2017-05-17 07:10:25 +08:00
|
|
|
//===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-06-06 19:49:48 +08:00
|
|
|
#include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2017-06-06 19:49:48 +08:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/ADT/Sequence.h"
|
|
|
|
#include "llvm/ADT/SetVector.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/ADT/Twine.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/Analysis/AssumptionCache.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/Analysis/LoopAnalysisManager.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
|
|
|
#include "llvm/Analysis/LoopPass.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/IR/BasicBlock.h"
|
|
|
|
#include "llvm/IR/Constant.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/IR/Constants.h"
|
|
|
|
#include "llvm/IR/Dominators.h"
|
|
|
|
#include "llvm/IR/Function.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/IR/InstrTypes.h"
|
|
|
|
#include "llvm/IR/Instruction.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/IR/Instructions.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/IR/Use.h"
|
|
|
|
#include "llvm/IR/Value.h"
|
|
|
|
#include "llvm/Pass.h"
|
|
|
|
#include "llvm/Support/Casting.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/Support/Debug.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
|
|
|
#include "llvm/Support/GenericDomTree.h"
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
|
|
|
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
|
|
|
|
#include "llvm/Transforms/Utils/LoopUtils.h"
|
2017-05-17 07:10:25 +08:00
|
|
|
#include <algorithm>
|
|
|
|
#include <cassert>
|
|
|
|
#include <iterator>
|
|
|
|
#include <utility>
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
|
|
|
|
#define DEBUG_TYPE "simple-loop-unswitch"
|
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
STATISTIC(NumBranches, "Number of branches unswitched");
|
|
|
|
STATISTIC(NumSwitches, "Number of switches unswitched");
|
|
|
|
STATISTIC(NumTrivial, "Number of unswitches that are trivial");
|
|
|
|
|
|
|
|
static void replaceLoopUsesWithConstant(Loop &L, Value &LIC,
|
|
|
|
Constant &Replacement) {
|
|
|
|
assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
|
|
|
|
|
|
|
|
// Replace uses of LIC in the loop with the given constant.
|
|
|
|
for (auto UI = LIC.use_begin(), UE = LIC.use_end(); UI != UE;) {
|
|
|
|
// Grab the use and walk past it so we can clobber it in the use list.
|
|
|
|
Use *U = &*UI++;
|
|
|
|
Instruction *UserI = dyn_cast<Instruction>(U->getUser());
|
|
|
|
if (!UserI || !L.contains(UserI))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Replace this use within the loop body.
|
|
|
|
*U = &Replacement;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Update the dominator tree after removing one exiting predecessor of a loop
|
|
|
|
/// exit block.
|
|
|
|
static void updateLoopExitIDom(BasicBlock *LoopExitBB, Loop &L,
|
2017-05-12 10:19:59 +08:00
|
|
|
DominatorTree &DT) {
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
assert(pred_begin(LoopExitBB) != pred_end(LoopExitBB) &&
|
|
|
|
"Cannot have empty predecessors of the loop exit block if we split "
|
|
|
|
"off a block to unswitch!");
|
|
|
|
|
|
|
|
BasicBlock *IDom = *pred_begin(LoopExitBB);
|
|
|
|
// Walk all of the other predecessors finding the nearest common dominator
|
|
|
|
// until all predecessors are covered or we reach the loop header. The loop
|
|
|
|
// header necessarily dominates all loop exit blocks in loop simplified form
|
|
|
|
// so we can early-exit the moment we hit that block.
|
|
|
|
for (auto PI = std::next(pred_begin(LoopExitBB)), PE = pred_end(LoopExitBB);
|
|
|
|
PI != PE && IDom != L.getHeader(); ++PI)
|
|
|
|
IDom = DT.findNearestCommonDominator(IDom, *PI);
|
|
|
|
|
|
|
|
DT.changeImmediateDominator(LoopExitBB, IDom);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Update the dominator tree after unswitching a particular former exit block.
|
|
|
|
///
|
|
|
|
/// This handles the full update of the dominator tree after hoisting a block
|
|
|
|
/// that previously was an exit block (or split off of an exit block) up to be
|
|
|
|
/// reached from the new immediate dominator of the preheader.
