For the given test SROA detects possible replacement and creates a correct alloca. After that SROA is adding lifetime markers for this new alloca. The function getNewAllocaSlicePtr is trying to deduce the pointer type based on the original alloca, which is split, to use it later in lifetime intrinsic.
For the test we ended up with such code (rA is initial alloca [10 x float], which is split, and rA.sroa.0.0 is a new split allocation)
```
%rA.sroa.0.0.rA.sroa_cast = bitcast i32* %rA.sroa.0 to [10 x float]* <----- this one causing the assertion and is an extra bitcast
%5 = bitcast [10 x float]* %rA.sroa.0.0.rA.sroa_cast to i8*
call void @llvm.lifetime.start.p0i8(i64 4, i8* %5)
```
isAllocaPromotable code assumes that a user of alloca may go into lifetime marker through bitcast but it must be the only one bitcast to i8* type. In the test it's not a i8* type, return false and throw the assertion.
As we are creating a pointer, which will be used in lifetime markers only, the proposed fix is to create a bitcast to i8* immediately to avoid extra bitcast creation.
The test is a greatly simplified to just reproduce the assertion.
Author: Igor Tsimbalist <igor.v.tsimbalist@intel.com>
Reviewers: chandlerc, craig.topper
Reviewed By: chandlerc
Differential Revision: https://reviews.llvm.org/D55934
llvm-svn: 351325
Increment statistics counter NumSwitches at unswitchNontrivialInvariants() for
unswitching a non-trivial switch instruction. This is to fix a bug that it
increments NumBranches even for the case of switch instruction.
There is no functional change in this patch.
Differential Revision: https://reviews.llvm.org/D56408
llvm-svn: 351193
Currently when a select has a constant value in one branch and the select feeds
a conditional branch (via a compare/ phi and compare) we unfold the select
statement. This results in threading the conditional branch later on. Similar
opportunity exists when a select (with a constant in one branch) feeds a
switch (via a phi node). The patch unfolds select under this condition.
A testcase is provided.
llvm-svn: 350931
Summary:
The original patch addressed the use of BlockRPONumber by forcing a sequence point when accessing that map in a conditional. In short we found cases where that map was being accessed with blocks that had not yet been added to that structure. For context, I've kept the wall of text below, to what we are trying to fix, by always ensuring a updated BlockRPONumber.
== Backstory ==
I was investigating an ICE (segfault accessing a DenseMap item). This failure happened non-deterministically, with no apparent reason and only on a Windows build of LLVM (from October 2018).
After looking into the crashes (multiple core files) and running DynamoRio, the cores and DynamoRio (DR) log pointed to the same code in `GVN::performScalarPRE()`. The values in the map are unsigned integers, the keys are `llvm::BasicBlock*`. Our test case that triggered this warning and periodic crash is rather involved. But the problematic line looks to be:
GVN.cpp: Line 2197
```
if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
```
To test things out, I cooked up a patch that accessed the items in the map outside of the condition, by forcing a sequence point between accesses. DynamoRio stopped warning of the issue, and the test didn't seem to crash after 1000+ runs.
My investigation was on an older version of LLVM, (source from October this year). What it looks like was occurring is the following, and the assembly from the latest pull of llvm in December seems to confirm this might still be an issue; however, I have not witnessed the crash on more recent builds. Of course the asm in question is generated from the host compiler on that Windows box (not clang), but it hints that we might want to consider how we access the BlockRPONumber map in this conditional (line 2197, listed above). In any case, I don't think the host compiler is wrong, rather I think it is pointing out a possibly latent bug in llvm.
1) There is no sequence point for the `>=` operation.
2) A call to a `DenseMapBase::operator[]` can have the side effect of the map reallocating a larger store (more Buckets, via a call to `DenseMap::grow`).
3) It seems perfectly legal for a host compiler to generate assembly that stores the result of a call to `operator[]` on the stack (that's what my host compile of GVN.cpp is doing) . A second call to `operator[]` //might// encourage the map to 'grow' thus making any pointers to the map's store invalid. The `>=` compares the first and second values. If the first happens to be a pointer produced from operator[], it could be invalid when dereferenced at the time of comparison.
