SimplifyTerminatorOnSelect didn't consider the possibility that the
condition might be related to one of PHI nodes.
This fixes PR25267.
llvm-svn: 250922
In some cases (as illustrated in the unittest), lineno can be less than the heade_lineno because the function body are included from some other files. In this case, offset will be negative. This patch makes clang still able to match the profile to IR in this situation.
http://reviews.llvm.org/D13914
llvm-svn: 250873
`normalizeForInvokeSafepoint` in RewriteStatepointsForGC.cpp, as it is
written today, deals with `gc.relocate` and `gc.result` uses of a
statepoint equally well. This change documents this fact and adds a
test case.
There is no functional change here -- only documentation of existing
functionality.
llvm-svn: 250784
Allow LLVM to optimize the sequence like the following:
%inc = add nsw i32 %i, 1
%cmp = icmp slt %n, %inc
into:
%cmp = icmp sle i32 %n, %i
The case is not handled previously due to the complexity of compuation of %n.
Hence, LLVM cannot swap operands of icmp accordingly.
llvm-svn: 250746
This was originally checked in at r250527, but reverted at r250570 because of PR25222.
There were at least 2 problems:
1. The cost check was checking for an instruction with an exact cost of TCC_Expensive;
that should have been >=.
2. The cause of the clang stage 1 failures was illegally sinking 'call' instructions;
we can't sink instructions that may have side effects / are not safe to execute speculatively.
Fixed those conditions in sinkSelectOperand() and added test cases.
Original commit message:
This is a follow-up to the discussion in D12882.
Ideally, we would like SimplifyCFG to be able to form select instructions even when the operands
are expensive (as defined by the TTI cost model) because that may expose further optimizations.
However, we would then like a later pass like CodeGenPrepare to undo that transformation if the
target would likely benefit from not speculatively executing an expensive op (this patch).
Once we have this safety mechanism in place, we can adjust SimplifyCFG to restore its
select-formation behavior that changed with r248439.
Differential Revision: http://reviews.llvm.org/D13297
llvm-svn: 250743
This patch improves support for combining the SSE4A EXTRQ(I) and INSERTQ(I) intrinsics:
1 - Converts INSERTQ/EXTRQ calls to INSERTQI/EXTRQI if the 'bit index' and 'length' operands are constant
2 - Converts INSERTQI/EXTRQI calls to shufflevector if the bit index/length are both byte aligned (we can already lower shuffles to INSERTQI/EXTRQI if its useful)
3 - Constant folding support
4 - Add zeroinitializer handling
Differential Revision: http://reviews.llvm.org/D13348
llvm-svn: 250609
The number of samples collected at the head of a function only make
sense for top-level functions (i.e., those actually called as opposed to
being inlined inside another).
Head samples essentially count the time spent inside the function's
prologue. This clearly doesn't make sense for inlined functions, so we
were always emitting 0 in those.
llvm-svn: 250539
Ideally, we would like SimplifyCFG to be able to form select instructions even when the operands
are expensive (as defined by the TTI cost model) because that may expose further optimizations.
However, we would then like a later pass like CodeGenPrepare to undo that transformation if the
target would likely benefit from not speculatively executing an expensive op (this patch).
Once we have this safety mechanism in place, we can adjust SimplifyCFG to restore its
select-formation behavior that changed with r248439.
Differential Revision: http://reviews.llvm.org/D13297
llvm-svn: 250527
The `"statepoint-id"` and `"statepoint-num-patch-bytes"` attributes are
used solely to determine properties of the `gc.statepoint` being
created. Once the `gc.statepoint` is in place, these should be removed.
llvm-svn: 250491
Summary:
This is a step towards using operand bundles to carry deopt state till
RewriteStatepointsForGC. The change adds a flag to
RewriteStatepointsForGC that teaches it to pick up deopt state from a
`"deopt"` operand bundle attached to the `call` or `invoke` it is
wrapping.
The command line flag added, `-rs4gc-use-deopt-bundles`, will only exist
for a short while. Once we are able to pipe deopt bundle state through
the full optimization pipeline without problems, we will "constant fold"
`-rs4gc-use-deopt-bundles` to `true`.
Reviewers: swaroop.sridhar, reames
Subscribers: llvm-commits, sanjoy
Differential Revision: http://reviews.llvm.org/D13372
llvm-svn: 250489
Summary:
`cloneArithmeticIVUser` currently trips over expression like `add %iv,
-1` when `%iv` is being zero extended -- it tries to construct the
widened use as `add %iv.zext, zext(-1)` and (correctly) fails to prove
equivalence to `zext(add %iv, -1)` (here the SCEV for `%iv` is
`{1,+,1}`).
This change teaches `IndVars` to try sign extending the non-IV operand
if that makes the newly constructed IV use equivalent to the widened
narrow IV use.