|
|
|
|
///
|
|
|
|
/// The common case is simple -- we just move the unswitched block to have an
|
|
|
|
/// immediate dominator of the old preheader. But in complex cases, there may
|
|
|
|
/// be other blocks reachable from the unswitched block that are immediately
|
|
|
|
/// dominated by some node between the unswitched one and the old preheader.
|
|
|
|
/// All of these also need to be hoisted in the dominator tree. We also want to
|
|
|
|
/// minimize queries to the dominator tree because each step of this
|
|
|
|
/// invalidates any DFS numbers that would make queries fast.
|
|
|
|
static void updateDTAfterUnswitch(BasicBlock *UnswitchedBB, BasicBlock *OldPH,
|
|
|
|
DominatorTree &DT) {
|
|
|
|
DomTreeNode *OldPHNode = DT[OldPH];
|
|
|
|
DomTreeNode *UnswitchedNode = DT[UnswitchedBB];
|
|
|
|
// If the dominator tree has already been updated for this unswitched node,
|
|
|
|
// we're done. This makes it easier to use this routine if there are multiple
|
|
|
|
// paths to the same unswitched destination.
|
|
|
|
if (UnswitchedNode->getIDom() == OldPHNode)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// First collect the domtree nodes that we are hoisting over. These are the
|
|
|
|
// set of nodes which may have children that need to be hoisted as well.
|
|
|
|
SmallPtrSet<DomTreeNode *, 4> DomChain;
|
|
|
|
for (auto *IDom = UnswitchedNode->getIDom(); IDom != OldPHNode;
|
|
|
|
IDom = IDom->getIDom())
|
|
|
|
DomChain.insert(IDom);
|
|
|
|
|
|
|
|
// The unswitched block ends up immediately dominated by the old preheader --
|
|
|
|
// regardless of whether it is the loop exit block or split off of the loop
|
|
|
|
// exit block.
|
|
|
|
DT.changeImmediateDominator(UnswitchedNode, OldPHNode);
|
|
|
|
|
[PM/Unswitch] Fix a bug in the domtree update logic for the new unswitch
pass.
The original logic only considered direct successors of the hoisted
domtree nodes, but that isn't really enough. If there are other basic
blocks that are completely within the subtree, their successors could
just as easily be impacted by the hoisting.
The more I think about it, the more I think the correct update here is
to hoist every block on the dominance frontier which has an idom in the
chain we hoist across. However, this is subtle enough that I'd
definitely appreciate some more eyes on it.
Sadly, if this is the correct algorithm, it requires computing a (highly
localized) dominance frontier. I've done this in the simplest (IE, least
code) way I could come up with, but that may be too naive. Suggestions
welcome here, dominance update algorithms are not an area I've studied
much, so I don't have strong opinions.
In good news, with this patch, turning on simple unswitch passes the
LLVM test suite for me with asserts enabled.
Differential Revision: https://reviews.llvm.org/D32740
llvm-svn: 303843
2017-05-25 14:33:36 +08:00
|
|
|
// For everything that moves up the dominator tree, we need to examine the
|
|
|
|
// dominator frontier to see if it additionally should move up the dominator
|
|
|
|
// tree. This lambda appends the dominator frontier for a node on the
|
|
|
|
// worklist.
|
|
|
|
//
|
|
|
|
// Note that we don't currently use the IDFCalculator here for two reasons:
|
|
|
|
// 1) It computes dominator tree levels for the entire function on each run
|
|
|
|
// of 'compute'. While this isn't terrible, given that we expect to update
|
|
|
|
// relatively small subtrees of the domtree, it isn't necessarily the right
|
|
|
|
// tradeoff.
|
|
|
|
// 2) The interface doesn't fit this usage well. It doesn't operate in
|
|
|
|
// append-only, and builds several sets that we don't need.
|
|
|
|
//
|
|
|
|
// FIXME: Neither of these issues are a big deal and could be addressed with
|
|
|
|
// some amount of refactoring of IDFCalculator. That would allow us to share
|
|
|
|
// the core logic here (which is solving the same core problem).
