Runtime unrolling will create a prologue to execute the extra
iterations which is can't divided by the unroll factor. It
generates an if-then-else sequence to jump into a factor -1
times unrolled loop body, like
extraiters = tripcount % loopfactor
if (extraiters == 0) jump Loop:
if (extraiters == loopfactor) jump L1
if (extraiters == loopfactor-1) jump L2
...
L1: LoopBody;
L2: LoopBody;
...
if tripcount < loopfactor jump End
Loop:
...
End:
It means if the unroll factor is 4, the loop body will be 7
times unrolled, 3 are in loop prologue, and 4 are in the loop.
This commit is to use a loop to execute the extra iterations
in prologue, like
extraiters = tripcount % loopfactor
if (extraiters == 0) jump Loop:
else jump Prol
Prol: LoopBody;
extraiters -= 1 // Omitted if unroll factor is 2.
if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.
if (tripcount < loopfactor) jump End
Loop:
...
End:
Then when unroll factor is 4, the loop body will be copied by
only 5 times, 1 in the prologue loop, 4 in the original loop.
And if the unroll factor is 2, new loop won't be created, just
as the original solution.
llvm-svn: 218604
The annotation instructions are dropped during codegen and have no
impact on size. In some cases, the annotations were preventing the
unroller from unrolling a loop because the annotation calls were
pushing the cost over the unrolling threshold.
Differential Revision: http://reviews.llvm.org/D5335
llvm-svn: 218525
The doFinalization method checks that the LoopToAliasSetMap is
empty. LICM populates that map as it runs through the loop nest,
deleting the entries for child loops as it goes. However, if a child
loop is deleted by another pass (e.g. unrolling) then the loop will
never be deleted from the map because LICM walks the loop nest to
find entries it can delete.
The fix is to delete the loop from the map and free the alias set
when the loop is deleted from the loop nest.
Differential Revision: http://reviews.llvm.org/D5305
llvm-svn: 218387
Rather than slurping in and splatting out the whole ctor list, preserve
the existing array entries without trying to understand them. Only
remove the entries that we know we can optimize away. This way we don't
need to wire through priority and comdats or anything else we might add.
Fixes a linker issue where the .init_array or .ctors entry would point
to discarded initialization code if the comdat group from the TU with
the faulty global_ctors entry was dropped.
llvm-svn: 218337
This improves other optimizations such as LSR. A sext may be added to the
compare's other operand, but this can often be hoisted outside of the loop.
llvm-svn: 217953
Example:
define i1 @foo(i32 %a) {
%shr = ashr i32 -9, %a
%cmp = icmp ne i32 %shr, -5
ret i1 %cmp
}
Before this fix, the instruction combiner wrongly thought that %shr
could have never been equal to -5. Therefore, %cmp was always folded to 'true'.
However, when %a is equal to 1, then %cmp evaluates to 'false'. Therefore,
in this example, it is not valid to fold %cmp to 'true'.
The problem was only affecting the case where the comparison was between
negative quantities where one of the quantities was obtained from arithmetic
shift of a negative constant.
This patch fixes the problem with the wrong folding (fixes PR20945).
With this patch, the 'icmp' from the example is now simplified to a
comparison between %a and 1. This still allows us to get rid of the arithmetic
shift (%shr).
llvm-svn: 217950
Some ICmpInsts when anded/ored with another ICmpInst trivially reduces
to true or false depending on whether or not all integers or no integers
satisfy the intersected/unioned range.
This sort of trivial looking code can come about when InstCombine
performs a range reduction-type operation on sdiv and the like.
This fixes PR20916.
llvm-svn: 217750
We used to crash processing any relevant @llvm.assume on a 32-bit target
(because we'd ask SE to subtract expressions of differing types). I've copied
our 'simple.ll' test, but with the data layout from arm-linux-gnueabihf to get
some meaningful test coverage here.
llvm-svn: 217574
The routine that determines an alignment given some SCEV returns zero if the
answer is unknown. In a case where we could determine the increment of an
AddRec but not the starting alignment, we would compute the integer modulus by
zero (which is illegal and traps). Prevent this by returning early if either
the start or increment alignment is unknown (zero).
llvm-svn: 217544
"Unroll" is not the appropriate name for this variable. Clang already uses
the term "interleave" in pragmas and metadata for this.
