Summary:
A recent change in the instruction simplifier enables a call to a function that just returns one of its parameter to be simplified as simply loading the parameter. This exposes a bug in the inliner where double inlining may be involved which in turn may cause compiler ICE when an already-inlined callsite is reused for further inlining.
To put it simply, in the following-like C program, when the function call second(t) is inlined, its code t = third(t) will be reduced to just loading the return value of the callsite first(). This causes the inliner internal data structure to register the first() callsite for the call edge representing the third() call, therefore incurs a double inlining when both call edges are considered an inline candidate. I'm making a fix to break the inliner from reusing a callsite for new call edges.
```
void top()
{
int t = first();
second(t);
}
void second(int t)
{
t = third(t);
fourth(t);
}
void third(int t)
{
return t;
}
```
The actual failing case is much trickier than the example here and is only reproducible with the legacy inliner. The way the legacy inliner works is to process each SCC in a bottom-up order. That means in reality function first may be already inlined into top, or function third is either inlined to second or is folded into nothing. To repro the failure seen from building a large application, we need to figure out a way to confuse the inliner so that the bottom-up inlining is not fulfilled. I'm doing this by making the second call indirect so that the alias analyzer fails to figure out the right call graph edge from top to second and top can be processed before second during the bottom-up. We also need to tweak the test code so that when the inlining of top happens, the function body of second is not that optimized, by delaying the pass of function attribute deducer (i.e, which tells function third has no side effect and just returns its parameter). Since the CGSCC pass is iterative, additional calls are added to top to postpone the inlining of second to the second round right after the first function attribute deducing pass is done. I haven't been able to repro the failure with the new pass manager since the processing order of ininlined callsites is a bit different, but in theory the issue could happen there too.
Note that this fix could introduce a side effect that blocks the simplification of inlined code, specifically for a call site that can be folded to another call site. I hope this can probably be complemented by subsequent inlining or folding, as shown in the attached unit test. The ideal fix should be to separate the use of VMap. However, in reality this failing pattern shouldn't happen often. And even if it happens, there should be a good chance that the non-folded call site will be refolded by iterative inlining or subsequent simplification.
Reviewers: wenlei, davidxl, tejohnson
Reviewed By: wenlei, davidxl
Subscribers: eraman, nikic, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76248
Consider a callee function that has a call (C) within it which feeds
into the return. When we inline that callee into a callsite that has
return attributes, we can backward propagate valid attributes to the
call (C) within that inlined callee body.
This is safe to do so only if we can guarantee transfer of execution to
successor in the window of instructions between return value (i.e. the
call C) and the return instruction.
Also, this is valid only for attributes which are a property of a
callsite and not those that are not dependent on the ABI, or a property
of the call itself.
Reviewed-By: reames, jdoerfert
Differential Revision: https://reviews.llvm.org/D76140
bitcast (shuf V, MaskC) --> shuf (bitcast V), MaskC'
We do not attempt this in InstCombine because we do not want to change
types and create new shuffle ops that are potentially not lowered as
well as the original code. Here, we can check the cost model to see if
it is worthwhile.
I've aggressively enabled this transform even if the types are the same
size and/or equal cost because moving the bitcast allows InstCombine to
make further simplifications.
In the motivating cases from PR35454:
https://bugs.llvm.org/show_bug.cgi?id=35454
...this is enough to let instcombine and the backend eliminate the
redundant shuffles, but we probably want to extend VectorCombine to
handle the inverse pattern (shuffle-of-bitcast) to get that
simplification directly in IR.
Differential Revision: https://reviews.llvm.org/D76727
These are versions of a function that regressed with:
rGf2fbdf76d8d0
That particular problem occurs with an instcombine-simplifycfg-instcombine
sequence, but we can show that it exists within instcombine only with
other variations of the pattern.
This reverts commit f2fbdf76d8.
As noted in the post-commit thread:
https://reviews.llvm.org/rGf2fbdf76d8d0
...this can obscure a min/max pattern where the components
have extra uses. We can show that the problem is independent
of this change with a slightly modified source example, so
this revert just delays/reduces the need to fix the real
problem.