Reviewers: atrick, hfinkel, reames
Subscribers: sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D13717
llvm-svn: 250483
Android libc provides a fixed TLS slot for the unsafe stack pointer,
and this change implements direct access to that slot on AArch64 via
__builtin_thread_pointer() + offset.
This change also moves more code into TargetLowering and its
target-specific subclasses to get rid of target-specific codegen
in SafeStackPass.
This change does not touch the ARM backend because ARM lowers
builting_thread_pointer as aeabi_read_tp, which is not available
on Android.
llvm-svn: 250456
Turns out this approach is buggy. In discussion about follow on work, Sanjoy pointed out that we could be subject to circular logic problems.
Consider:
if (i u< L) leave()
if ((i + 1) u< L) leave()
print(a[i] + a[i+1])
If we know that L is less than UINT_MAX, we could possible prove (in a control dependent way) that i + 1 does not overflow. This gives us:
if (i u< L) leave()
if ((i +nuw 1) u< L) leave()
print(a[i] + a[i+1])
If we now do the transform this patch proposed, we end up with:
if ((i +nuw 1) u< L) leave_appropriately()
print(a[i] + a[i+1])
That would be a miscompile when i==-1. The problem here is that the control dependent nuw bits got used to prove something about the first condition. That's obviously invalid.
This won't happen today, but since I plan to enhance LVI/CVP with exactly that transform at some point in the not too distant future...
llvm-svn: 250430
With r250345 and r250343, we start to observe the following failure
when bootstrap clang with lto and pgo:
PHI node entries do not match predecessors!
%.sroa.029.3.i = phi %"class.llvm::SDNode.13298"* [ null, %30953 ], [ null, %31017 ], [ null, %30998 ], [ null, %_ZN4llvm8dyn_castINS_14ConstantSDNodeENS_7SDValueEEENS_10cast_rettyIT_T0_E8ret_typeERS5_.exit.i.1804 ], [ null, %30975 ], [ null, %30991 ], [ null, %_ZNK4llvm3EVT13getScalarTypeEv.exit.i.1812 ], [ %..sroa.029.0.i, %_ZN4llvm11SmallVectorIiLj8EED1Ev.exit.i.1826 ], !dbg !451895
label %30998
label %_ZNK4llvm3EVTeqES0_.exit19.thread.i
LLVM ERROR: Broken function found, compilation aborted!
I will re-commit this if the bot does not recover.
llvm-svn: 250366
Currently in JumpThreading pass, the branch weight metadata is not updated after CFG modification. Consider the jump threading on PredBB, BB, and SuccBB. After jump threading, the weight on BB->SuccBB should be adjusted as some of it is contributed by the edge PredBB->BB, which doesn't exist anymore. This patch tries to update the edge weight in metadata on BB->SuccBB by scaling it by 1 - Freq(PredBB->BB) / Freq(BB->SuccBB).
This is the third attempt to submit this patch, while the first two led to failures in some FDO tests. After investigation, it is the edge weight normalization that caused those failures. In this patch the edge weight normalization is fixed so that there is no zero weight in the output and the sum of all weights can fit in 32-bit integer. Several unit tests are added.
Differential revision: http://reviews.llvm.org/D10979
llvm-svn: 250345
This is a cleaned up patch from the one written by John Regehr based on the findings of the Souper superoptimizer.
The basic idea here is that input bits that are known zero reduce the maximum count that the intrinsic could return. We know that the number of bits required to represent a particular count is at most log2(N)+1.
Differential Revision: http://reviews.llvm.org/D13253
llvm-svn: 250338
Binary encoded profiles used to encode all function names inline at
every reference. This is clearly suboptimal in terms of space. This
patch fixes this by adding a name table to the header of the file.
llvm-svn: 250241
We forgot to append the terminatepad's arguments which resulted in us
treating the old terminatepad as an argument to the new terminatepad
causing us to crash immediately. Instead, add the old terminatepad's
arguments to the new terminatepad.
This fixes PR25155.
llvm-svn: 250234
As discussed in D13348 - the INSERTQI range combining code is wrong in that it confuses the insertion bit index with an extraction bit index.
The remaining legal combines are very unlikely (especially once we've converted to shuffles in D13348) so I'm removing the optimization.
llvm-svn: 250160
In JumpThreading pass, the branch weight metadata is not updated after CFG modification. Consider the jump threading on PredBB, BB, and SuccBB. After jump threading, the weight on BB->SuccBB should be adjusted as some of it is contributed by the edge PredBB->BB, which doesn't exist anymore. This patch tries to update the edge weight in metadata on BB->SuccBB by scaling it by 1 - Freq(PredBB->BB) / Freq(BB->SuccBB).