Differential Revision: http://reviews.llvm.org/D5066
llvm-svn: 217528
From a combination of @llvm.assume calls (and perhaps through other means, such
as range metadata), it is possible that all bits of a return value might be
known. Previously, InstCombine did not check for this (which is understandable
given assumptions of constant propagation), but means that we'd miss simple
cases where assumptions are involved.
llvm-svn: 217346
This change teaches LazyValueInfo to use the @llvm.assume intrinsic. Like with
the known-bits change (r217342), this requires feeding a "context" instruction
pointer through many functions. Aside from a little refactoring to reuse the
logic that turns predicates into constant ranges in LVI, the only new code is
that which can 'merge' the range from an assumption into that otherwise
computed. There is also a small addition to JumpThreading so that it can have
LVI use assumptions in the same block as the comparison feeding a conditional
branch.
With this patch, we can now simplify this as expected:
int foo(int a) {
__builtin_assume(a > 5);
if (a > 3) {
bar();
return 1;
}
return 0;
}
llvm-svn: 217345
This adds a ScalarEvolution-powered transformation that updates load, store and
memory intrinsic pointer alignments based on invariant((a+q) & b == 0)
expressions. Many of the simple cases we can get with ValueTracking, but we
still need something like this for the more complicated cases (such as those
with an offset) that require some algebra. Note that gcc's
__builtin_assume_aligned's optional third argument provides exactly for this
kind of 'misalignment' offset for which this kind of logic is necessary.
The primary motivation is to fixup alignments for vector loads/stores after
vectorization (and unrolling). This pass is added to the optimization pipeline
just after the SLP vectorizer runs (which, admittedly, does not preserve SE,
although I imagine it could). Regardless, I actually don't think that the
preservation matters too much in this case: SE computes lazily, and this pass
won't issue any SE queries unless there are any assume intrinsics, so there
should be no real additional cost in the common case (SLP does preserve DT and
LoopInfo).
llvm-svn: 217344
This builds on r217342, which added the infrastructure to compute known bits
using assumptions (@llvm.assume calls). That original commit added only a few
patterns (to catch common cases related to determining pointer alignment); this
change adds several other patterns for simple cases.
r217342 contained that, for assume(v & b = a), bits in the mask
that are known to be one, we can propagate known bits from the a to v. It also
had a known-bits transfer for assume(a = b). This patch adds:
assume(~(v & b) = a) : For those bits in the mask that are known to be one, we
can propagate inverted known bits from the a to v.
assume(v | b = a) : For those bits in b that are known to be zero, we can
propagate known bits from the a to v.
assume(~(v | b) = a): For those bits in b that are known to be zero, we can
propagate inverted known bits from the a to v.
assume(v ^ b = a) : For those bits in b that are known to be zero, we can
propagate known bits from the a to v. For those bits in
b that are known to be one, we can propagate inverted
known bits from the a to v.
assume(~(v ^ b) = a) : For those bits in b that are known to be zero, we can
propagate inverted known bits from the a to v. For those
bits in b that are known to be one, we can propagate
known bits from the a to v.
assume(v << c = a) : For those bits in a that are known, we can propagate them
to known bits in v shifted to the right by c.
assume(~(v << c) = a) : For those bits in a that are known, we can propagate
them inverted to known bits in v shifted to the right by c.
assume(v >> c = a) : For those bits in a that are known, we can propagate them
to known bits in v shifted to the right by c.
assume(~(v >> c) = a) : For those bits in a that are known, we can propagate
them inverted to known bits in v shifted to the right by c.
assume(v >=_s c) where c is non-negative: The sign bit of v is zero
assume(v >_s c) where c is at least -1: The sign bit of v is zero
assume(v <=_s c) where c is negative: The sign bit of v is one
assume(v <_s c) where c is non-positive: The sign bit of v is one
assume(v <=_u c): Transfer the known high zero bits
assume(v <_u c): Transfer the known high zero bits (if c is know to be a power
of 2, transfer one more)
A small addition to InstCombine was necessary for some of the test cases. The
problem is that when InstCombine was simplifying and, or, etc. it would fail to
check the 'do I know all of the bits' condition before checking less specific
conditions and would not fully constant-fold the result. I'm not sure how to
trigger this aside from using assumptions, so I've just included the change
here.