We need to improve our analysis of negation or -- more
generally -- subtraction using patches like D77230 or D68408.
Summary:
Splitting Knowledge retention into Queries in Analysis and Builder into Transform/Utils
allows Queries and Transform/Utils to use Analysis.
Reviewers: jdoerfert, sstefan1
Reviewed By: jdoerfert
Subscribers: mgorny, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77171
Summary:
This allows doing `memcmp(p, q, 7)` with 2 loads instead of a call to
memcmp.
This fixes part of PR45147.
Reviewers: spatel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76133
As pointed out by @thakis, currently CallSiteSplitting bails out after
checking the first PHI node. We should check all PHI nodes, until we
find one where call site splitting is beneficial.
This patch also slightly simplifies the code using BasicBlock::phis().
Reviewers: davidxl, junbuml, thakis
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D77089
There was a TODO in genericValueTraversal to provide the context
instruction and due to the lack of it users that wanted one just used
something available. Unfortunately, using a fixed instruction is wrong
in the presence of PHIs so we need to update the context instruction
properly.
Reviewed By: uenoku
Differential Revision: https://reviews.llvm.org/D76870
Use DL & ABI information for better alignment deduction, e.g., if a type
is accessed and the ABI specifies an alignment requirement for such an
access we can use it. This is based on a patch by @lebedev.ri and
inspired by getBaseAlign in Loads.cpp.
Depends on D76673.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D76674
If we have a must-tail call the callee and caller need to have matching
ABIs. Part of that is alignment which we might modify when we deduce
alignment of arguments of either. Since we would need to keep them in
sync, which is not as simple, we simply avoid deducing alignment for
arguments of the must-tail caller or callee.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D76673
Negation is equivalent to bitwise-not + 1, so try to convert more
subtracts into adds using this relationship:
0 - (A ^ C) => ((A ^ C) ^ -1) + 1 => A ^ ~C + 1
I doubt this will recover the regression noted in rGf2fbdf76d8d0,
but seems like we're going to need to improve here and/or revive D68408?
Alive2 proofs:
http://volta.cs.utah.edu:8080/z/Re5tMUhttp://volta.cs.utah.edu:8080/z/An-uns
Differential Revision: https://reviews.llvm.org/D77230
find() was altering the UserChain, even in cases where it subsequently
discovered that the resulting constant was a 0. This confuses
rebuildWithoutConstOffset() when it attempts to walk the chain later, since it
is expected that the chain itself be a path down the use-def edges of an
expression.
Make the attributor pass aware of aligned_alloc for converting heap
allocations to stack ones.
Depends on D76971.
Differential Revision: https://reviews.llvm.org/D76974
Summary:
Aggregate types containing scalable vectors aren't supported and as far
as I can tell this pass is mostly concerned with optimisations on
aggregate types, so the majority of this pass isn't very useful for
scalable vectors.
This patch modifies SROA such that mem2reg is run on allocas with
scalable types that are promotable, but nothing else such as slicing is
done.
The use of TypeSize in this pass has also been updated to be explicitly
fixed size. When invoking the following methods in DataLayout:
* getTypeSizeInBits
* getTypeStoreSize
* getTypeStoreSizeInBits
* getTypeAllocSize
we now called getFixedSize on the resultant TypeSize. This is quite an
extensive change with around 50 calls to these functions, and also the
first change of this kind (being explicit about fixed vs scalable
size) as far as I'm aware, so feedback welcome.
A test is included containing IR with scalable vectors that this pass is
able to optimise.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D76720
Currently ConstantRange::binaryAnd/binaryOr results are too pessimistic
for single element constant ranges.
If both operands are single element ranges, we can use APInt's AND and
OR implementations directly.
Note that some other binary operations on constant ranges can cover the
single element cases naturally, but for OR and AND this unfortunately is
not the case.
Reviewers: nikic, spatel, lebedev.ri
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D76446
Summary:
Select folding in JumpThreading can create a conditional branch on a
code patch that did not have one in the original program. This is not a
valid transformation in sanitize_memory functions.
