Summary:
The SLP vectorizer should propagate IR-level optimization hints/flags
(nsw, nuw, exact, fast-math) when converting scalar horizontal
reductions instructions into vectors, just like for other vectorized
instructions.
It doe not include IR propagation for extra arguments, we need to handle
original scalar operations for extra args to propagate correct flags.
Reviewers: mkuper, mzolotukhin, hfinkel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30418
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296614 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
We should preserve IR flags for extra args. These IR flags should be
taken from original scalar operations, not from the reduction
operations.
Reviewers: mkuper, mzolotukhin, hfinkel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30447
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296613 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
If horizontal reduction tree starts from the binary operation that is
used in PHI node, but this PHI is not used in horizontal reduction, we
may end up with extra addition of this PHI node after vectorization.
Here is an example:
```
%phi = phi i32 [ %tmp, %end], ...
...
%tmp = add i32 %tmp1, %tmp2
end:
```
after vectorization we always have something like:
```
%phi = phi i32 [ %tmp, %end], ...
...
%red = extractelement <8 x 32> %vec.red, 0
%tmp = add i32 %red, %phi
end:
```
even if `%phi` is not used in reduction tree. Patch considers these PHI
nodes as extra arguments and considers them in the final result iff they
really used in reduction.
Reviewers: mkuper, hfinkel, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30409
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296606 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
Solves PR 31990.
The bad rewrite could replace a memcpy of one word with
store i4 -1
while it should actually be
store i8 -1
Hopefully opt and llc has improved enough so the original optimization
done by the code isn't needed anymore.
One already existing testcase is affected. It originally tested that
the memcpy was replaced with
load double
but since we now remove that rewrite it will be
load i64
instead.
Patch suggestion by Eli Friedman.
Reviewers: eli.friedman, majnemer, efriedma
Reviewed By: efriedma
Subscribers: efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D30254
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296585 91177308-0d34-0410-b5e6-96231b3b80d8
The practice in LV is that we emit analysis remarks and then finally report
either a missed or applied remark on the final decision whether vectorization
is taking place. On this code path, we were closing with an analysis remark.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296578 91177308-0d34-0410-b5e6-96231b3b80d8
for VectorizeTree() API.This API uses it for proper mask computation to be used in shufflevector IR.
The fix is to compute the mask for out of order memory accesses while building the vectorizable tree
instead of actual vectorization of vectorizable tree.
Reviewers: mkuper
Differential Revision: https://reviews.llvm.org/D30159
Change-Id: Id1e287f073fa4959713ba545fa4254db5da8b40d
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296575 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
Currently, our post-dom tree tries to ignore and remove the effects of
infinite loops. It fails miserably at this, because it tries to do it
ahead of time, and thus can only detect self-loops, and any other type
of infinite loop, it pretends doesn't exist at all.
This can, in a bunch of cases, lead to wrong answers and a completely
empty post-dom tree.
Wrong answer:
```
declare void foo()
define internal void @f() {
entry:
br i1 undef, label %bb35, label %bb3.i
bb3.i:
call void @foo()
br label %bb3.i
bb35.loopexit3:
br label %bb35
bb35:
ret void
}
```
We get:
```
Inorder PostDominator Tree:
[1] <<exit node>> {0,7}
[2] %bb35 {1,6}
[3] %bb35.loopexit3 {2,3}
[3] %entry {4,5}
```
This is a trivial modification of the testcase for PR 6047
Note that we pretend bb3.i doesn't exist.
We also pretend that bb35 post-dominates entry.
While it's true that it does not exit in a theoretical sense, it's not
really helpful to try to ignore the effect and pretend that bb35
post-dominates entry. Worse, we pretend the infinite loop does
nothing (it's usually considered a side-effect), and doesn't even
exist, even when it calls a function. Sadly, this makes it impossible
to use when you are trying to move code safely. All compilers also
create virtual or real single exit nodes (including us), and connect
infinite loops there (which this patch does). In fact, others have
worked around our behavior here, to the point of building their own
post-dom trees:
https://zneak.github.io/fcd/2016/02/17/structuring.html and pointing
out the region infrastructure is near-useless for them with postdom in
this state :(
Completely empty post-dom tree:
```
define void @spam() #0 {
bb:
br label %bb1
bb1: ; preds = %bb1, %bb
br label %bb1
bb2: ; No predecessors!
ret void
}
```
Printing analysis 'Post-Dominator Tree Construction' for function 'foo':
=============================--------------------------------
Inorder PostDominator Tree:
[1] <<exit node>> {0,1}
:(
(note that even if you ignore the effects of infinite loops, bb2
should be present as an exit node that post-dominates nothing).