llvm-svn: 217343
This change, which allows @llvm.assume to be used from within computeKnownBits
(and other associated functions in ValueTracking), adds some (optional)
parameters to computeKnownBits and friends. These functions now (optionally)
take a "context" instruction pointer, an AssumptionTracker pointer, and also a
DomTree pointer, and most of the changes are just to pass this new information
when it is easily available from InstSimplify, InstCombine, etc.
As explained below, the significant conceptual change is that known properties
of a value might depend on the control-flow location of the use (because we
care that the @llvm.assume dominates the use because assumptions have
control-flow dependencies). This means that, when we ask if bits are known in a
value, we might get different answers for different uses.
The significant changes are all in ValueTracking. Two main changes: First, as
with the rest of the code, new parameters need to be passed around. To make
this easier, I grouped them into a structure, and I made internal static
versions of the relevant functions that take this structure as a parameter. The
new code does as you might expect, it looks for @llvm.assume calls that make
use of the value we're trying to learn something about (often indirectly),
attempts to pattern match that expression, and uses the result if successful.
By making use of the AssumptionTracker, the process of finding @llvm.assume
calls is not expensive.
Part of the structure being passed around inside ValueTracking is a set of
already-considered @llvm.assume calls. This is to prevent a query using, for
example, the assume(a == b), to recurse on itself. The context and DT params
are used to find applicable assumptions. An assumption needs to dominate the
context instruction, or come after it deterministically. In this latter case we
only handle the specific case where both the assumption and the context
instruction are in the same block, and we need to exclude assumptions from
being used to simplify their own ephemeral values (those which contribute only
to the assumption) because otherwise the assumption would prove its feeding
comparison trivial and would be removed.
This commit adds the plumbing and the logic for a simple masked-bit propagation
(just enough to write a regression test). Future commits add more patterns
(and, correspondingly, more regression tests).
llvm-svn: 217342
This adds a set of utility functions for collecting 'ephemeral' values. These
are LLVM IR values that are used only by @llvm.assume intrinsics (directly or
indirectly), and thus will be removed prior to code generation, implying that
they should be considered free for certain purposes (like inlining). The
inliner's cost analysis, and a few other passes, have been updated to account
for ephemeral values using the provided functionality.
This functionality is important for the usability of @llvm.assume, because it
limits the "non-local" side-effects of adding llvm.assume on inlining, loop
unrolling, etc. (these are hints, and do not generate code, so they should not
directly contribute to estimates of execution cost).
llvm-svn: 217335
The special case did not work when run under -reassociate and can easily
be expressed by a further generalization of an existing pattern.
llvm-svn: 217227
LinearFunctionTestReplace tries to use the *next* indvar to compare
against when possible. However, it may be the case that the calculation
for the next indvar has NUW/NSW flags and that it may only be safely
used inside the loop. Using it in a comparison to calculate the exit
condition could result in observing poison.
This fixes PR20680.
Differential Revision: http://reviews.llvm.org/D5174
llvm-svn: 217102
Fixes two latent bugs:
- There was no fence inserted before expanded seq_cst load (unsound on Power)
- There was only a fence release before seq_cst stores (again unsound, in particular on Power)
It is not even clear if this is correct on ARM swift processors (where release fences are
DMB ishst instead of DMB ish). This behaviour is currently preserved on ARM Swift
as it is not clear whether it is incorrect. I would love to get documentation stating
whether it is correct or not.
These two bugs were not triggered because Power is not (yet) using this pass, and these
behaviours happen to be (mostly?) working on ARM
(although they completely butchered the semantics of the llvm IR).
See:
http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075821.html
for an example of the problems that can be caused by the second of these bugs.
I couldn't see a way of fixing these in a completely target-independent way without
adding lots of unnecessary fences on ARM, hence the target-dependent parts of this
patch.