Note that JumpThreading does select folding in 3 different places. Two
of them seem safe - they apply to a select instruction in a BB that ends
with an unconditional branch to another BB, which (in turn) ends with a
conditional branch or a switch with the same condition.
Fixes PR45220.
Reviewers: glider, dvyukov, efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76332
This reverts commit 28518d9ae3.
There is a failure in MsgPackReader.cpp when built with clang. It
complains about "signext and zeroext" are incompatible. Investigating
offline if it is infact a UB in the MsgPackReader code.
The attached test case is simplified from tcmalloc. Both function calls should be optimized as tailcall. But llvm can only optimize the first call. The second call can't be optimized because function dupRetToEnableTailCallOpts failed to duplicate ret into block case2.
There 2 problems blocked the duplication:
1 Intrinsic call llvm.assume is not handled by dupRetToEnableTailCallOpts.
2 The control flow is more complex than expected, dupRetToEnableTailCallOpts can only duplicate ret into its predecessor, but here we have an intermediate block between call and ret.
The solutions:
1 Since CodeGenPrepare is already at the end of LLVM IR phase, we can simply delete the intrinsic call to llvm.assume.
2 A general solution to the complex control flow is hard, but for this case, after exit2 is duplicated into case1, exit2 is the only successor of exit1 and exit1 is the only predecessor of exit2, so they can be combined through eliminateFallThrough. But this function is called too late, there is no more dupRetToEnableTailCallOpts after it. We can add an earlier call to eliminateFallThrough to solve it.
Differential Revision: https://reviews.llvm.org/D76539
Consider a callee function that has a call (C) within it which feeds
into the return. When we inline that callee into a callsite that has
return attributes, we can backward propagate those attributes to the
call (C) within that inlined callee body.
This is safe to do so only if we can guarantee transfer of execution to
successor in the window of instructions between return value (i.e. the
call C) and the return instruction.
See added test cases.
Reviewed-By: reames, jdoerfert
Differential Revision: https://reviews.llvm.org/D76140
Make InstCombine aware of the aligned_alloc library function.
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Depends on D76970.
Differential Revision: https://reviews.llvm.org/D76971
Summary: this patch preserve information from various places in EarlyCSE into assume bundles.
Reviewers: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76769
The script now includes extra info about command-line options used
when generating its advertisement heading, but we don't want that
here. This is a special-case because we have enhanced the check
lines (as noted in the 2nd comment line).
This patch updates ValueLattice to distinguish between ranges that are
guaranteed to not include undef and ranges that may include undef.
A constant range guaranteed to not contain undef can be used to simplify
instructions to arbitrary values. A constant range that may contain
undef can only be used to simplify to a constant. If the value can be
undef, it might take a value outside the range. For example, consider
the snipped below
define i32 @f(i32 %a, i1 %c) {
br i1 %c, label %true, label %false
true:
%a.255 = and i32 %a, 255
br label %exit
false:
br label %exit
exit:
%p = phi i32 [ %a.255, %true ], [ undef, %false ]
%f.1 = icmp eq i32 %p, 300
call void @use(i1 %f.1)
%res = and i32 %p, 255
ret i32 %res
}
In the exit block, %p would be a constant range [0, 256) including undef as
%p could be undef. We can use the range information to replace %f.1 with
false because we remove the compare, effectively forcing the use of the
constant to be != 300. We cannot replace %res with %p however, because
if %a would be undef %cond may be true but the second use might not be
< 256.
Currently LazyValueInfo uses the new behavior just when simplifying AND
instructions and does not distinguish between constant ranges with and
without undef otherwise. I think we should address the remaining issues
in LVI incrementally.
Reviewers: efriedma, reames, aqjune, jdoerfert, sstefan1
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D76931
Canonicalize the case when a scalar extracted from a vector is
truncated. Transform such cases to bitcast-then-extractelement.
This will enable erasing the truncate operation.