The assembly generated from the Window's host compiler does show the result of the first access to the map via `operator[]` produces a pointer to an unsigned int. And that pointer is being stored on the stack. If a second call to the map (which does occur) causes the map to grow, that address (on the stack) is now invalid.
Reviewers: t.p.northover, efriedma
Reviewed By: efriedma
Subscribers: efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D55974
llvm-svn: 350880
Summary:
Step 2 in using MemorySSA in LICM:
Use MemorySSA in LICM to do sinking and hoisting, all under "EnableMSSALoopDependency" flag.
Promotion is disabled.
Enable flag in LICM sink/hoist tests to test correctness of this change. Moved one test which
relied on promotion, in order to test all sinking tests.
Reviewers: sanjoy, davide, gberry, george.burgess.iv
Subscribers: llvm-commits, Prazek
Differential Revision: https://reviews.llvm.org/D40375
llvm-svn: 350879
That is, remove many of the calls to Type::getNumContainedTypes(),
Type::subtypes(), and Type::getContainedType(N).
I'm not intending to remove these accessors -- they are
useful/necessary in some cases. However, removing the pointee type
from pointers would potentially break some uses, and reducing the
number of calls makes it easier to audit.
llvm-svn: 350835
Current strategy of dropping `InstructionPrecedenceTracking` cache is to
invalidate the entire basic block whenever we change its contents. In fact,
`InstructionPrecedenceTracking` has 2 internal strictures: `OrderedInstructions`
that is needed to be invalidated whenever the contents changes, and the map
with first special instructions in block. This second map does not need an
update if we add/remove a non-special instuction because it cannot
affect the contents of this map.
This patch changes API of `InstructionPrecedenceTracking` so that it now
accounts for reasons under which we invalidate blocks. This should lead
to much less recalculations of the map and should save us some compile time
because in practice we don't typically add/remove special instructions.
Differential Revision: https://reviews.llvm.org/D54462
Reviewed By: efriedma
llvm-svn: 350694
minted `CallBase` class instead of the `CallSite` wrapper.
This moves the largest interwoven collection of APIs that traffic in
`CallSite`s. While a handful of these could have been migrated with
a minorly more shallow migration by converting from a `CallSite` to
a `CallBase`, it hardly seemed worth it. Most of the APIs needed to
migrate together because of the complex interplay of AA APIs and the
fact that converting from a `CallBase` to a `CallSite` isn't free in its
current implementation.
Out of tree users of these APIs can fairly reliably migrate with some
combination of `.getInstruction()` on the `CallSite` instance and
casting the resulting pointer. The most generic form will look like `CS`
-> `cast_or_null<CallBase>(CS.getInstruction())` but in most cases there
is a more elegant migration. Hopefully, this migrates enough APIs for
users to fully move from `CallSite` to the base class. All of the
in-tree users were easily migrated in that fashion.
Thanks for the review from Saleem!
Differential Revision: https://reviews.llvm.org/D55641
llvm-svn: 350503
In addition to finding dead uses of instructions, also find dead uses
of function arguments, and replace them with zero as well.
I'm changing the way the known bits are computed here to remove the
coupling between the transfer function and the algorithm. It previously
relied on the first op being visited first and computing known bits --
unless the first op is not an instruction, in which case they're computed
on the second op. I could have adjusted this to check for "instruction
or argument", but I think it's better to avoid the repeated calculation
with an explicit flag.
Differential Revision: https://reviews.llvm.org/D56247
llvm-svn: 350435
In some cases the order that we hoist instructions in means that when rehoisting
(which uses the same order as hoisting) we can rehoist to a block A, then a
block B, then block A again. This currently causes an assertion failure as it
expects that when changing the hoist point it only ever moves to a block that
dominates the hoist point being moved from.
Fix this by moving the re-hoist point when it doesn't dominate the dominator of
hoisted instruction, or in other words when it wouldn't dominate the uses of
the instruction being rehoisted.
Differential Revision: https://reviews.llvm.org/D55266
llvm-svn: 350408
If an instruction has no demanded bits, remove it directly during BDCE,
instead of leaving it for something else to clean up.