This patch implements the new target-dependent parts only for ARM (the default
of not doing anything is enough for AArch64), other architectures will use this
infrastructure in later patches.
llvm-svn: 217076
The SLP vectorizer should propagate IR-level optimization hints/flags (nsw, nuw, exact, fast-math)
when converting scalar instructions into vectors. But this isn't a simple copy - we need to take
the intersection (the logical 'and') of the sets of flags on the scalars.
The solution is further complicated because we can have non-uniform (non-SIMD) vector ops after:
http://reviews.llvm.org/D4015http://llvm.org/viewvc/llvm-project?view=revision&revision=211339
The vast majority of changed files are existing tests that were not propagating IR flags, but I've
also added a new test file for focused testing of IR flag possibilities.
Differential Revision: http://reviews.llvm.org/D5172
llvm-svn: 217051
When folding a fused multiply-add builtin call, make sure that we propagate the
correct result in the case where the addend is zero, and the two other operands
are finite non-zero.
Example:
define double @test() {
%1 = call double @llvm.fma.f64(double 7.0, double 8.0, double 0.0)
ret double %1
}
Before this patch, the instruction simplifier wrongly folded the builtin call
in function @test to constant 'double 7.0'.
With this patch, method 'fusedMultiplyAdd' correctly evaluates the multiply and
propagates the expected result (i.e. 56.0).
Added test fold-builtin-fma.ll with the reproducible from PR20832 plus extra
test cases to verify the behavior of method 'fusedMultiplyAdd' in the presence
of NaN/Inf operands.
This fixes PR20832.
Differential Revision: http://reviews.llvm.org/D5152
llvm-svn: 216913
Summary:
BBs might contain non-LCSSA'd values after the LCSSA pass is run if they
are unreachable from the entry block.
Normally, the users of the instruction would be PHIs but the unreachable
BBs have normal users; rewrite their uses to be undef values.
An alternative fix could involve fixing this at LCSSA but that would
require this invariant to hold after subsequent transforms. If a BB
created an unreachable block, they would be in violation of this.
This fixes PR19798.
Differential Revision: http://reviews.llvm.org/D5146
llvm-svn: 216911
SROA may decide that it needs to insert a bitcast and would set it's
insertion point before a PHI. This will create an invalid module
right quick.
Instead, choose the first insertion point in the basic block that holds
our PHI.
This fixes PR20822.
Differential Revision: http://reviews.llvm.org/D5141
llvm-svn: 216891
This reverts commit r216698 which reverted r216523 and r216598.
We would attempt to perform the transformation even if the match()
failed because, as a side effect, it would set V. This would trick us
into believing that we correctly found a place to correctly apply the
transform.
An additional test case was added to getelementptr.ll so that we might
not regress in the future.
llvm-svn: 216890
The loop vectorizer preserves wrapping, exact, and fast-math properties of scalar instructions.
This patch adds a convenience method to make that operation easier because we need to do this
in the loop vectorizer, SLP vectorizer, and possibly other places.
Although this is a 'no functional change' patch, I've added a testcase to verify that the exact
flag is preserved by the loop vectorizer. The wrapping and fast-math flags are already checked
in existing testcases.
Differential Revision: http://reviews.llvm.org/D5138
llvm-svn: 216886
chain became completely broken here as *all* intrinsic users ended up
being skipped, and the ones that seemed to be singled out were actually
the exact wrong set.
This is a great example of why long else-if chains can be easily
confusing. Switch the entire code to use early exits and early continues
to have simpler (and more importantly, correct) logic here, as well as
fixing the reversed logic for detecting and continuing on lifetime
intrinsics.
I've also significantly cleaned up the test case and added another test
case demonstrating an example where the optimization is not (trivially)
safe to perform.
llvm-svn: 216871
Previously, the hint mechanism relied on clean up passes to remove redundant
metadata, which still showed up if running opt at low levels of optimization.
That also has shown that multiple nodes of the same type, but with different
values could still coexist, even if temporary, and cause confusion if the
next pass got the wrong value.
This patch makes sure that, if metadata already exists in a loop, the hint
mechanism will never append a new node, but always replace the existing one.
It also enhances the algorithm to cope with more metadata types in the future
by just adding a new type, not a lot of code.