This commit fixes PR45314.
reviewers: spatel
Differential revision: https://reviews.llvm.org/D76983
Add a new llvm.amdgcn.ballot intrinsic modeled on the ballot function
in GLSL and other shader languages. It returns a bitfield containing the
result of its boolean argument in all active lanes, and zero in all
inactive lanes.
This is intended to replace the existing llvm.amdgcn.icmp and
llvm.amdgcn.fcmp intrinsics after a suitable transition period.
Use the new intrinsic in the atomic optimizer pass.
Differential Revision: https://reviews.llvm.org/D65088
For casts with constant range operands, we can use
ConstantRange::castOp.
Reviewers: davide, efriedma, mssimpso
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D71938
Compbinary format uses MD5 to represent strings in name table. That gives smaller profile without the need of compression/decompression when writing/reading the profile. The patch adds the support in extbinary format. It is off by default but user can choose to enable it.
Note the feature of using MD5 in name table can bring very small chance of name conflict leading to profile mismatch. Besides, profile using the feature won't have the profile remapping support.
Differential Revision: https://reviews.llvm.org/D76255
InstCombine has a mess of logic that tries to preserve min/max patterns,
but AFAICT, this one is not necessary because we can always narrow the
corresponding select in this sequence to match the narrow compare.
The biggest danger for this patch is inducing infinite looping or
assert from exceeding max iterations. If any bots hit that in the
vicinity of this commit, this is the likely patch to blame.
Summary:
Code frequently relies upon the results of "is.constant" intrinsics to
DCE invalid code paths. We don't want the intrinsic to be made control-
dependent on any additional values. For instance, we can't split a PHI
into a "constant" and "non-constant" part via jump threading in order
to "optimize" the constant part, because the "is.constant" intrinsic is
meant to return "false".
Reviewers: wmi, kazu, MaskRay
Reviewed By: kazu
Subscribers: jdoerfert, efriedma, joerg, lebedev.ri, nikic, xbolva00, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75799
For each natural loop with multiple exit blocks, this pass creates a
new block N such that all exiting blocks now branch to N, and then
control flow is redistributed to all the original exit blocks.
The bulk of the tranformation is a new function introduced in
BasicBlockUtils that an redirect control flow from a set of incoming
blocks to a set of outgoing blocks via a common "hub".
This is a useful workaround for a limitation in the structurizer which
incorrectly orders blocks when processing a nest of loops. This pass
bypasses that issue by ensuring that each natural loop is recognized
as a separate region. Since the structurizer is a region pass, it no
longer sees a nest of loops in a single region, and instead processes
each "level" in the nesting as a separate region.
The AMDGPU backend provides a new option to enable this pass before
the structurizer, which may eventually be enabled by default.
Reviewers: madhur13490, arsenm, nhaehnle
Reviewed By: nhaehnle
Differential Revision: https://reviews.llvm.org/D75865
In InnerLoopVectorizer::getOrCreateTripCount, when the backedge taken
count is a SCEV add expression, its type is defined by the type of the
last operand of the add expression.
In the test case from PR45259, this last operand happens to be a
pointer, which (according to llvm::Type) does not have a primitive size
in bits. In this case, LoopVectorize fails to truncate the SCEV and
crashes as a result.
Uing ScalarEvolution::getTypeSizeInBits makes the truncation work as expected.
https://bugs.llvm.org/show_bug.cgi?id=45259
Differential Revision: https://reviews.llvm.org/D76669
Aligned_alloc is a standard lib function and has been in glibc since
2.16 and in the C11 standard. It has semantics similar to malloc/calloc
for several analyses/transforms. This patch introduces aligned_alloc
in target library info and memory builtins. Subsequent ones will
make other passes aware and fix https://bugs.llvm.org/show_bug.cgi?id=44062
This change will also be useful to LLVM generators that need to allocate
buffers of vector elements larger than 16 bytes (for eg. 256-bit ones),
element boundary alignment for which is not typically provided by glibc malloc.
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Differential Revision: https://reviews.llvm.org/D76970
Summary:
On targets with different pointer sizes, -alignment-from-assumptions could attempt to create SCEV expressions which use different effective SCEV types. The provided test illustrates the issue.