Differential Revision: https://reviews.llvm.org/D56185
llvm-svn: 350257
This (mostly) fixes https://bugs.llvm.org/show_bug.cgi?id=39771.
BDCE currently detects instructions that don't have any demanded bits
and replaces their uses with zero. However, if an instruction has
multiple uses, then some of the uses may be dead (have no demanded bits)
even though the instruction itself is still live. This patch extends
DemandedBits/BDCE to detect such uses and replace them with zero.
While this will not immediately render any instructions dead, it may
lead to simplifications (in the motivating case, by converting a rotate
into a simple shift), break dependencies, etc.
The implementation tries to strike a balance between analysis power and
complexity/memory usage. Originally I wanted to track demanded bits on
a per-use level, but ultimately we're only really interested in whether
a use is entirely dead or not. I'm using an extra set to track which uses
are dead. However, as initially all uses are dead, I'm not storing uses
those user is also dead. This case is checked separately instead.
The previous attempt to land this lead to miscompiles, because cases
where uses were initially dead but were later found to be live during
further analysis were not always correctly removed from the DeadUses
set. This is fixed now and the added test case demanstrates such an
instance.
Differential Revision: https://reviews.llvm.org/D55563
llvm-svn: 350188
Deletion of dead blocks in arbitrary order may lead to failure
of assertion in `DeleteDeadBlock` that requires that we have
deleted all predecessors before we can delete the current block.
We should instead delete them in RPO order.
llvm-svn: 350116
Summary:
Existing LIR recognizes CTLZ where shifting input variable right until it is zero. (Shift-Until-Zero idiom)
This commit:
1. Augments Shift-Until-Zero idiom to recognize CTTZ where input variable is shifted left.
2. Prepare for BitScan idiom recognition.
Patch by Yuanfang Chen (tabloid.adroit)
Reviewers: craig.topper, evstupac
Reviewed By: craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D55876
llvm-svn: 350074
This patch teaches LoopSimplifyCFG to remove dead exiting edges
from loops.
Differential Revision: https://reviews.llvm.org/D54025
Reviewed By: fedor.sergeev
llvm-svn: 350049
Both of these places reference memset-like loops. Memset is precise.
Trying to keep these patches super small so they're easily post-commit
verifiable, as requested in D44748.
llvm-svn: 350044
Trying to keep these patches super small so they're easily post-commit
verifiable, as requested in D44748.
srcSize is derived from the size of an alloca, and we quit out if the
size of that is > the size of the thing we're copying to. Hence, we
should always copy everything over, so these sizes are precise.
Don't make srcSize itself a LocationSize, since optionality isn't
helpful, and we do some comparisons against other sizes elsewhere in
that function.
llvm-svn: 350019
Instruction::isLifetimeStartOrEnd() checks whether an Instruction is an
llvm.lifetime.start or an llvm.lifetime.end intrinsic.
This was suggested as a cleanup in D55967.
Differential Revision: https://reviews.llvm.org/D56019
llvm-svn: 349964
The current llvm.mem.parallel_loop_access metadata has a problem in that
it uses LoopIDs. LoopID unfortunately is not loop identifier. It is
neither unique (there's even a regression test assigning the some LoopID
to multiple loops; can otherwise happen if passes such as LoopVersioning
make copies of entire loops) nor persistent (every time a property is
removed/added from a LoopID's MDNode, it will also receive a new LoopID;
this happens e.g. when calling Loop::setLoopAlreadyUnrolled()).
Since most loop transformation passes change the loop attributes (even
if it just to mark that a loop should not be processed again as
llvm.loop.isvectorized does, for the versioned and unversioned loop),
the parallel access information is lost for any subsequent pass.
This patch unlinks LoopIDs and parallel accesses.
llvm.mem.parallel_loop_access metadata on instruction is replaced by
llvm.access.group metadata. llvm.access.group points to a distinct
MDNode with no operands (avoiding the problem to ever need to add/remove
operands), called "access group". Alternatively, it can point to a list
of access groups. The LoopID then has an attribute
llvm.loop.parallel_accesses with all the access groups that are parallel
(no dependencies carries by this loop).