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
SmallSetVector<BasicBlock *, 4> Worklist;
|
[PM/Unswitch] Fix a bug in the domtree update logic for the new unswitch
pass.
The original logic only considered direct successors of the hoisted
domtree nodes, but that isn't really enough. If there are other basic
blocks that are completely within the subtree, their successors could
just as easily be impacted by the hoisting.
The more I think about it, the more I think the correct update here is
to hoist every block on the dominance frontier which has an idom in the
chain we hoist across. However, this is subtle enough that I'd
definitely appreciate some more eyes on it.
Sadly, if this is the correct algorithm, it requires computing a (highly
localized) dominance frontier. I've done this in the simplest (IE, least
code) way I could come up with, but that may be too naive. Suggestions
welcome here, dominance update algorithms are not an area I've studied
much, so I don't have strong opinions.
In good news, with this patch, turning on simple unswitch passes the
LLVM test suite for me with asserts enabled.
Differential Revision: https://reviews.llvm.org/D32740
llvm-svn: 303843
2017-05-25 14:33:36 +08:00
|
|
|
SmallVector<DomTreeNode *, 4> DomNodes;
|
|
|
|
SmallPtrSet<BasicBlock *, 4> DomSet;
|
|
|
|
auto AppendDomFrontier = [&](DomTreeNode *Node) {
|
|
|
|
assert(DomNodes.empty() && "Must start with no dominator nodes.");
|
|
|
|
assert(DomSet.empty() && "Must start with an empty dominator set.");
|
|
|
|
|
|
|
|
// First flatten this subtree into sequence of nodes by doing a pre-order
|
|
|
|
// walk.
|
|
|
|
DomNodes.push_back(Node);
|
|
|
|
// We intentionally re-evaluate the size as each node can add new children.
|
|
|
|
// Because this is a tree walk, this cannot add any duplicates.
|
|
|
|
for (int i = 0; i < (int)DomNodes.size(); ++i)
|
|
|
|
DomNodes.insert(DomNodes.end(), DomNodes[i]->begin(), DomNodes[i]->end());
|
|
|
|
|
|
|
|
// Now create a set of the basic blocks so we can quickly test for
|
|
|
|
// dominated successors. We could in theory use the DFS numbers of the
|
|
|
|
// dominator tree for this, but we want this to remain predictably fast
|
|
|
|
// even while we mutate the dominator tree in ways that would invalidate
|
|
|
|
// the DFS numbering.
|
|
|
|
for (DomTreeNode *InnerN : DomNodes)
|
|
|
|
DomSet.insert(InnerN->getBlock());
|
|
|
|
|
|
|
|
// Now re-walk the nodes, appending every successor of every node that isn't
|
|
|
|
// in the set. Note that we don't append the node itself, even though if it
|
|
|
|
// is a successor it does not strictly dominate itself and thus it would be
|
|
|
|
// part of the dominance frontier. The reason we don't append it is that
|
|
|
|
// the node passed in came *from* the worklist and so it has already been
|
|
|
|
// processed.
|
|
|
|
for (DomTreeNode *InnerN : DomNodes)
|
|
|
|
for (BasicBlock *SuccBB : successors(InnerN->getBlock()))
|
|
|
|
if (!DomSet.count(SuccBB))
|
|
|
|
Worklist.insert(SuccBB);
|
|
|
|
|
|
|
|
DomNodes.clear();
|
|
|
|
DomSet.clear();
|
|
|
|
};
|
|
|
|
|
|
|
|
// Append the initial dom frontier nodes.
|
|
|
|
AppendDomFrontier(UnswitchedNode);
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
|
|
|
|
// Walk the worklist. We grow the list in the loop and so must recompute size.
|
|
|
|
for (int i = 0; i < (int)Worklist.size(); ++i) {
|
|
|
|
auto *BB = Worklist[i];
|
|
|
|
|
|
|
|
DomTreeNode *Node = DT[BB];
|
|
|
|
assert(!DomChain.count(Node) &&
|
|
|
|
"Cannot be dominated by a block you can reach!");
|
[PM/Unswitch] Fix a bug in the domtree update logic for the new unswitch
pass.