In `getNewAlignment`, AASCEV would be the (only) alloca, which would have an effective SCEV type of i32. But PtrSCEV, the GEP in this case, due to being in the flat/default address space, will have an effective SCEV of i64.
This patch resolves the issue by truncating PtrSCEV to AASCEV's effective type.
Reviewers: hfinkel, jdoerfert
Reviewed By: jdoerfert
Subscribers: jvesely, nhaehnle, hiraditya, javed.absar, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75471
To make sure that replaced operands get DCEd. This drops one
iteration from gepphigep.ll, which is still not optimal.
This was the last test case performing more than 3 iterations.
NFC-ish, only worklist order should change.
Because this code does not use the IC-aware replaceInstUsesWith()
helper, we need to manually push users to the worklist.
This is NFC-ish, in that it may only change worklist order.
This change implements constant folding to constrained versions of
intrinsics, implementing rounding: floor, ceil, trunc, round, rint and
nearbyint.
Differential Revision: https://reviews.llvm.org/D72930
Previously, we would ignore alloca alignment when building the frame
and just use the natural alignment of the allocated type. If an alloca
is over-aligned for its IR type, this could lead to a frame entry with
inadequate alignment for the downstream uses of the alloca.
Since highly-aligned fields also tend to produce poor layouts under a
naive layout algorithm, I've also switched coroutine frames to use the
new optimal struct layout algorithm.
In order to communicate the frame size and alignment to later passes,
I needed to set align+dereferenceable attributes on the frame-pointer
parameter of the resume function. This is clearly the right thing to
do, but the align attribute currently seems to result in assumptions
being added during inlining that the optimizer cannot easily remove.
Summary:
This patch allows code-sinking in InstCombine to be performed when instruction have uses in llvm.assume.
Use are considered droppable when it is preferable to modify the User such that the use disappears rather than to prevent a transformation because of the use.
for now uses are considered droppable if they are in an llvm.assume.
Reviewers: jdoerfert, nikic, spatel, lebedev.ri, sstefan1
Reviewed By: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73832
Because using -print-imports is not thread-safe, make the test rely on llvm-dis instead.
Also cover the ICALL-PROM part as intended originally.
Differential Revision: https://reviews.llvm.org/D76775
This patch integrates operand bundle llvm.assumes [0] with the
Attributor. Most IRAttributes will now look at uses of the associated
value and if there are llvm.assume operand bundle uses with the right
tag we will check if they are in the must-be-executed-context (around
the context instruction). Droppable users, which is currently only
llvm::assume, are handled special in some places now as well.
[0] http://lists.llvm.org/pipermail/llvm-dev/2019-December/137632.html
Reviewed By: uenoku
Differential Revision: https://reviews.llvm.org/D74888
Summary:
DivRemPairs is unsound with respect to undef values.
```
// bb1:
// %rem = srem %x, %y
// bb2:
// %div = sdiv %x, %y
// -->
// bb1:
// %div = sdiv %x, %y
// %mul = mul %div, %y
// %rem = sub %x, %mul
```
If X can be undef, X should be frozen first.
For example, let's assume that Y = 1 & X = undef:
```
%div = sdiv undef, 1 // %div = undef
%rem = srem undef, 1 // %rem = 0
=>
%div = sdiv undef, 1 // %div = undef
%mul = mul %div, 1 // %mul = undef
%rem = sub %x, %mul // %rem = undef - undef = undef
```
http://volta.cs.utah.edu:8080/z/m7Xrx5
Same for Y. If X = 1 and Y = (undef | 1), %rem in src is either 1 or 0,
but %rem in tgt can be one of many integer values.
This resolves https://bugs.llvm.org/show_bug.cgi?id=42619 .
This miscompilation disappears if undef value is removed, but it may take a while.
DivRemPair happens pretty late during the optimization pipeline, so this optimization seemed as a good candidate to fix without major regression using freeze than other broken optimizations.