This intentionally avoid any kind of "ID". Loops that are clones/have
their attributes modifies retain the llvm.loop.parallel_accesses
attribute. Access instructions that a cloned point to the same access
group. It is not necessary for each access to have it's own "ID" MDNode,
but those memory access instructions with the same behavior can be
grouped together.
The behavior of llvm.mem.parallel_loop_access is not changed by this
patch, but should be considered deprecated.
Differential Revision: https://reviews.llvm.org/D52116
llvm-svn: 349725
This (mostly) fixes https://bugs.llvm.org/show_bug.cgi?id=39771.
BDCE currently detects instructions that don't have any demanded bits
and replaces their uses with zero. However, if an instruction has
multiple uses, then some of the uses may be dead (have no demanded bits)
even though the instruction itself is still live. This patch extends
DemandedBits/BDCE to detect such uses and replace them with zero.
While this will not immediately render any instructions dead, it may
lead to simplifications (in the motivating case, by converting a rotate
into a simple shift), break dependencies, etc.
The implementation tries to strike a balance between analysis power and
complexity/memory usage. Originally I wanted to track demanded bits on
a per-use level, but ultimately we're only really interested in whether
a use is entirely dead or not. I'm using an extra set to track which uses
are dead. However, as initially all uses are dead, I'm not storing uses
those user is also dead. This case is checked separately instead.
The test case has a couple of cases that are not simplified yet. In
particular, we're only looking at uses of instructions right now. I think
it would make sense to also extend this to arguments. Furthermore
DemandedBits doesn't yet know some of the tricks that InstCombine does
for the demanded bits or bitwise or/and/xor in combination with known
bits information.
Differential Revision: https://reviews.llvm.org/D55563
llvm-svn: 349674
When using clang with `-fno-unroll-loops` (implicitly added with `-O1`),
the LoopUnrollPass is not not added to the (legacy) pass pipeline. This
also means that it will not process any loop metadata such as
llvm.loop.unroll.enable (which is generated by #pragma unroll or
WarnMissedTransformationsPass emits a warning that a forced
transformation has not been applied (see
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20181210/610833.html).
Such explicit transformations should take precedence over disabling
heuristics.
This patch unconditionally adds LoopUnrollPass to the optimizing
pipeline (that is, it is still not added with `-O0`), but passes a flag
indicating whether automatic unrolling is dis-/enabled. This is the same
approach as LoopVectorize uses.
The new pass manager's pipeline builder has no option to disable
unrolling, hence the problem does not apply.
Differential Revision: https://reviews.llvm.org/D55716
llvm-svn: 349509
When splitting up an alloca's uses we were dropping any explicit
alignment tags, which means they default to the ABI-required default
alignment and this can cause miscompiles if the real value was smaller.
Also refactor the TBAA metadata into a parent class since it's shared by
both children anyway.
llvm-svn: 349465
The current code relies on LeaderUseCount to determine if we can remove
an SSA copy, but in that the LeaderUseCount does not refer to the SSA
copy. If a SSA copy is a dominating leader, we use the operand as dominating
leader instead. This means we removed a user of a ssa copy and we should
decrement its use count, so we can remove the ssa copy once it becomes dead.
Fixes PR38804.
Reviewers: efriedma, davide
Reviewed By: davide
Differential Revision: https://reviews.llvm.org/D51595
llvm-svn: 349217
Optimization transformations are intentionally disabled by the 'optnone'
function attribute. Therefore do not warn if transformation metadata is
still present.
Using the legacy pass manager structure, the `skipFunction` method takes
care for the optnone attribute (already called before this patch). For
the new pass manager, there is no equivalent, so we check for the
'optnone' attribute manually.
Differential Revision: https://reviews.llvm.org/D55690
llvm-svn: 349184
Currently memcpyopt optimizes cases like
memset(a, byte, N);
memcpy(b, a, M);
to
memset(a, byte, N);
memset(b, byte, M);
if M <= N. Often this allows further simplifications down the line,
which drop the first memset entirely.
This patch extends this optimization for the case where M > N, but we
know that the bytes a[N..M] are undef due to alloca/lifetime.start.
This situation arises relatively often for Rust code, because Rust does
not initialize trailing structure padding and loves to insert redundant
memcpys. This also fixes https://bugs.llvm.org/show_bug.cgi?id=39844.
