We're sinking stores, which is a good thing, but in the process creating selects for the store address operand, which SROA/Mem2Reg can't look through, which caused serious regressions.
The real fix is in SROA, which I'll be looking into.
llvm-svn: 280470
r280425 | dehao | 2016-09-01 16:15:50 -0700 (Thu, 01 Sep 2016) | 9 lines
Refactor LICM pass in preparation for LoopSink pass.
Summary: LoopSink pass uses some common function in LICM. This patch refactor the LICM code to make it usable by LoopSink pass (https://reviews.llvm.org/D22778).
r280429 | dehao | 2016-09-01 16:31:25 -0700 (Thu, 01 Sep 2016) | 9 lines
Refactor LICM to expose canSinkOrHoistInst to LoopSink pass.
Summary: LoopSink pass shares the same canSinkOrHoistInst functionality with LICM pass. This patch exposes this function in preparation of https://reviews.llvm.org/D22778
llvm-svn: 280453
Summary: LoopSink pass shares the same canSinkOrHoistInst functionality with LICM pass. This patch exposes this function in preparation of https://reviews.llvm.org/D22778
Reviewers: chandlerc, davidxl, danielcdh
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24171
llvm-svn: 280429
Summary: This is in preparation for LoopSink pass which calls replaceDominatedUsesWith to update after sinking.
Reviewers: chandlerc, davidxl, danielcdh
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24170
llvm-svn: 280427
Summary: LoopSink pass uses some common function in LICM. This patch refactor the LICM code to make it usable by LoopSink pass (https://reviews.llvm.org/D22778).
Reviewers: chandlerc, davidxl, danielcdh
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24168
llvm-svn: 280425
We can now maintain scalar values in VectorLoopValueMap. Thus, we no longer
have to create temporary vectors with insertelement instructions when handling
pointer induction variables. This case was mistakenly missed from r279649 when
refactoring the other scalarization code.
llvm-svn: 280405
This patch moves the allocation of VectorParts for PHI nodes into the actual
PHI widening code. Previously, we allocated these VectorParts in
vectorizeBlockInLoop, and passed them by reference to widenPHIInstruction. Upon
returning, we would then map the VectorParts in VectorLoopValueMap. This
behavior is problematic for the cases where we only want to generate a scalar
version of a PHI node. For example, if in the future we only generate a scalar
version of an induction variable, we would end up inserting an empty vector
entry into the map once we return to vectorizeBlockInLoop. We now no longer
need to pass VectorParts to the various PHI widening functions, and we can keep
VectorParts allocation as close as possible to the point at which they are
actually mapped in VectorLoopValueMap.
llvm-svn: 280390
Previous change broke the C API for creating an EarlyCSE pass w/
MemorySSA by adding a bool parameter to control whether MemorySSA was
used or not. This broke the OCaml bindings. Instead, change the old C
API entry point back and add a new one to request an EarlyCSE pass with
MemorySSA.
llvm-svn: 280379
While removing a scalar shackle from an icmp fold, I noticed that I couldn't find any tests to trigger
this code path.
The 'and' shrinking transform should be handled by InstCombiner::foldCastedBitwiseLogic()
or eliminated with InstSimplify. The icmp narrowing is part of InstCombiner::foldICmpWithCastAndCast().
Differential Revision: https://reviews.llvm.org/D24031
llvm-svn: 280370
This was a real restriction in the original version of SinkIfThenCodeToEnd. Now it's been rewritten, the restriction can be lifted.
As part of this, we handle a very common and useful case where one of the incoming branches is actually conditional. Consider:
if (a)
x(1);
else if (b)
x(2);
This produces the following CFG:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ | /
[ end ]
[end] has two unconditional predecessor arcs and one conditional. The conditional refers to the implicit empty 'else' arc. This same pattern can also be caused by an empty default block in a switch.
We can't sink the call to x() down to end because no call to x() happens on the third incoming arc (assume that x() has sideeffects for the sake of argument; if something is safe to speculate we could indeed sink nevertheless but this cannot happen in the general case and causes many extra selects).