Reviewers: spatel, lebedev.ri, george.burgess.iv
Reviewed By: spatel
Subscribers: wuzish, regehr, nlopes, nemanjai, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76483
These should really be moved over to a ConstantFolding test file,
but since this may overlap with the in-progress D76010 and similar
tests already exist here, we can do that as a later cleanup.
coroutine frame
Currently we move all allocas into the frame when build coroutine frame in
CoroSplit pass. However, this can be relaxed.
Since CoroSplit pass run after Inline pass, we can use lifetime intrinsic to
do such analysis: If the scope of lifetime intrinsic is not across any suspend
point, rather than move the allocas to frame, we can just move them to entry bb
of corresponding function. This reduce the frame size.
More importantly, this also avoid data race in multithread environment.
Consider one inline function by coroutine: it starts a thread which access
local variables, while after inline the movement of allocs to frame also access
them. cause data race.
Differential Revision: https://reviews.llvm.org/D75664
PR35760 shows an example program which, when compiled with `clang -O0`
or gcc at any optimization level, prints '0'. However, llvm transforms
the program in a way that causes it to print '1'.
Fix the issue by having `AllUsesOfValueWillTrapIfNull` return false when
analyzing a load from a global which is used by an `icmp`. This special
case was untested [0] so this is just deleting dead code.
An alternative fix might be to change the GlobalStatus analysis for the
global to report "Stored" instead of "StoredOnce". However, "StoredOnce"
is appropriate when only one value other than the initializer is stored
to the global.
[0]
http://lab.llvm.org:8080/coverage/coverage-reports/coverage/Users/buildslave/jenkins/workspace/coverage/llvm-project/llvm/lib/Transforms/IPO/GlobalOpt.cpp.html#L662
Differential Revision: https://reviews.llvm.org/D76645
There seems to be a small benefit to the legalized sequence for v2f16
round with packed instructions, so allow vectorizing it by reducing
the cost.
An unintended side effect is vectorization of f32 round also
happens. The current FMA logic seems off to me, and isn't checking for
packed instructions.
Since intrinsics can now specify when an argument is required to be
constant, it is now OK to replace arguments with variables if they
aren't. This means intrinsics must now be accurately marked with
immarg.
Two one-use checks were added with rGfdcb27105537,
but only the first one is necessary to limit an
increase in instruction count. The second transform
only creates one instruction, so it is always a
reasonable canonicalization/optimization.
Validation of the found runtime library functions declarations types
(return and argument types) with the expected types.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D76058
We special cased must-tail calls all over the place because they cannot
be modified as other calls can be. However, we already centralized the
modification API so we can centralize the handling as well. This
simplifies the code and allows to remove must-tail calls completely.
This patch attempts to more accurately model the reduction of
power of 2 vectors of types we natively support. This takes into
account the narrowing of vectors that occur as we go from 512
bits to 256 bits, to 128 bits. It also takes into account the use
of wider elements in the shuffles for the first 2 steps of a
reduction from 128 bits. And uses a v8i16 shift for the final step
of vXi8 reduction.
The default implementation uses the legalized type for the arithmetic
for all levels. And uses the single source permute cost of the
legalized type for all levels. This penalizes things like
lack of v16i8 pshufb on pre-sse3 targets and the splitting and
joining that needs to be done for integer types on AVX1. We never
need v16i8 shuffle for a reduction and we only need split AVX1 ops
when type the type wide and needs to be split. I think we're still
over costing splits and joins for AVX1, but we're closer now.
I've also removed all pairwise special casing because I don't
think we ever want to generate that on X86. I've also adjusted
the add handling to more accurately account for any type splitting
that occurs before we reach a legal type.
Differential Revision: https://reviews.llvm.org/D76478
D75801 removed the last and only user of this option, so we can
drop it now. The original idea behind this was to only run expensive
transforms under -O3, but apart from the one known bits transform,
this has never really taken off. I believe nowadays the recommendation
is to put expensive transforms in AggressiveInstCombine instead,
though that isn't terribly popular either :)
Differential Revision: https://reviews.llvm.org/D76540
Summary:
DataLayout::getTypeStoreSize() returns TypeSize.