The previous version of this patch did not perform dependency checking
properly: While the dependency is checked at the position of the memset,
the used size must be that of the memcpy. Previously the size of the
memset was used, which missed modification in the region
MemSetSize..CopySize, resulting in miscompiles. The added tests cover
variations of this issue.
Differential Revision: https://reviews.llvm.org/D55120
llvm-svn: 349078
When multiple loop transformation are defined in a loop's metadata, their order of execution is defined by the order of their respective passes in the pass pipeline. For instance, e.g.
#pragma clang loop unroll_and_jam(enable)
#pragma clang loop distribute(enable)
is the same as
#pragma clang loop distribute(enable)
#pragma clang loop unroll_and_jam(enable)
and will try to loop-distribute before Unroll-And-Jam because the LoopDistribute pass is scheduled after UnrollAndJam pass. UnrollAndJamPass only supports one inner loop, i.e. it will necessarily fail after loop distribution. It is not possible to specify another execution order. Also,t the order of passes in the pipeline is subject to change between versions of LLVM, optimization options and which pass manager is used.
This patch adds 'followup' attributes to various loop transformation passes. These attributes define which attributes the resulting loop of a transformation should have. For instance,
!0 = !{!0, !1, !2}
!1 = !{!"llvm.loop.unroll_and_jam.enable"}
!2 = !{!"llvm.loop.unroll_and_jam.followup_inner", !3}
!3 = !{!"llvm.loop.distribute.enable"}
defines a loop ID (!0) to be unrolled-and-jammed (!1) and then the attribute !3 to be added to the jammed inner loop, which contains the instruction to distribute the inner loop.
Currently, in both pass managers, pass execution is in a fixed order and UnrollAndJamPass will not execute again after LoopDistribute. We hope to fix this in the future by allowing pass managers to run passes until a fixpoint is reached, use Polly to perform these transformations, or add a loop transformation pass which takes the order issue into account.
For mandatory/forced transformations (e.g. by having been declared by #pragma omp simd), the user must be notified when a transformation could not be performed. It is not possible that the responsible pass emits such a warning because the transformation might be 'hidden' in a followup attribute when it is executed, or it is not present in the pipeline at all. For this reason, this patche introduces a WarnMissedTransformations pass, to warn about orphaned transformations.
Since this changes the user-visible diagnostic message when a transformation is applied, two test cases in the clang repository need to be updated.
To ensure that no other transformation is executed before the intended one, the attribute `llvm.loop.disable_nonforced` can be added which should disable transformation heuristics before the intended transformation is applied. E.g. it would be surprising if a loop is distributed before a #pragma unroll_and_jam is applied.
With more supported code transformations (loop fusion, interchange, stripmining, offloading, etc.), transformations can be used as building blocks for more complex transformations (e.g. stripmining+stripmining+interchange -> tiling).
Reviewed By: hfinkel, dmgreen
Differential Revision: https://reviews.llvm.org/D49281
Differential Revision: https://reviews.llvm.org/D55288
llvm-svn: 348944
IR-printing AfterPass instrumentation might be called on a loop
that has just been invalidated. We should skip printing it to
avoid spurious asserts.
Reviewed By: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D54740
llvm-svn: 348887
Currently memcpyopt optimizes cases like
memset(a, byte, N);
memcpy(b, a, M);
to
memset(a, byte, N);
memset(b, byte, M);
if M <= N. Often this allows further simplifications down the line,
which drop the first memset entirely.
This patch extends this optimization for the case where M > N, but we
know that the bytes a[N..M] are undef due to alloca/lifetime.start.
This situation arises relatively often for Rust code, because Rust does
not initialize trailing structure padding and loves to insert redundant
memcpys. This also fixes https://bugs.llvm.org/show_bug.cgi?id=39844.
For the implementation, I'm reusing a bit of code for a similar existing
optimization (direct memcpy of undef). I've also added memset support to
MemDepAnalysis GetLocation -- Instead, getPointerDependencyFrom could be
used, but it seems to make more sense to add this to GetLocation and thus
make the computation cachable.
