going to be removed at the vectorization. Also, no need to
generate new extractelement instruction, we may just regenerate the
original one. It may improve the final vectorization.
Differential Revision: https://reviews.llvm.org/D102933
tryToVectorizeList function allows to reorder only 2 scalars. Patch
allows to reorder >2 scalars. Also, to avoid possible regressions, it
allows extra vectorization of the remaining parts of the scalars
elements if possible.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103247
As noticed by NAKAMURA Takumi back in 2017, we cannot use
properlyDominates for std::stable_sort as properlyDominates only
partially orders blocks. That is, for blocks A, B, C, D, where A
dominates B and C dominates D, we have A == C, B == C, but A < B. This
is not a valid comparison function for std::stable_sort and causes
different results between libstdc++ and libc++. This change uses DFS
numbering to give deterministic results for all reachable blocks.
Unreachable blocks are ignored already, so do not need special
consideration.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103441
Coro-split functions with an active suspend point have their scope line set to
the line of the suspend point. However for compiler generated functions, this
results in debug info with unconventional results: a file named
`<compiler-generated>` with a non-zero line number. The convention for
`<compiler-generated>` is that the line number is zero.
This change propagates the scope line only for non-compiler generated
functions.
Differential Revision: https://reviews.llvm.org/D102412
This patch was split from https://reviews.llvm.org/D102246
[SampleFDO] New hierarchical discriminator for Flow Sensitive SampleFDO
This is mainly for ProfileData part of change. It will load
FS Profile when such profile is detected. For an extbinary format profile,
create_llvm_prof tool will add a flag to profile summary section.
For other format profiles, the users need to use an internal option
(-profile-isfs) to tell the compiler that the profile uses FS discriminators.
This patch also simplified the bit API used by FS discriminators.
Differential Revision: https://reviews.llvm.org/D103041
During Loop Strength Reduce, if the terminating condition for the loop
is not immediately adjacent to the terminating branch and it has more
than one use, a clone of the condition will be created just before the
terminating branch and will be used as the branch condition. Currently,
whether the instructions are "immediately adjacent" is determined by
checking whether the next instruction after the condition is the
terminating branch; this is incorrect however, as the presence of a
debug intrinsic between the two will result in a change to the output.
This is fixed by using getNextNonDebugInstruction() instead.
Differential Revision: https://reviews.llvm.org/D103033
Transfer the swiftasync attribute to the resume partial function according to
suspend.async specification. It's first argument denotes which argument is the
async context.
rdar://71499498
Differential Revision: https://reviews.llvm.org/D103285
This patch uses the calculated maximum scalable VFs to build VPlans,
cost them and select a suitable scalable VF.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D98722
This is a pre-commit of a test case D99439 which is a patch that
updates @llvm.powi to handle different int sizes for the exponent.
Problem is that @llvm.powi is used as an IR construct that maps
to RT libcalls to __powi* functions, and those lib functions depend
on sizeof(int) to use correct type for the exponent.
The test cases show that we use i32 for the powi expenent, which
later would result in wrong type being used in libcalls (miscompile).
But there are also a couple of the negative test cases that show
that we rewrite into using powi when having a uitofp conversion
from i16, which would be wrong when doing the libcall as an
"unsigned int" isn't guaranteed to fit inside the "int" argument
in the called libcall function.
Differential Revision: https://reviews.llvm.org/D102919
When rewriting
powf(2.0, itofp(x)) -> ldexpf(1.0, x)
exp2(sitofp(x)) -> ldexp(1.0, sext(x))
exp2(uitofp(x)) -> ldexp(1.0, zext(x))
the wrong type was used for the second argument in the ldexp/ldexpf
libc call, for target architectures with 16 bit "int" type.
The transform incorrectly used a bitcasted function pointer with
a 32-bit argument when emitting the ldexp/ldexpf call for such
targets.
The fault is solved by using the correct function prototype
in the call, by asking TargetLibraryInfo about the size of "int".