For cases where it can not be scalable vector (e.g., GlobalVariable),
explicitly call TypeSize::getFixedSize().
For cases where scalable property doesn't matter, (e.g., check for
zero-sized type), use TypeSize::isNonZero().
Reviewers: sdesmalen, efriedma, apazos, reames
Reviewed By: efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76454
Summary:
Return None if GEP index type is scalable vector. Size of scalable vectors
are multiplied by a runtime constant.
Avoid transforming:
%a = bitcast i8* %p to <vscale x 16 x i8>*
%tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 0
store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp0
%tmp1 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 1
store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp1
into:
%a = bitcast i8* %p to <vscale x 16 x i8>*
%tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 0
%1 = bitcast <vscale x 16 x i8>* %tmp0 to i8*
call void @llvm.memset.p0i8.i64(i8* align 16 %1, i8 0, i64 32, i1 false)
Reviewers: sdesmalen, efriedma, apazos, reames
Reviewed By: sdesmalen
Subscribers: tschuett, hiraditya, rkruppe, arphaman, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76464
If ExpensiveCombines is enabled (which is the case with -O3 on the
legacy PM and always on the new PM), InstCombine tries to compute
the known bits of all instructions in the hope that all bits end up
being known, which is fairly expensive.
How effective is it? If we add some statistics on how often the
constant folding succeeds and how many KnownBits calculations are
performed and run test-suite we get:
"instcombine.NumConstPropKnownBits": 642,
"instcombine.NumConstPropKnownBitsComputed": 18744965,
In other words, we get one fold for every 30000 KnownBits calculations.
However, the truth is actually much worse: Currently, known bits are
computed before performing other folds, so there is a high chance
that cases that get folded by known bits would also have been
handled by other folds.
What happens if we compute known bits after all other folds
(hacky implementation: https://gist.github.com/nikic/751f25b3b9d9e0860db5dde934f70f46)?
"instcombine.NumConstPropKnownBits": 0,
"instcombine.NumConstPropKnownBitsComputed": 18105547,
So it turns out despite doing 18 million known bits calculations,
the known bits fold does not do anything useful on test-suite.
I was originally planning to move this into AggressiveInstCombine
so it only runs once in the pipeline, but seeing this, I think
we're better off removing it entirely.
As this is the only use of the "expensive combines" mechanism,
it may be removed afterwards, but I'll leave that to a separate patch.
Differential Revision: https://reviews.llvm.org/D75801
Ideally SimplifyDemanded should compute the same known bits as
computeKnownBits(). This patch addresses one discrepancy, where
ValueTracking is more powerful: If we have a shl nsw shift, we
know that the sign bit of the input and output must be the same.
If this results in a conflict, the result is poison.
This is implemented in
2c4ca6832f/lib/Analysis/ValueTracking.cpp (L1175-L1179)
and
2c4ca6832f/lib/Analysis/ValueTracking.cpp (L904-L908).
This implements the same basic logic in SimplifyDemanded. It's
slightly stronger, because I return undef instead of zero for the
poison case (which is not an option inside ValueTracking).
As mentioned in https://reviews.llvm.org/D75801#inline-698484,
we could detect poison in more cases, this just establishes parity
with the existing logic.
Differential Revision: https://reviews.llvm.org/D76489
The sll/srl/sra scalar vector shifts can be replaced with generic shifts if the shift amount is known to be in range.
This also required public DemandedElts variants of llvm::computeKnownBits to be exposed (PR36319).
If a call argument has the "returned" attribute, we can simplify
the call to the value of that argument. This was already partially
handled by InstSimplify/InstCombine for the case where the argument
is an integer constant, and the result is thus known via known bits.
The non-constant (or non-int) argument cases weren't handled though.
This previously landed as an InstSimplify transform, but was reverted
due to assertion failures when compiling the Linux kernel. The reason
is that simplifying a call to another call breaks assumptions in
call graph updating during inlining. As the code is not easy to fix,
and there is no particularly strong motivation for having this in
InstSimplify, the transform is only performed in InstCombine instead.
