Introduced masks where they are not added and improved target dependent
cost models to avoid returning of the incorrect cost results after
adding masks.
Differential Revision: https://reviews.llvm.org/D100486
Introduced masks where they are not added and improved target dependent
cost models to avoid returning of the incorrect cost results after
adding masks.
Differential Revision: https://reviews.llvm.org/D100486
tryToVectorize() method implements one of searching paths for vectorizable tree roots in SLP vectorizer,
specifically for binary and comparison operations. Order of making probes for various scalar pairs
was defined by its implementation: the instruction operands, then climb over one operand if
the instruction is its sole user and then perform same actions for another operand if previous
attempts failed. Problem with this approach is that among these options we can have more than a
single vectorizable tree candidate and it is not necessarily the one that encountered first.
Trying to build vectorizable tree for each possible combination for just evaluation is expensive.
But we already have lookahead heuristics mechanism which we use for finding best pick among
operands of commutative instructions. It calculates cumulative score for candidates in two
consecutive lanes. This patch introduces use of the heuristics for choosing the best pair among
several combinations. We only try one that looks as most promising for vectorization.
Additional benefit is that we reduce total number of vectorization trees built for probes
because we skip those looking non-profitable early.
Reviewed By: Alexey Bataev (ABataev), Vasileios Porpodas (vporpo)
Differential Revision: https://reviews.llvm.org/D124309
Test case to show not quite optimal SLP vectorization.
Reviewed By: Vasileios Porpodas (vporpo)
Differential Revision: https://reviews.llvm.org/D124293
The original patch (https://reviews.llvm.org/D121354) targets x86 and adjusts
the lookahead score of splat loads ad they can be done by the `movddup`
instruction that combines the load and the broadcast and is cheap to execute.
A similar issue shows up on AArch64. The `ld1r` instruction performs a broadcast
load and is cheap to execute.
This patch implements the TargetTransformInfo hooks for AArch64.
Differential Revision: https://reviews.llvm.org/D123638
SLP uses the distance between pointers to optimize
the getShallowScore. However the current code misses
the case where we are trying to vectorize for VF=4, and the distance
between pointers is 2. In that case the returned score
reflects the case of contiguous loads, when it's not actually
contiguous.
The attached unit tests have 5 loads, where the program order
is not the same as the offset order in the GEPs. So, the choice
of which 4 loads to bundle together matters. If we pick the
first 4, then we can vectorize with VF=4. If we pick the
last 4, then we can only vectorize with VF=2.
This patch makes a more conservative choice, to consider
all distances>1 to not be a case of contiguous load, and
give those cases a lower score.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D123516
Until now we would only accept a broadcast load pattern if it is only used
by a single vector of instructions.
This patch relaxes this, and allows for the broadcast to have more than one
user vector, as long as all of its uses are internal to the SLP graph and
vectorized.
Differential Revision: https://reviews.llvm.org/D121940
Currently SLP vectorizer walks through the instructions and selects
3 main classes of values: 1) reduction operations - instructions with same
reduction opcode (add, mul, min/max, etc.), which build the reduction,
2) reduced values - instructions with the same opcodes, but different
from the reduction opcode, 3) extra arguments - all other values,
instructions from the different basic block rather than the root node,
instructions with to many/less uses.
This scheme is not very efficient. It excludes some instructions and all
non-instruction values from the reductions (constants, proficient
gathers), to many possibly reduced values are marked as extra arguments.
Patch improves this process by introducing a bit extended analysis
stage. During this stage, we still try to select 3 classes of the
values: 1) reduction operations - same as before, 2) possibly reduced
values - all instructions from the current block/non-instructions, which
may build a vectorization tree, 3) extra arguments - instructions from
the different basic blocks. Additionally, an extra sorting of the
possibly reduced values occurs to build the scalar sequences which
highly likely will bed vectorized, e.g. loads are grouped by the
distance between them, constants are grouped together, cmp instructions
are sorted by their compare types and predicates, extractelement
instructions are sorted by the vector operand, etc. Also, these groups
are reordered by their length so the longest group is the first in the
list of the possibly reduced values.
The vectorization process tries to emit the reductions for all these
groups. These reductions, remaining non-vectorized possible reduced
values and extra arguments are then combined into the final expression
just like it was before.
Differential Revision: https://reviews.llvm.org/D114171
We're failing to vectorize several comparison reduction patterns.
Issue #43090 was based off this, but while that simplified test case is now folding, the original still fails due to poor cost model values for vXi1 extractions
If we vectorize a e.g. store, we leave around a bunch of getelementptrs for the individual scalar stores which we removed. We can go ahead and delete them as well.
This is purely for test output quality and readability. It should have no effect in any sane pipeline.