This patch changes post-dom to properly handle infinite loops and does
root finding during calculation to prevent empty tress in such cases.
We match gcc's (and the canonical theoretical) behavior for infinite
loops (find the backedge, connect it to the exit block).
Testcases coming as soon as i finish running this on a ton of random graphs :)
Reviewers: chandlerc, davide
Subscribers: bryant, llvm-commits
Differential Revision: https://reviews.llvm.org/D29705
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296535 91177308-0d34-0410-b5e6-96231b3b80d8
Summary: For SamplePGO, the profile may contain cross-module inline stacks. As we need to make sure the profile annotation happens when all the hot inline stacks are expanded, we need to pass this info to the module importer so that it can import proper functions if necessary. This patch implemented this feature by emitting cross-module targets as part of function entry metadata. In the module-summary phase, the metadata is used to build call edges that points to functions need to be imported.
Reviewers: mehdi_amini, tejohnson
Reviewed By: tejohnson
Subscribers: davidxl, llvm-commits
Differential Revision: https://reviews.llvm.org/D30053
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296498 91177308-0d34-0410-b5e6-96231b3b80d8
The LLVM backend cannot produce any debug info for an llvm::Function
without a DISubprogram attachment. When inlining a debug-info-carrying
function into a nodebug function, there is therefore no reason to keep
any debug info intrinsic calls or debug locations on the instructions.
This fixes a problem discovered in PR32042.
rdar://problem/30679307
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296488 91177308-0d34-0410-b5e6-96231b3b80d8
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296416 91177308-0d34-0410-b5e6-96231b3b80d8
We may get a VL where the first element is a load, but the others
aren't. Trying to sort such VLs can only lead to sorrow.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296411 91177308-0d34-0410-b5e6-96231b3b80d8
This was suggested in D27855: have the inliner add assumptions, so we don't
lose nonnull info provided by argument attributes.
This still doesn't solve PR28430 (dyn_cast), but this gets us closer.
https://reviews.llvm.org/D29999
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296366 91177308-0d34-0410-b5e6-96231b3b80d8
This is a fix for a loop predication bug which resulted in malformed IR generation.
Loop invariant side of the widened condition is not guaranteed to be available in the preheader as is, so we need to expand it as well. See added unsigned_loop_0_to_n_hoist_length test for example.
Reviewed By: sanjoy, mkazantsev
Differential Revision: https://reviews.llvm.org/D30099
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296345 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
Previously we used to return a bogus result, 0, for IR like `ashr %val,
-1`.
I've also added an assert checking that `ComputeNumSignBits` at least
returns 1. That assert found an already checked in test case where we
were returning a bad result for `ashr %val, -1`.
Fixes PR32045.
Reviewers: spatel, majnemer
Reviewed By: spatel, majnemer
Subscribers: efriedma, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D30311
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296273 91177308-0d34-0410-b5e6-96231b3b80d8
Current internal option -static-func-full-module-prefix keeps all the
directory path the profile counter names for static functions. The default
of this option is false. This strips the directory names from the source
filename which is problematic:
(1) it creates linker errors for profile-generation compilation, exposed in
our internal benchmarks. We are seeing messages like
"warning: relocation refers to discarded section".
This is due to the name conflicts after the stripping.
(2) the stripping only applies to getPGOFuncName.
Current Thin-LTO module importing for the indirect-calls assumes
the source directory name not being stripped. Current default value
for this option can potentially prevent some inter-module
indirect-call-promotions.
This patch turns the default value for -static-func-full-module-prefix to true.
The second part of the patch is to have an alternative implementation under
the internal option -static-func-strip-dirname-prefix=<value>
This options specifies level of directories to be stripped from the source
filename. Using a large value as the parameter has the same effect as
-static-func-full-module-prefix.
Differential Revision: http://reviews.llvm.org/D29512
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296206 91177308-0d34-0410-b5e6-96231b3b80d8
When we construct addressing modes, we use isNoopAddrSpaceCast to ignore
addrspacecast instructions. Make sure we insert the correct addrspacecast
when we reconstruct the addressing mode.
Differential Revision: https://reviews.llvm.org/D30114
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296167 91177308-0d34-0410-b5e6-96231b3b80d8
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296149 91177308-0d34-0410-b5e6-96231b3b80d8
This patch merges the existing floating-point induction variable widening code
into the integer induction variable widening code, creating a single set of
functions for both kinds of inductions. The primary motivation for doing this
is to enable vector phi node creation for floating-point induction variables.
Differential Revision: https://reviews.llvm.org/D30211
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296145 91177308-0d34-0410-b5e6-96231b3b80d8
The Fuchsia ABI defines slots from the thread pointer where the
stack-guard value for stack-protector, and the unsafe stack pointer
for safe-stack, are stored. This parallels the Android ABI support.