Re-applying again due to MSVC 2013 being minimum requirement, and this patch
having C++11 that MSVC 2012 didn't support.
Fixes PR20655.
llvm-svn: 216870
This feeds AA through the IFI structure into the inliner so that
AddAliasScopeMetadata can use AA->getModRefBehavior to figure out which
functions only access their arguments (instead of just hard-coding some
knowledge of memory intrinsics). Most of the information is only available from
BasicAA; this is important for preserving alias scoping information for
target-specific intrinsics when doing the noalias parameter attribute to
metadata conversion.
llvm-svn: 216866
I thought that I had fixed this problem in r216818, but I did not do a very
good job. The underlying issue is that when we add alias.scope metadata we are
asserting that this metadata completely describes the aliasing relationships
within the current aliasing scope domain, and so in the context of translating
noalias argument attributes, the pointers must all be based on noalias
arguments (as underlying objects) and have no other kind of underlying object.
In r216818 excluding appropriate accesses from getting alias.scope metadata is
done by looking for underlying objects that are not identified function-local
objects -- but that's wrong because allocas, etc. are also function-local
objects and we need to explicitly check that all underlying objects are the
noalias arguments for which we're adding metadata aliasing scopes.
This fixes the underlying-object check for adding alias.scope metadata, and
does some refactoring of the related capture-checking eligibility logic (and
adds more comments; hopefully making everything a bit clearer).
Fixes self-hosting on x86_64 with -mllvm -enable-noalias-to-md-conversion (the
feature is still disabled by default).
llvm-svn: 216863
The previous implementation of AddAliasScopeMetadata, which adds noalias
metadata to preserve noalias parameter attribute information when inlining had
a flaw: it would add alias.scope metadata to accesses which might have been
derived from pointers other than noalias function parameters. This was
incorrect because even some access known not to alias with all noalias function
parameters could easily alias with an access derived from some other pointer.
Instead, when deriving from some unknown pointer, we cannot add alias.scope
metadata at all. This fixes a miscompile of the test-suite's tramp3d-v4.
Furthermore, we cannot add alias.scope to functions unless we know they
access only argument-derived pointers (currently, we know this only for
memory intrinsics).
Also, we fix a theoretical problem with using the NoCapture attribute to skip
the capture check. This is incorrect (as explained in the comment added), but
would not matter in any code generated by Clang because we get only inferred
nocapture attributes in Clang-generated IR.
This functionality is not yet enabled by default.
llvm-svn: 216818
consider: (and (icmp X, Y), (and Z, (icmp A, B)))
It may be possible to combine (icmp X, Y) with (icmp A, B).
If we successfully combine, create an 'and' instruction with Z.
This fixes PR20814.
N.B. There is room for improvement after this change but I'm not
convinced it's worth chasing yet.
llvm-svn: 216814
Even loads/stores that have a stronger ordering than monotonic can be safe.
The rule is no release-acquire pair on the path from the QueryInst, assuming that
the QueryInst is not atomic itself.
llvm-svn: 216771
Don't promote byval pointer arguments when when their size in bits is
not equal to their alloc size in bits. This can happen for x86_fp80,
where the size in bits is 80 but the alloca size in bits in 128.
Promoting these types can break passing unions of x86_fp80s and other
types.
Patch by Thomas Jablin!
Reviewed By: rnk
Differential Revision: http://reviews.llvm.org/D5057
llvm-svn: 216693
For a detailed description of the problem see the comment in the test file.
The problematic moveBefore() calls are not required anymore because the new
scheduling algorithm ensures a correct ordering anyway.
llvm-svn: 216656
Several combines involving icmp (shl C2, %X) C1 can be simplified
without introducing any new instructions. Move them to InstSimplify;
while we are at it, make them more powerful.
llvm-svn: 216642
We try to perform this transform in InstSimplify but we aren't always
able to. Sometimes, we need to insert a bitcast if X and Y don't have
the same time.
llvm-svn: 216598
It's incorrect to perform this simplification if the types differ.
A bitcast would need to be inserted for this to work.
This fixes PR20771.
llvm-svn: 216597
Kevin Qin