We are now able to detect this case and split off the unconditional arcs to a common successor:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ / |
[sink.split] |
\ /
[ end ]
Now we can sink the call to x() into %sink.split. This can cause significant code simplification in many testcases.
llvm-svn: 280364
r279460 rewrote this function to be able to handle more than two incoming edges and took pains to ensure this didn't regress anything.
This time we change the logic for determining if an instruction should be sunk. Previously we used a single pass greedy algorithm - sink instructions until one requires more than one PHI node or we run out of instructions to sink.
This had the problem that sinking instructions that had non-identical but trivially the same operands needed extra logic so we sunk them aggressively. For example:
%a = load i32* %b %d = load i32* %b
%c = gep i32* %a, i32 0 %e = gep i32* %d, i32 1
Sinking %c and %e would naively require two PHI merges as %a != %d. But the loads are obviously equivalent (and maybe can't be hoisted because there is no common predecessor).
This is why we implemented the fairly complex function areValuesTriviallySame(), to look through trivial differences like this. However it's just not clever enough.
Instead, throw areValuesTriviallySame away, use pointer equality to check equivalence of operands and switch to a two-stage algorithm.
In the "scan" stage, we look at every sinkable instruction in isolation from end of block to front. If it's sinkable, we keep track of all operands that required PHI merging.
In the "sink" stage, we iteratively sink the last non-terminator in the source blocks. But when calculating how many PHIs are actually required to be inserted (to work out if we should stop or not) we remove any values that have already been sunk from the set of PHI-merges required, which allows us to be more aggressive.
This turns an algorithm with potentially recursive lookahead (looking through GEPs, casts, loads and any other instruction potentially not CSE'd) to two linear scans.
llvm-svn: 280351
We iterate over the result from SafeToMergeTerminators, so make it a SmallSetVector instead of a SmallPtrSet.
Should fix stage3 convergence builds.
llvm-svn: 280342
A very important case is not handled here: multiple arcs to a single block with a PHI. Consider:
a:
%1 = icmp %b, 1
br %1, label %c, label %e
c:
%2 = icmp %b, 2
br %2, label %d, label %e
d:
br %e
e:
phi [0, %a], [1, %c], [2, %d]
FoldValueComparisonIntoPredecessors will refuse to fold this, as it doesn't know how to deal with two arcs to a common destination with different PHI values. The answer is obvious - just split all conflicting arcs.
llvm-svn: 280338
-fprofile-dir=path allows the user to specify where .gcda files should be
emitted when the program is run. In particular, this is the first flag that
causes the .gcno and .o files to have different paths, LLVM is extended to
support this. -fprofile-dir= does not change the file name in the .gcno (and
thus where lcov looks for the source) but it does change the name in the .gcda
(and thus where the runtime library writes the .gcda file). It's different from
a GCOV_PREFIX because a user can observe that the GCOV_PREFIX_STRIP will strip
paths off of -fprofile-dir= but not off of a supplied GCOV_PREFIX.
To implement this we split -coverage-file into -coverage-data-file and
-coverage-notes-file to specify the two different names. The !llvm.gcov
metadata node grows from a 2-element form {string coverage-file, node dbg.cu}
to 3-elements, {string coverage-notes-file, string coverage-data-file, node
dbg.cu}. In the 3-element form, the file name is already "mangled" with
.gcno/.gcda suffixes, while the 2-element form left that to the middle end
pass.
llvm-svn: 280306
This is prep work before changing the callers to also use APInt which will
allow folds for splat vectors. Currently, the callers have ConstantInt
guards in place, so no functional change intended with this commit.
llvm-svn: 280282
Summary:
Current implementation of LI verifier isn't ideal and fails to detect
some cases when LI is incorrect. For instance, it checks that all
recorded loops are in a correct form, but it has no way to check if
there are no more other (unrecorded in LI) loops in the function. This
patch adds a way to detect such bugs.
Reviewers: chandlerc, sanjoy, hfinkel
Subscribers: llvm-commits, silvas, mzolotukhin
Differential Revision: https://reviews.llvm.org/D23437
llvm-svn: 280280
Summary:
Use MemorySSA, if requested, to do less conservative memory dependency
checking.