The original logic only considered direct successors of the hoisted
domtree nodes, but that isn't really enough. If there are other basic
blocks that are completely within the subtree, their successors could
just as easily be impacted by the hoisting.
The more I think about it, the more I think the correct update here is
to hoist every block on the dominance frontier which has an idom in the
chain we hoist across. However, this is subtle enough that I'd
definitely appreciate some more eyes on it.
Sadly, if this is the correct algorithm, it requires computing a (highly
localized) dominance frontier. I've done this in the simplest (IE, least
code) way I could come up with, but that may be too naive. Suggestions
welcome here, dominance update algorithms are not an area I've studied
much, so I don't have strong opinions.
In good news, with this patch, turning on simple unswitch passes the
LLVM test suite for me with asserts enabled.
Differential Revision: https://reviews.llvm.org/D32740
llvm-svn: 303843
2017-05-25 14:33:36 +08:00
|
|
|
|
|
|
|
// If this block had an immediate dominator somewhere in the chain
|
|
|
|
// we hoisted over, then its position in the domtree needs to move as it is
|
|
|
|
// reachable from a node hoisted over this chain.
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
if (!DomChain.count(Node->getIDom()))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
DT.changeImmediateDominator(Node, OldPHNode);
|
[PM/Unswitch] Fix a bug in the domtree update logic for the new unswitch
pass.
The original logic only considered direct successors of the hoisted
domtree nodes, but that isn't really enough. If there are other basic
blocks that are completely within the subtree, their successors could
just as easily be impacted by the hoisting.
The more I think about it, the more I think the correct update here is
to hoist every block on the dominance frontier which has an idom in the
chain we hoist across. However, this is subtle enough that I'd
definitely appreciate some more eyes on it.
Sadly, if this is the correct algorithm, it requires computing a (highly
localized) dominance frontier. I've done this in the simplest (IE, least
code) way I could come up with, but that may be too naive. Suggestions
welcome here, dominance update algorithms are not an area I've studied
much, so I don't have strong opinions.
In good news, with this patch, turning on simple unswitch passes the
LLVM test suite for me with asserts enabled.
Differential Revision: https://reviews.llvm.org/D32740
llvm-svn: 303843
2017-05-25 14:33:36 +08:00
|
|
|
|
|
|
|
// Now add this node's dominator frontier to the worklist as well.
|
|
|
|
AppendDomFrontier(Node);
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-05-12 10:19:59 +08:00
|
|
|
/// Check that all the LCSSA PHI nodes in the loop exit block have trivial
|
|
|
|
/// incoming values along this edge.
|
|
|
|
static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
|
|
|
|
BasicBlock &ExitBB) {
|
|
|
|
for (Instruction &I : ExitBB) {
|
|
|
|
auto *PN = dyn_cast<PHINode>(&I);
|
|
|
|
if (!PN)
|
|
|
|
// No more PHIs to check.
|
|
|
|
return true;
|
|
|
|
|
|
|
|
// If the incoming value for this edge isn't loop invariant the unswitch
|
|
|
|
// won't be trivial.
|
|
|
|
if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
llvm_unreachable("Basic blocks should never be empty!");
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Rewrite the PHI nodes in an unswitched loop exit basic block.
|
|
|
|
///
|
|
|
|
/// Requires that the loop exit and unswitched basic block are the same, and
|
|
|
|
/// that the exiting block was a unique predecessor of that block. Rewrites the
|
|
|
|
/// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
|
|
|
|
/// PHI nodes from the old preheader that now contains the unswitched
|
|
|
|
/// terminator.
|
|
|
|
static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
|
|
|
|
BasicBlock &OldExitingBB,
|
|
|
|
BasicBlock &OldPH) {
|
|
|
|
for (Instruction &I : UnswitchedBB) {
|
|
|
|
auto *PN = dyn_cast<PHINode>(&I);
|
|
|
|
if (!PN)
|
|
|
|
// No more PHIs to check.