TargetLibraryInfo by default derives the size of the int type by
assuming that it is 16 bits for 16-bit architectures, and
32 bits otherwise. If this isn't true for a target it should be
possible to override that default in the TargetLibraryInfo
initializer.
Differential Revision: https://reviews.llvm.org/D99438
We can look through invariant group intrinsics for the purposes of
simplifying the result of a load.
Since intrinsics can't be constants, but we also don't want to
completely rewrite load constant folding, we convert the load operand to
a constant. For GEPs and bitcasts we just treat them as constants. For
invariant group intrinsics, we treat them as a bitcast.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D101103
As the existing test unreachable.ll shows, we should be doing more
work to avoid entering unreachable blocks: we should not stop
vectorization just because a PHI incoming value from an unreachable
block cannot be vectorized. We know that particular value will never
be used so we can just replace it with poison.
D101841 added this test. It appears to generate different outcome on different platforms.
Make it to only call -coro-split instead of entire O2 pipeline to simplify the test flow.
Hope this will make the test more robust.
Reviewed By: djtodoro
Differential Revision: https://reviews.llvm.org/D103418
Implemented better scheme for perfect/shuffled matches of the gather
nodes which allows to fix the performance regressions introduced by
earlier patches. Starting detecting matches for broadcast nodes and
extractelement gathering.
Differential Revision: https://reviews.llvm.org/D102920
InstCombine didn't perform the transformations when fmul's operands were
the same instruction because it required to have one use for each of them
which is false in the case. This patch fixes this + adds tests for them
and introduces a new function isOnlyUserOfAnyOperand to check these cases
in a single place.
This patch is a result of discussion in D102574.
Differential Revision: https://reviews.llvm.org/D102698
The current loop or any of its sub-loops may be infinite. Unless the
function or the loops are marked as mustprogress, this in itself makes
the loop *not* dead.
This patch moves the logic to check whether the current loop is finite
or mustprogress to `isLoopDead` and also extends it to check the
sub-loops. This should fix PR50511.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D103382
If the index itself is already poison, the poison propagates through
instructions clamping the index to a valid range. This still causes
introducing a load of poison, as flagged by Alive2 and pointed out
at 575e2aff55.
This patch updates the code to freeze the index, unless it is proven to
not be poison.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D103378
This reverts commit 4f2fd3818b.
The Linux kernel fails to build after this commit. See
https://reviews.llvm.org/D99481 for a reproducer.
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
This patch fixes pr43326 and pr48212.
Currently when we move reduction phis to the right place,
loop interchange assumes the first phi in loop headers is
an induction phi, skips the first phi and assumes the rest
of phis are candidate reduction phis to move. However, it
may not always be the case.
This patch loops over all phis in loop headers and considers
a phi node as a candidate reduction phi to move only when it
is indeed a reduction phi across outer and inner loop.
Reviewed By: Whitney
Differential Revision: https://reviews.llvm.org/D102743
Update isFirstOrderRecurrence to explore all uses of a recurrence phi
and check if we can sink them. If there are multiple users to sink, they
are all mapped to the previous instruction.
Fixes PR44286 (and another PR or two).
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D84951
This is a patch that replaces shufflevector and insertelement's placeholder value with poison.
Underlying motivation is to fix the semantics of shufflevector with undef mask to return poison instead
(D93818)
The consensus has been made in the late 2020 via mailing list as well as the thread in https://bugs.llvm.org/show_bug.cgi?id=44185 .
This patch is a simple syntactic change to the existing code, hence directly pushed as a commit.
DSE will currently only remove stores in the same block unless they can
be guaranteed to be loop invariant. This expands that to any stores that
are in the same Loop, at the same loop level. This should still account
for where AA/MSSA will not handle aliasing between loops, but allow the
dead stores to be removed where they overlap in the same loop iteration.
It requires adding loop info to DSE, but that looks fairly harmless.