Differential Revision: https://reviews.llvm.org/D75815
This is the same change as D75824, but for two cases where
InstCombine performs the same optimization: Replacing an instruction
whose bits are fully known with a constant. This is not (generally)
legal for musttail calls.
Differential Revision: https://reviews.llvm.org/D76457
For MemoryPhis, we have to avoid that the MemoryPhi may be executed
before before the access we are currently looking at.
To do this we do a post-order numbering of the basic blocks in the
function and bail out once we reach a MemoryPhi with a larger (or equal)
post-order block number than the current MemoryAccess.
This changes the order in which we visit stores for elimination.
This patch also adds support for exploring multiple paths. We keep a worklist (ToCheck) of memory accesses that might be eliminated by our starting MemoryDef or MemoryPhis for further exploration. For MemoryPhis, we add the incoming values to the worklist, for MemoryDefs we add the defining access.
Reviewers: dmgreen, rnk, efriedma, bryant, asbirlea
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D72148
For now, when final suspend can be simplified by simplifySuspendPoint,
handleFinalSuspend is executed as well to remove last case in switch
instruction. This patch fixes it.
Differential Revision: https://reviews.llvm.org/D76345
Summary:
This patch fixes https://bugs.llvm.org/show_bug.cgi?id=44611 by
preventing an infinite loop in the jump threading pass when
-jump-threading-across-loop-headers is on. Specifically, without this
patch, jump threading through two basic blocks would trigger on the
same area of the CFG over and over, resulting in an infinite loop.
Consider testcase PR44611-across-header-hang.ll in this patch. The
first opportunity to thread through two basic blocks is:
from bb_body2 through bb_header and bb_body1 to bb_body2.
The pass duplicates bb_header and bb_body1 as, say, bb_header.thread1
and bb_body1.thread1. Since bb_header contains a successor edge back
to itself, bb_header.thread1 also contains a successor edge to
bb_header, immediately giving rise to the next jump threading
opportunity:
from bb_header.thread1 through bb_header and bb_body1 to bb_body2.
After that, we repeatedly thread an incoming edge into bb_header
through bb_header and bb_body1 to bb_body2. In other words, we keep
peeling one iteration from bb_header's self loop.
The patch fixes the problem by preventing the pass from duplicating a
basic block containing a self loop.
Reviewers: wmi, junparser, efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76390
The slli/srli/srai 'immediate' vector shifts (although its not immediate anymore to match gcc) can be replaced with generic shifts if the shift amount is known to be in range.
For PHIs with multiple incoming values, we can improve precision by
using constant ranges for integers. We can over-approximate phis
by merging the incoming values.
Reviewers: davide, efriedma, mssimpso
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D71933
If one of the operands of a binary operator is a constant range, we can
use ConstantRange::binaryOp to approximate the result.
We still handle single element constant ranges as we did previously,
with ConstantExpr::get(), because ConstantRange::binaryOp still gives
worse results in a few cases for single element ranges.
Also note that we bail out early if any of the operands is still unknown.
Reviewers: davide, efriedma, mssimpso
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D71936
On Powerpc fma is faster than fadd + fmul for some types,
(PPCTargetLowering::isFMAFasterThanFMulAndFAdd). we should implement target
hook isProfitableToHoist to prevent simplifyCFGpass from breaking fma
pattern by hoisting fmul to predecessor block.
Reviewed By: nemanjai
Differential Revision: https://reviews.llvm.org/D76207
Summary:
DataLayout::getTypeAllocSize() return TypeSize. For cases where scalable
property doesn't matter (check for zero-sized alloca), we should explicitly
call getKnownMinSize() to avoid implicit type conversion to uint64_t, which is
invalid for scalable vector type.
Reviewers: sdesmalen, efriedma, spatel, apazos
Reviewed By: efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76386
For selects with an unknown condition, we can approximate the result by
merging the state of both options. This automatically takes care of
the case where on operand is undef.
Reviewers: davide, efriedma, mssimpso
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D71935
Hongtao Yu