Differential revision: http://reviews.llvm.org/D10979
llvm-svn: 250089
GlobalOpt currently merges stores into the initialisers of internal,
externally_initialized globals, but should not do so as the value of the global
may change between the initialiser and any code in the module being run.
llvm-svn: 250035
C semantics force sub-int-sized values (e.g. i8, i16) to be promoted to int
type (e.g. i32) whenever arithmetic is performed on them.
For targets with native i8 or i16 operations, usually InstCombine can shrink
the arithmetic type down again. However InstCombine refuses to create illegal
types, so for targets without i8 or i16 registers, the lengthening and
shrinking remains.
Most SIMD ISAs (e.g. NEON) however support vectors of i8 or i16 even when
their scalar equivalents do not, so during vectorization it is important to
remove these lengthens and truncates when deciding the profitability of
vectorization.
The algorithm this uses starts at truncs and icmps, trawling their use-def
chains until they terminate or instructions outside the loop are found (or
unsafe instructions like inttoptr casts are found). If the use-def chains
starting from different root instructions (truncs/icmps) meet, they are
unioned. The demanded bits of each node in the graph are ORed together to form
an overall mask of the demanded bits in the entire graph. The minimum bitwidth
that graph can be truncated to is the bitwidth minus the number of leading
zeroes in the overall mask.
The intention is that this algorithm should "first do no harm", so it will
never insert extra cast instructions. This is why the use-def graphs are
unioned, so that subgraphs with different minimum bitwidths do not need casts
inserted between them.
This algorithm works hard to reduce compile time impact. DemandedBits are only
queried if there are extends of illegal types and if a truncate to an illegal
type is seen. In the general case, this results in a simple linear scan of the
instructions in the loop.
No non-noise compile time impact was seen on a clang bootstrap build.
llvm-svn: 250032
Doing so could cause the post-unswitching convergent ops to be
control-dependent on the unswitch condition where they were not before.
This check could be refined to allow unswitching where the convergent
operation was already control-dependent on the unswitch condition.
llvm-svn: 249874
With this patch we can now read and write inline stacks in sample
profiles to the binary encoded profiles.
In a subsequent patch, I will add a string table to the binary encoding.
Right now function names are emitted as strings every time we find them.
This is too bloated and will produce large files in applications with
lots of inlining.
llvm-svn: 249861
Pass MemCpyOpt doesn't check if a store instruction is nontemporal.
As a consequence, adjacent nontemporal stores are always merged into a
memset call.
Example:
;;;
define void @foo(<4 x float>* nocapture %p) {
entry:
store <4 x float> zeroinitializer, <4 x float>* %p, align 16, !nontemporal !0
%p1 = getelementptr inbounds <4 x float>, <4 x float>* %dst, i64 1
store <4 x float> zeroinitializer, <4 x float>* %p1, align 16, !nontemporal !0
ret void
}
!0 = !{i32 1}
;;;
In this example, the two nontemporal stores are combined to a memset of zero
which does not preserve the nontemporal hint. Later on the backend (tested on a
x86-64 corei7) expands that memset call into a sequence of two normal 16-byte
aligned vector stores.
opt -memcpyopt example.ll -S -o - | llc -mcpu=corei7 -o -
Before:
xorps %xmm0, %xmm0
movaps %xmm0, 16(%rdi)
movaps %xmm0, (%rdi)
With this patch, we no longer merge nontemporal stores into calls to memset.
In this example, llc correctly expands the two stores into two movntps:
xorps %xmm0, %xmm0
movntps %xmm0, 16(%rdi)
movntps %xmm0, (%rdi)
In theory, we could extend the usage of !nontemporal metadata to memcpy/memset
calls. However a change like that would only have the effect of forcing the
backend to expand !nontemporal memsets back to sequences of store instructions.
A memset library call would not have exactly the same semantic of a builtin
!nontemporal memset call. So, SelectionDAG will have to conservatively expand
it back to a sequence of !nontemporal stores (effectively undoing the merging).
Differential Revision: http://reviews.llvm.org/D13519
llvm-svn: 249820
Summary:
These non-semantic changes will help make a later change adding
support for deopt operand bundles more streamlined.
Reviewers: reames, swaroop.sridhar
Subscribers: sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D13491
llvm-svn: 249779
Summary:
`getConstantEvolutionLoopExitValue` and `ComputeExitCountExhaustively`
assumed all phi nodes in the loop header have the same order of incoming
values. This is not correct, and this commit changes
`getConstantEvolutionLoopExitValue` and `ComputeExitCountExhaustively`
to lookup the backedge value of a phi node using the loop's latch block.
Unfortunately, there is still some code duplication
`getConstantEvolutionLoopExitValue` and `ComputeExitCountExhaustively`.