Patch by Roland McGrath
Differential Revision: https://reviews.llvm.org/D30237
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296081 91177308-0d34-0410-b5e6-96231b3b80d8
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296060 91177308-0d34-0410-b5e6-96231b3b80d8
Summary: In case we do not know what the condition is in an unswitched loop, but we know its definitely NOT a known constant. We can perform simplifcations based on this information.
Reviewers: sanjoy, hfinkel, chenli, efriedma
Reviewed By: efriedma
Subscribers: david2050, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D28968
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296041 91177308-0d34-0410-b5e6-96231b3b80d8
While not CVP's fault, this caused miscompiles (PR31181). Reverting
until those are resolved.
(This also reverts the follow-ups r288154 and r288161 which removed the
flag.)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296030 91177308-0d34-0410-b5e6-96231b3b80d8
Summary: SamplePGO uses branch_weight annotation to represent callsite hotness. When ICP promotes an indirect call to direct call, we need to make sure the direct call is annotated with branch_weight in SamplePGO mode, so that downstream function inliner can use hot callsite heuristic.
Reviewers: davidxl, eraman, xur
Reviewed By: davidxl, xur
Subscribers: mehdi_amini, llvm-commits
Differential Revision: https://reviews.llvm.org/D30282
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296028 91177308-0d34-0410-b5e6-96231b3b80d8
In OptimizeAdd, we scan the operand list to see if there are any common factors
between operands that can be factored out to reduce the number of multiplies
(e.g., 'A*A+A*B*C+D' -> 'A*(A+B*C)+D'). For each operand of the operand list, we
only consider unique factors (which is tracked by the Duplicate set). Now if we
find a factor that is a negative constant, we add the negated value as a factor
as well, because we can percolate the negate out. However, we mistakenly don't
add this negated constant to the Duplicates set.
Consider the expression A*2*-2 + B. Obviously, nothing to factor.
For the added value A*2*-2 we over count 2 as a factor without this change,
which causes the assert reported in PR30256. The problem is that this code is
assuming that all the multiply operands of the add are already reassociated.
This change avoids the issue by making OptimizeAdd tolerate multiplies which
haven't been completely optimized; this sort of works, but we're doing wasted
work: we'll end up revisiting the add later anyway.
Another possible approach would be to enforce RPO iteration order more strongly.
If we have RedoInsts, we process them immediately in RPO order, rather than
waiting until we've finished processing the whole function. Intuitively, it
seems like the natural approach: reassociation works on expression trees, so
the optimization only works in one direction. That said, I'm not sure how
practical that is given the current Reassociate; the "optimal" form for an
expression depends on its use list (see all the uses of "user_back()"), so
Reassociate is really an iterative optimization of sorts, so any changes here
would probably get messy.
PR30256
Differential Revision: https://reviews.llvm.org/D30228
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296003 91177308-0d34-0410-b5e6-96231b3b80d8
Summary: The discriminator has been encoded, and only the base discriminator should be used during profile matching.
Reviewers: dblaikie, davidxl
Reviewed By: dblaikie, davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30218
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295999 91177308-0d34-0410-b5e6-96231b3b80d8
result
Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
for (int x = 0; x < 8; x++) {
val = val + input[y * 8 + x] + 3;
}
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.
Reviewers: mkuper, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30262
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295972 91177308-0d34-0410-b5e6-96231b3b80d8
result
Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
for (int x = 0; x < 8; x++) {
val = val + input[y * 8 + x] + 3;
}
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.
Reviewers: mkuper, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30262
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295956 91177308-0d34-0410-b5e6-96231b3b80d8
result
Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
for (int x = 0; x < 8; x++) {
val = val + input[y * 8 + x] + 3;
}
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.
Reviewers: mkuper, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30262
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295949 91177308-0d34-0410-b5e6-96231b3b80d8
Implement isLegalToVectorizeLoadChain for AMDGPU to avoid
producing private address spaces accesses that will need to be
split up later. This was doing the wrong thing in the case
where the queried chain was an even number of elements.
A possible <4 x i32> store was being split into
store <2 x i32>
store i32
store i32
rather than
store <2 x i32>
store <2 x i32>
when legal.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295933 91177308-0d34-0410-b5e6-96231b3b80d8
Summary:
Depends on D29606 and D29682
Makes us pass GVN's edge.ll (we also will pass a few other testcases
they just need cleaning up).
Thoughts on the Predicate* hiearchy of classes especially welcome :)
(it's not clear to me how best to organize it, and currently, the getBlock* seems ... uglier than maybe wasting a field somewhere or something).
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29747
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295889 91177308-0d34-0410-b5e6-96231b3b80d8
Hans Wennborg