The test case this helps is from code like this, which can come up in
certain matrix operations:
for(i=..)
dst[i] = 0;
for(j=..)
dst[i] += src[i*n+j];
After LICM, this becomes:
for(i=..)
dst[i] = 0;
sum = 0;
for(j=..)
sum += src[i*n+j];
dst[i] = sum;
The first store is dead, and with this patch is now removed.
Differntial Revision: https://reviews.llvm.org/D100464
As noted in PR45210: https://bugs.llvm.org/show_bug.cgi?id=45210
...the bug is triggered as Eli say when sext(idx) * ElementSize overflows.
```
// assume that GV is an array of 4-byte elements
GEP = gep GV, 0, Idx // this is accessing Idx * 4
L = load GEP
ICI = icmp eq L, value
=>
ICI = icmp eq Idx, NewIdx
```
The foldCmpLoadFromIndexedGlobal function simplifies GEP+load operation to icmp.
And there is a problem because Idx * ElementSize can overflow.
Let's assume that the wanted value is at offset 0.
Then, there are actually four possible values for Idx to match offset 0: 0x00..00, 0x40..00, 0x80..00, 0xC0..00.
We should return true for all these values, but currently, the new icmp only returns true for 0x00..00.
This problem can be solved by masking off (trailing zeros of ElementSize) bits from Idx.
```
...
=>
Idx' = and Idx, 0x3F..FF
ICI = icmp eq Idx', NewIdx
```
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D99481
This is similar to the fix in c590a9880d ( PR49832 ), but
we missed handling the pattern for select of bools (no compare
inst).
We can't substitute a vector value because the equality condition
replacement that we are attempting requires that the condition
is true/false for the entire value. Vector select can be partly
true/false.
I added an assert for vector types, so we shouldn't hit this again.
Fixed formatting while auditing the callers.
https://llvm.org/PR50500
Using the proper API automatically sets `__stack_chk_guard` to `dso_local` if
`Reloc::Static`. This wasn't strictly necessary until recently when dso_local was
no longer implied by `TargetMachine::shouldAssumeDSOLocal` for
`__stack_chk_guard`. By using the proper API, we can avoid generating unnecessary
GOT relocations.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D102646
This does not solve PR17101, but it is one of the
underlying diffs noted here:
https://bugs.llvm.org/show_bug.cgi?id=17101#c8
We could ease the one-use checks for the 'clear'
(no 'not' op) half of the transform, but I do not
know if that asymmetry would make things better
or worse.
Proofs:
https://rise4fun.com/Alive/uVB
Name: masked bit set
%sh1 = shl i32 1, %y
%and = and i32 %sh1, %x
%cmp = icmp ne i32 %and, 0
%r = zext i1 %cmp to i32
=>
%s = lshr i32 %x, %y
%r = and i32 %s, 1
Name: masked bit clear
%sh1 = shl i32 1, %y
%and = and i32 %sh1, %x
%cmp = icmp eq i32 %and, 0
%r = zext i1 %cmp to i32
=>
%xn = xor i32 %x, -1
%s = lshr i32 %xn, %y
%r = and i32 %s, 1
Note: this is a re-post of a patch that I committed at:
rGa041c4ec6f7a
The commit was reverted because it exposed another bug:
rGb212eb7159b40
But that has since been corrected with:
rG8a156d1c2795189 ( D101191 )
Differential Revision: https://reviews.llvm.org/D72396
Follow the same strategy used for atomic loads/stores by converting the operands to equally-sized integer types.
This change prevents the atomic expansion pass from generating illegal LL/SC pairs when targeting AArch64: `expand-atomicrmw-xchg-fp.ll` would previously instantiate intrinsics such as `llvm.aarch64.ldaxr.p0f32` that cannot be lowered.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D103232
This is to show that we currently only convert the terminator to
unreachable, but don't clean up instructions before it (unless
trivial DCE removes them).
Also clean up excessive whitespace in this test.
This is split off from D102002, and I think it is clear that
the difference in behavior was not intended. Options were
added to SimplifyCFG over time, but different chunks of
the pass pipelines were not kept in sync.
Nikita Popov