This change doesn't enable the MemorySSA enhanced EarlyCSE in the
default pipelines, so should be NFC.
Reviewers: dberlin, sanjoy, reames, majnemer
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D19821
llvm-svn: 280279
Passing the types/opcode check still doesn't guarantee we'll actually vectorize.
Therefore, just make it clear we're attempting to vectorize.
llvm-svn: 280263
This is a first step towards supporting deopt value lowering and reporting entirely with the register allocator. I hope to build on this in the near future to support live-on-return semantics, but I have a use case which allows me to test and investigate code quality with just the live-in semantics so I've chosen to start there. For those curious, my use cases is our implementation of the "__llvm_deoptimize" function we bind to @llvm.deoptimize. I'm choosing not to hard code that fact in the patch and instead make it configurable via function attributes.
The basic approach here is modelled on what is done for the "Live In" values on stackmaps and patchpoints. (A secondary goal here is to remove one of the last barriers to merging the pseudo instructions.) We start by adding the operands directly to the STATEPOINT SDNode. Once we've lowered to MI, we extend the remat logic used by the register allocator to fold virtual register uses into StackMap::Indirect entries as needed. This does rely on the fact that the register allocator rematerializes. If it didn't along some code path, we could end up with more vregs than physical registers and fail to allocate.
Today, we *only* fold in the register allocator. This can create some weird effects when combined with arguments passed on the stack because we don't fold them appropriately. I have an idea how to fix that, but it needs this patch in place to work on that effectively. (There's some weird interaction with the scheduler as well, more investigation needed.)
My near term plan is to land this patch off-by-default, experiment in my local tree to identify any correctness issues and then start fixing codegen problems one by one as I find them. Once I have the live-in lowering fully working (both correctness and code quality), I'm hoping to move on to the live-on-return semantics. Note: I don't have any *known* miscompiles with this patch enabled, but I'm pretty sure I'll find at least a couple. Thus, the "experimental" tag and the fact it's off by default.
Differential Revision: https://reviews.llvm.org/D24000
llvm-svn: 280250
We check that a sinking candidate is used by only one PHI node during our legality checks. However for instructions that are used by other sinking candidates our heuristic is less conservative. This can result in a candidate actually being illegal when we come to sink it because of how we sunk a predecessor. Do the used-by-only-one-PHI checks again during sinking to ensure we don't crash.
llvm-svn: 280228
We're sinking stores, which is a good thing, but in the process creating selects for the store address operand, which SROA/Mem2Reg can't look through, which caused serious regressions.
The real fix is in SROA, which I'll be looking into.
llvm-svn: 280219
A very important case is not handled here: multiple arcs to a single block with a PHI. Consider:
a:
%1 = icmp %b, 1
br %1, label %c, label %e
c:
%2 = icmp %b, 2
br %2, label %d, label %e
d:
br %e
e:
phi [0, %a], [1, %c], [2, %d]
FoldValueComparisonIntoPredecessors will refuse to fold this, as it doesn't know how to deal with two arcs to a common destination with different PHI values. The answer is obvious - just split all conflicting arcs.
llvm-svn: 280218
This was a real restriction in the original version of SinkIfThenCodeToEnd. Now it's been rewritten, the restriction can be lifted.
As part of this, we handle a very common and useful case where one of the incoming branches is actually conditional. Consider:
if (a)
x(1);
else if (b)
x(2);
This produces the following CFG:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ | /
[ end ]
[end] has two unconditional predecessor arcs and one conditional. The conditional refers to the implicit empty 'else' arc. This same pattern can also be caused by an empty default block in a switch.
We can't sink the call to x() down to end because no call to x() happens on the third incoming arc (assume that x() has sideeffects for the sake of argument; if something is safe to speculate we could indeed sink nevertheless but this cannot happen in the general case and causes many extra selects).