Differential Revision: https://reviews.llvm.org/D122493
The original commit exposed several missing dependencies (e.g. latent bugs in SLP scheduling). Most of these were fixed over the weekend and have had several days to bake. The last was fixed this morning after being noticed in manual review of test changes yesterday. See the review thread for links to each change.
Original commit message follows:
SLP currently schedules all instructions within a scheduling window which stretches from the first instruction potentially vectorized to the last. This window can include a very large number of unrelated instructions which are not being considered for vectorization. This change switches the code to only schedule the sub-graph consisting of the instructions being vectorized and their transitive users.
This has the effect of greatly reducing the amount of work performed in large basic blocks, and thus greatly improves compile time on degenerate examples. To understand the effects, I added some statistics (not planned for upstream contribution). Here's an illustration from my motivating example:
Before this patch:
704357 SLP - Number of calcDeps actions
699021 SLP - Number of schedule calls
5598 SLP - Number of ReSchedule actions
59 SLP - Number of ReScheduleOnFail actions
10084 SLP - Number of schedule resets
8523 SLP - Number of vector instructions generated
After this patch:
102895 SLP - Number of calcDeps actions
161916 SLP - Number of schedule calls
5637 SLP - Number of ReSchedule actions
55 SLP - Number of ReScheduleOnFail actions
10083 SLP - Number of schedule resets
8403 SLP - Number of vector instructions generated
I do want to highlight that there is a small difference in number of generated vector instructions. This example is hitting the bailout due to maximum window size, and the change in scheduling is slightly perturbing when and how we hit it. This can be seen in the RescheduleOnFail counter change. Given that, I think we can safely ignore.
The downside of this change can be seen in the large test diff. We group all vectorizable instructions together at the bottom of the scheduling region. This means that vector instructions can move quite far from their original point in code. While maybe undesirable, I don't see this as being a major problem as this pass is not intended to be a general scheduling pass.
For context, it's worth noting that the pre-scheduling that SLP does while building the vector tree is exactly the sub-graph scheduling implemented by this patch.
Differential Revision: https://reviews.llvm.org/D118538
The semantics of an inalloca alloca instruction requires that it not be reordered with a preceeding stacksave intrinsic call. Unfortunately, there's no def/use edge or memory dependence edge. (THe memory point is slightly subtle, but in general a new allocation can't alias with a call which executes strictly before it comes into existance.)
I'd tried to tackle this same case previously in 689babdf6, but the fix chosen there turned out to be incomplete. As such, this change contains a fully revert of the first fix attempt.
This was noticed when investigating problems which surfaced with D118538, but this is definitely an existing bug. This time around, I managed to reduce a couple of additional cases, including one which was being actively miscompiled even without the new scheduling change. (See test diffs)
Compile time wise, we only spend extra time when seeing a stacksave (rare), and even then we walk the block at most once per schedule window extension. Likely a non-issue.
There are some slight changes to the test lines due to different cost threshold choices in the two command lines, but I don't believe these to be interesting the purpose of the tests.
The existing scheduling doesn't account for the scheduling restrictions implied by inalloca allocas combined with stacksave/stackrestore. This adds coverage including one currently miscompiling case.
This fixes an active miscompile visible in the test changes. The basic problem is that the scheduling dependency graph didn't have any edges for control dependence within a single basic block. The result is that we could (and in some rare cases *did*) perform reorderings within a block which could introduce new undefined behavior along paths which didn't previously contain any.
Impact wise, we have two major cases where control is not guaranteed to reach a later instruction in the block: may throw calls, and calls containing infinite loops.
* The former case was mostly covered by the memory dependencies, and to trigger require a function which can throw, but not write to memory. In theory, such a case is possible, but not likely in practice.
* The later case is likely more of an issue in practice. After this code was first written, we changed the IR semantics to allow well defined infinite loops without satisifying mustprogress. Even for C/C++ - which do imply mustprogress - recent changes to how we treat atomics (e.g. an atomic read does not always imply a write) could expose this issue. I'm a bit shocked we don't seem to have a bug report which hit this in real code actually.
Compile time wise, this results in a single extra scan of the scheduling window in the common case. Since we stop scanning at the next instruction which isn't guaranteed to execute, no matter what order we traverse instructions in, we scan the block once. The exception to this is that when we extend the scheduling window downwards, we invalidate all dependencies, and thus rescan. So the potentially expensive case is when we a call in a big schedule window which is frequently extended. We could optimize this case (by caching the last instruction not guaranteeed to transfer execution and scanning only the extended window) and starting there), but I decided to leave the complexity until it mattered. That same case is already degenerate with memory dependences which is more expensive than the control dependence scan.
We could also consider combining the memory dependence and control dependence sets to reduce memory usage, but since it complicates the code slightly and makes debugging a bit harder, I went with the simplest scheme for now.