Differential Revision: https://reviews.llvm.org/D55120
llvm-svn: 348645
DemandedBits and BDCE currently only support scalar integers. This
patch extends them to also handle vector integer operations. In this
case bits are not tracked for individual vector elements, instead a
bit is demanded if it is demanded for any of the elements. This matches
the behavior of computeKnownBits in ValueTracking and
SimplifyDemandedBits in InstCombine.
Unlike the previous iteration of this patch, getDemandedBits() can now
again be called on arbirary (sized) instructions, even if they don't
have integer or vector of integer type. (For vector types the size of the
returned mask will now be the scalar size in bits though.)
The added LoopVectorize test case shows a case which triggered an
assertion failure with the previous attempt, because getDemandedBits()
was called on a pointer-typed instruction.
Differential Revision: https://reviews.llvm.org/D55297
llvm-svn: 348602
This patch introduces a new instinsic `@llvm.experimental.widenable_condition`
that allows explicit representation for guards. It is an alternative to using
`@llvm.experimental.guard` intrinsic that does not contain implicit control flow.
We keep finding places where `@llvm.experimental.guard` is not supported or
treated too conservatively, and there are 2 reasons to that:
- `@llvm.experimental.guard` has memory write side effect to model implicit control flow,
and this sometimes confuses passes and analyzes that work with memory;
- Not all passes and analysis are aware of the semantics of guards. These passes treat them
as regular throwing call and have no idea that the condition of guard may be used to prove
something. One well-known place which had caused us troubles in the past is explicit loop
iteration count calculation in SCEV. Another example is new loop unswitching which is not
aware of guards. Whenever a new pass appears, we potentially have this problem there.
Rather than go and fix all these places (and commit to keep track of them and add support
in future), it seems more reasonable to leverage the existing optimizer's logic as much as possible.
The only significant difference between guards and regular explicit branches is that guard's condition
can be widened. It means that a guard contains (explicitly or implicitly) a `deopt` block successor,
and it is always legal to go there no matter what the guard condition is. The other successor is
a guarded block, and it is only legal to go there if the condition is true.
This patch introduces a new explicit form of guards alternative to `@llvm.experimental.guard`
intrinsic. Now a widenable guard can be represented in the CFG explicitly like this:
%widenable_condition = call i1 @llvm.experimental.widenable.condition()
%new_condition = and i1 %cond, %widenable_condition
br i1 %new_condition, label %guarded, label %deopt
guarded:
; Guarded instructions
deopt:
call type @llvm.experimental.deoptimize(<args...>) [ "deopt"(<deopt_args...>) ]
The new intrinsic `@llvm.experimental.widenable.condition` has semantics of an
`undef`, but the intrinsic prevents the optimizer from folding it early. This form
should exploit all optimization boons provided to `br` instuction, and it still can be
widened by replacing the result of `@llvm.experimental.widenable.condition()`
with `and` with any arbitrary boolean value (as long as the branch that is taken when
it is `false` has a deopt and has no side-effects).
For more motivation, please check llvm-dev discussion "[llvm-dev] Giving up using
implicit control flow in guards".
This patch introduces this new intrinsic with respective LangRef changes and a pass
that converts old-style guards (expressed as intrinsics) into the new form.
The naming discussion is still ungoing. Merging this to unblock further items. We can
later change the name of this intrinsic.
Reviewed By: reames, fedor.sergeev, sanjoy
Differential Revision: https://reviews.llvm.org/D51207
llvm-svn: 348593
The current algorithm that collects live/dead/inloop blocks relies on some invariants
related to RPO and PO traversals. In particular, the important fact it requires is that
the only loop's latch is the first block in PO traversal. It also relies on fact that during
RPO we visit all prececessors of a block before we visit this block (backedges ignored).
If a loop has irreducible non-loop cycle inside, both these assumptions may break.
This patch adds detection for this situation and prohibits the terminator folding
for loops with irreducible CFG.
We can in theory support this later, for this some algorithmic changes are needed.
Besides, irreducible CFG is not a frequent situation and we can just don't bother.
Thanks @uabelho for finding this!
Differential Revision: https://reviews.llvm.org/D55357
Reviewed By: skatkov
llvm-svn: 348567
Max Kazantsev