We are now able to detect this case and split off the unconditional arcs to a common successor:
[if]
/ \
[x(1)] [if]
| | \
| | \
| [x(2)] |
\ / |
[sink.split] |
\ /
[ end ]
Now we can sink the call to x() into %sink.split. This can cause significant code simplification in many testcases.
llvm-svn: 280217
r279460 rewrote this function to be able to handle more than two incoming edges and took pains to ensure this didn't regress anything.
This time we change the logic for determining if an instruction should be sunk. Previously we used a single pass greedy algorithm - sink instructions until one requires more than one PHI node or we run out of instructions to sink.
This had the problem that sinking instructions that had non-identical but trivially the same operands needed extra logic so we sunk them aggressively. For example:
%a = load i32* %b %d = load i32* %b
%c = gep i32* %a, i32 0 %e = gep i32* %d, i32 1
Sinking %c and %e would naively require two PHI merges as %a != %d. But the loads are obviously equivalent (and maybe can't be hoisted because there is no common predecessor).
This is why we implemented the fairly complex function areValuesTriviallySame(), to look through trivial differences like this. However it's just not clever enough.
Instead, throw areValuesTriviallySame away, use pointer equality to check equivalence of operands and switch to a two-stage algorithm.
In the "scan" stage, we look at every sinkable instruction in isolation from end of block to front. If it's sinkable, we keep track of all operands that required PHI merging.
In the "sink" stage, we iteratively sink the last non-terminator in the source blocks. But when calculating how many PHIs are actually required to be inserted (to work out if we should stop or not) we remove any values that have already been sunk from the set of PHI-merges required, which allows us to be more aggressive.
This turns an algorithm with potentially recursive lookahead (looking through GEPs, casts, loads and any other instruction potentially not CSE'd) to two linear scans.
llvm-svn: 280216
This was deliberately disabled during my rewrite of SinkIfThenToEnd to keep behaviour
at least vaguely consistent with the previous version and keep it as close to NFC as
I could.
There's no real reason not to merge sideeffect calls though, so let's do it! Small fixup
along the way to ensure we don't create indirect calls.
Should fix PR28964.
llvm-svn: 280215
Summary:
1) CoroEarly now lowers llvm.coro.promise intrinsic that allows to obtain
a coroutine promise pointer from a coroutine frame and vice versa.
2) CoroFrame now interprets Promise argument of llvm.coro.begin to
place CoroutinPromise alloca at a deterministic offset from the coroutine frame.
Now, the coroutine promise example from docs\Coroutines.rst compiles and produces expected result (see test/Transform/Coroutines/ex4.ll).
Reviewers: majnemer
Subscribers: llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D23993
llvm-svn: 280184
It's much less code and easier to read if we don't duplicate
everything between the 'Inside' and not 'Inside' cases.
As noted with the FIXME, the goal is to make this vector-friendly
in a follow-up patch.
llvm-svn: 280183
Summary:
LSV was using two vector sets (heads and tails) to track pairs of adjiacent position to vectorize.
A recent optimization is trying to obtain the longest chain to vectorize and assumes the positions
in heads(H) and tails(T) match, which is not the case is there are multiple tails for the same head.
e.g.:
i1: store a[0]
i2: store a[1]
i3: store a[1]
Leads to:
H: i1
T: i2 i3
Instead of:
H: i1 i1
T: i2 i3
So the positions for instructions that follow i3 will have different indexes in H/T.
This patch resolves PR29148.
This issue also surfaced the fact that if the chain is too long, and TLI
returns a "not-fast" answer, the whole chain will be abandoned for
vectorization, even though a smaller one would be beneficial.
Added a testcase and FIXME for this.
Reviewers: tstellarAMD, arsenm, jlebar
Subscribers: mzolotukhin, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D24057
llvm-svn: 280179
We don't need to limit predication to blocks that have a single incoming
edge, we just need to use the right mask.
This fixes PR30172.
Differential Revision: https://reviews.llvm.org/D24009
llvm-svn: 280148
Reverse iterators to doubly-linked lists can be simpler (and cheaper)
than std::reverse_iterator. Make it so.
In particular, change ilist<T>::reverse_iterator so that it is *never*
invalidated unless the node it references is deleted. This matches the
guarantees of ilist<T>::iterator.