At some point in the future we should extract out a helper class /
method that can evolve constant evolution phi nodes across iterations.
Fixes 25060. Thanks to Mattias Eriksson for the spot-on analysis!
Depends on D13457.
Reviewers: atrick, hfinkel
Subscribers: materi, llvm-commits
Differential Revision: http://reviews.llvm.org/D13458
llvm-svn: 249712
This is a partial fix for PR24886:
https://llvm.org/bugs/show_bug.cgi?id=24886
Without this IR transform, the backend (x86 at least) was producing inefficient code.
This patch is making 2 assumptions:
1. The canonical form of a fabs() operation is, in fact, the LLVM fabs() intrinsic.
2. The high bit of an FP value is always the sign bit; as noted in the bug report, this isn't specified by the LangRef.
Differential Revision: http://reviews.llvm.org/D13076
llvm-svn: 249702
This was requested in D13076: if we're going to canonicalize to fabs(), ValueTracking
should know that fabs() clears sign bits.
In this patch (as in D13076), we're not handling vectors yet even though computeKnownBits'
fabs() case itself should be vector-ready via the splat in this patch.
Fixing this will require follow-on patches to correct other logic that uses 'getScalarType'.
Differential Revision: http://reviews.llvm.org/D13222
llvm-svn: 249701
Summary:
After r249211, SCEV can see through some LCSSA phis. Add a
`replacementPreservesLCSSAForm` check before replacing uses of these phi
nodes with a simplified use of the induction variable to avoid breaking
LCSSA.
Fixes 25047.
Depends on D13460.
Reviewers: atrick, hfinkel
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D13461
llvm-svn: 249575
Summary:
Some target intrinsics can access multiple elements, using the pointer as a
base address (e.g. AArch64 ld4). When trying to CSE such instructions,
it must be checked the available value comes from a compatible instruction
because the pointer is not enough to discriminate whether the value is
correct.
Reviewers: ssijaric
Subscribers: mcrosier, llvm-commits, aemerson
Differential Revision: http://reviews.llvm.org/D13475
llvm-svn: 249523
This will allow us to optimize code such as:
int f(int *p) {
int x;
return p == &x;
}
as well as:
int *allocate(void);
int f() {
int x;
int *p = allocate();
return p == &x;
}
The folding can only be done under certain circumstances. Even though p and &x
cannot alias, the comparison must still return true if the pointer
representations are equal. If a user successfully generates a p that's a
correct guess for &x, comparison should return true even though p is an invalid
pointer.
This patch argues that if the address of the alloca isn't observable outside the
function, the function can act as-if the address is impossible to guess from the
outside. The tricky part is keeping the act consistent: if we fold p == &x to
false in one place, we must make sure to fold any other comparisons based on
those pointers similarly. To ensure that, we only fold when &x is involved
exactly once in comparison instructions.
Differential Revision: http://reviews.llvm.org/D13358
llvm-svn: 249490
Summary:
After r249211, `getSCEV(X) == getSCEV(Y)` does not guarantee that X and
Y are related in the dominator tree, even if X is an operand to Y (I've
included a toy example in comments, and a real example as a test case).
This commit changes `SimplifyIndVar` to require a `DominatorTree`. I
don't think this is a problem because `ScalarEvolution` requires it
anyway.
Fixes PR25051.
Depends on D13459.
Reviewers: atrick, hfinkel
Subscribers: joker.eph, llvm-commits, sanjoy
Differential Revision: http://reviews.llvm.org/D13460
llvm-svn: 249471
This is a cleaned up patch from the one written by John Regehr based on the findings of the Souper superoptimizer.
When writing tests, I was surprised to find that instsimplify apparently doesn't know how to collapse bit test sequences based purely on known bits. This required me to split my tests across both instsimplify and instcombine.
Differential Revision: http://reviews.llvm.org/D13250
llvm-svn: 249453
As mentioned in the bug, I'd missed the presence of a getScalarType in the caller of the new implies method. As a result, when we ended up with a implication over two vectors, we'd trip an assert and crash.
Differential Revision: http://reviews.llvm.org/D13441
llvm-svn: 249442
If the mask of a select instruction is a ConstantVector, method
SimplifyDemandedVectorElts iterates over the mask elements to identify which
values are selected from the select inputs.
Before this patch, method SimplifyDemandedVectorElts always used method
Constant::isNullValue() to check if a value in the mask was zero. Unfortunately
that method always returns false when called on a ConstantExpr.
This patch fixes the problem in SimplifyDemandedVectorElts by adding an explicit
check for ConstantExpr values. Now, if a value in the mask is a ConstantExpr, we
avoid calling isNullValue() on it.
Fixes PR24922.
Differential Revision: http://reviews.llvm.org/D13219
llvm-svn: 249390