This was noticed while trying to understand the failures reported against D118538, but is not otherwise related to that change.
SLP is currently assuming that control dependence in these cases is irrelevant. This is only valid if none of the lib-funcs involved can throw or infinite loop in the scalar forms. This appears to be true (or at least we infer the respective attributes) for the libfuncs I spot checked. This change is mostly for shrunking the diff on an upcoming patch.
Splat loads are inexpensive in X86. For a 2-lane vector we need just one
instruction: `movddup (%reg), xmm0`. Using the standard Splat score leads
to worse code. This patch adds a new score dedicated for splat loads.
Please note that a splat is usually three IR instructions:
- It is usually a load and 2 inserts:
%ld = load double, double* %gep
%ins1 = insertelement <2 x double> poison, double %ld, i32 0
%ins2 = insertelement <2 x double> %ins1, double %ld, i32 1
- But it can also be a load, an insert and a shuffle:
%ld = load double, double* %gep
%ins = insertelement <2 x double> poison, double %ld, i32 0
%shf = shufflevector <2 x double> %ins, <2 x double> poison, <2 x i32> zeroinitializer
Because of this some of the lit tests contain more IR instructions.
Differential Revision: https://reviews.llvm.org/D121354
No need to schedule entry nodes where all instructions are not memory
read/write instructions and their operands are either constants, or
arguments, or phis, or instructions from others blocks, or their users
are phis or from the other blocks.
The resulting vector instructions can be placed at
the beginning of the basic block without scheduling (if operands does
not need to be scheduled) or at the end of the block (if users are
outside of the block).
It may save some compile time and scheduling resources.
Differential Revision: https://reviews.llvm.org/D121121
No need to schedule entry nodes where all instructions are not memory
read/write instructions and their operands are either constants, or
arguments, or phis, or instructions from others blocks, or their users
are phis or from the other blocks.
The resulting vector instructions can be placed at
the beginning of the basic block without scheduling (if operands does
not need to be scheduled) or at the end of the block (if users are
outside of the block).
It may save some compile time and scheduling resources.
Differential Revision: https://reviews.llvm.org/D121121
The costs of vector shifts was 2 as opposed to 1, as the nodes are
marked custom. Fix this like the others and mark the nodes as cheap.
Differential Revision: https://reviews.llvm.org/D120773
Root issue which triggered the revert was fixed in 689bab. No changes in the reapplied patch.
Original commit message follows:
SLP currently schedules all instructions within a scheduling window which stretches from the first instr
uction potentially vectorized to the last. This window can include a very large number of unrelated instruct
ions which are not being considered for vectorization. This change switches the code to only schedule the su
b-graph consisting of the instructions being vectorized and their transitive users.
This has the effect of greatly reducing the amount of work performed in large basic blocks, and thus greatly improves compile time on degenerate examples. To understand the effects, I added some statistics (not planned for upstream contribution). Here's an illustration from my motivating example:
Before this patch:
704357 SLP - Number of calcDeps actions
699021 SLP - Number of schedule calls
5598 SLP - Number of ReSchedule actions
59 SLP - Number of ReScheduleOnFail actions
10084 SLP - Number of schedule resets
8523 SLP - Number of vector instructions generated
After this patch:
102895 SLP - Number of calcDeps actions
161916 SLP - Number of schedule calls
5637 SLP - Number of ReSchedule actions
55 SLP - Number of ReScheduleOnFail actions
10083 SLP - Number of schedule resets
8403 SLP - Number of vector instructions generated
I do want to highlight that there is a small difference in number of generated vector instructions. This example is hitting the bailout due to maximum window size, and the change in scheduling is slightly perturbing when and how we hit it. This can be seen in the RescheduleOnFail counter change. Given that, I think we can safely ignore.
The downside of this change can be seen in the large test diff. We group all vectorizable instructions together at the bottom of the scheduling region. This means that vector instructions can move quite far from their original point in code. While maybe undesirable, I don't see this as being a major problem as this pass is not intended to be a general scheduling pass.
For context, it's worth noting that the pre-scheduling that SLP does while building the vector tree is exactly the sub-graph scheduling implemented by this patch.
Differential Revision: https://reviews.llvm.org/D118538
While a collection of allocas are technically vectorizeable - by forming a wider alloca - this was not a transform SLP actually knows how to do. Instead, we were forming a bundle with missing dependencies, and then relying on the scheduling code to preserve program order if multiple instructions were scheduleable at once. I haven't been able to write a test case, but I'm 99% sure this was wrong in some edge case.
The unknown op case was flowing down the shufflevector path. This did result in some splat handling being lost with this change, but the same lack of splat handling is visible in a whole bunch of simple examples for the gather path. I didn't consider this interesting to fix given how narrow the splat of allocas case is.