(Note: MachineBasicBlock::iterator is *not* an ilist iterator, but a
MachineInstrBundleIterator<MachineInstr>. This commit does not change
MachineBasicBlock::reverse_iterator, but it does update
MachineBasicBlock::reverse_instr_iterator. See note at end of commit
message for details on bundle iterators.)
Given the list (with the Sentinel showing twice for simplicity):
[Sentinel] <-> A <-> B <-> [Sentinel]
the following is now true:
1. begin() represents A.
2. begin() holds the pointer for A.
3. end() represents [Sentinel].
4. end() holds the poitner for [Sentinel].
5. rbegin() represents B.
6. rbegin() holds the pointer for B.
7. rend() represents [Sentinel].
8. rend() holds the pointer for [Sentinel].
The changes are #6 and #8. Here are some properties from the old
scheme (which used std::reverse_iterator):
- rbegin() held the pointer for [Sentinel] and rend() held the pointer
for A;
- operator*() cost two dereferences instead of one;
- converting from a valid iterator to its valid reverse_iterator
involved a confusing increment; and
- "RI++->erase()" left RI invalid. The unintuitive replacement was
"RI->erase(), RE = end()".
With vector-like data structures these properties are hard to avoid
(since past-the-beginning is not a valid pointer), and don't impose a
real cost (since there's still only one dereference, and all iterators
are invalidated on erase). But with lists, this was a poor design.
Specifically, the following code (which obviously works with normal
iterators) now works with ilist::reverse_iterator as well:
for (auto RI = L.rbegin(), RE = L.rend(); RI != RE;)
fooThatMightRemoveArgFromList(*RI++);
Converting between iterator and reverse_iterator for the same node uses
the getReverse() function.
reverse_iterator iterator::getReverse();
iterator reverse_iterator::getReverse();
Why doesn't iterator <=> reverse_iterator conversion use constructors?
In order to catch and update old code, reverse_iterator does not even
have an explicit conversion from iterator. It wouldn't be safe because
there would be no reasonable way to catch all the bugs from the changed
semantic (see the changes at call sites that are part of this patch).
Old code used this API:
std::reverse_iterator::reverse_iterator(iterator);
iterator std::reverse_iterator::base();
Here's how to update from old code to new (that incorporates the
semantic change), assuming I is an ilist<>::iterator and RI is an
ilist<>::reverse_iterator:
[Old] ==> [New]
reverse_iterator(I) (--I).getReverse()
reverse_iterator(I) ++I.getReverse()
--reverse_iterator(I) I.getReverse()
reverse_iterator(++I) I.getReverse()
RI.base() (--RI).getReverse()
RI.base() ++RI.getReverse()
--RI.base() RI.getReverse()
(++RI).base() RI.getReverse()
delete &*RI, RE = end() delete &*RI++
RI->erase(), RE = end() RI++->erase()
=======================================
Note: bundle iterators are out of scope
=======================================
MachineBasicBlock::iterator, also known as
MachineInstrBundleIterator<MachineInstr>, is a wrapper to represent
MachineInstr bundles. The idea is that each operator++ takes you to the
beginning of the next bundle. Implementing a sane reverse iterator for
this is harder than ilist. Here are the options:
- Use std::reverse_iterator<MBB::i>. Store a handle to the beginning of
the next bundle. A call to operator*() runs a loop (usually
operator--() will be called 1 time, for unbundled instructions).
Increment/decrement just works. This is the status quo.
- Store a handle to the final node in the bundle. A call to operator*()
still runs a loop, but it iterates one time fewer (usually
operator--() will be called 0 times, for unbundled instructions).
Increment/decrement just works.
- Make the ilist_sentinel<MachineInstr> *always* store that it's the
sentinel (instead of just in asserts mode). Then the bundle iterator
can sniff the sentinel bit in operator++().
I initially tried implementing the end() option as part of this commit,
but updating iterator/reverse_iterator conversion call sites was
error-prone. I have a WIP series of patches that implements the final
option.
llvm-svn: 280032