|
|
|
|
break;
|
|
|
|
|
|
|
|
// When the loop exit is directly unswitched we just need to update the
|
|
|
|
// incoming basic block. We loop to handle weird cases with repeated
|
|
|
|
// incoming blocks, but expect to typically only have one operand here.
|
2017-05-17 07:10:25 +08:00
|
|
|
for (auto i : seq<int>(0, PN->getNumOperands())) {
|
2017-05-12 10:19:59 +08:00
|
|
|
assert(PN->getIncomingBlock(i) == &OldExitingBB &&
|
|
|
|
"Found incoming block different from unique predecessor!");
|
|
|
|
PN->setIncomingBlock(i, &OldPH);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Rewrite the PHI nodes in the loop exit basic block and the split off
|
|
|
|
/// unswitched block.
|
|
|
|
///
|
|
|
|
/// Because the exit block remains an exit from the loop, this rewrites the
|
|
|
|
/// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
|
|
|
|
/// nodes into the unswitched basic block to select between the value in the
|
|
|
|
/// old preheader and the loop exit.
|
|
|
|
static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
|
|
|
|
BasicBlock &UnswitchedBB,
|
|
|
|
BasicBlock &OldExitingBB,
|
|
|
|
BasicBlock &OldPH) {
|
|
|
|
assert(&ExitBB != &UnswitchedBB &&
|
|
|
|
"Must have different loop exit and unswitched blocks!");
|
|
|
|
Instruction *InsertPt = &*UnswitchedBB.begin();
|
|
|
|
for (Instruction &I : ExitBB) {
|
|
|
|
auto *PN = dyn_cast<PHINode>(&I);
|
|
|
|
if (!PN)
|
|
|
|
// No more PHIs to check.
|
|
|
|
break;
|
|
|
|
|
|
|
|
auto *NewPN = PHINode::Create(PN->getType(), /*NumReservedValues*/ 2,
|
|
|
|
PN->getName() + ".split", InsertPt);
|
|
|
|
|
|
|
|
// Walk backwards over the old PHI node's inputs to minimize the cost of
|
|
|
|
// removing each one. We have to do this weird loop manually so that we
|
|
|
|
// create the same number of new incoming edges in the new PHI as we expect
|
|
|
|
// each case-based edge to be included in the unswitched switch in some
|
|
|
|
// cases.
|
|
|
|
// FIXME: This is really, really gross. It would be much cleaner if LLVM
|
|
|
|
// allowed us to create a single entry for a predecessor block without
|
|
|
|
// having separate entries for each "edge" even though these edges are
|
|
|
|
// required to produce identical results.
|
|
|
|
for (int i = PN->getNumIncomingValues() - 1; i >= 0; --i) {
|
|
|
|
if (PN->getIncomingBlock(i) != &OldExitingBB)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
Value *Incoming = PN->removeIncomingValue(i);
|
|
|
|
NewPN->addIncoming(Incoming, &OldPH);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now replace the old PHI with the new one and wire the old one in as an
|
|
|
|
// input to the new one.
|
|
|
|
PN->replaceAllUsesWith(NewPN);
|
|
|
|
NewPN->addIncoming(PN, &ExitBB);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
/// Unswitch a trivial branch if the condition is loop invariant.
|
|
|
|
///
|
|
|
|
/// This routine should only be called when loop code leading to the branch has
|
|
|
|
/// been validated as trivial (no side effects). This routine checks if the
|
|
|
|
/// condition is invariant and one of the successors is a loop exit. This
|
|
|
|
/// allows us to unswitch without duplicating the loop, making it trivial.
|
|
|
|
///
|
|
|
|
/// If this routine fails to unswitch the branch it returns false.
|
|
|
|
///
|
|
|
|
/// If the branch can be unswitched, this routine splits the preheader and
|
|
|
|
/// hoists the branch above that split. Preserves loop simplified form
|
|
|
|
/// (splitting the exit block as necessary). It simplifies the branch within
|
|
|
|
/// the loop to an unconditional branch but doesn't remove it entirely. Further
|
|
|
|
/// cleanup can be done with some simplify-cfg like pass.
|
|
|
|
static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
|
|
|
|
LoopInfo &LI) {
|
|
|
|
assert(BI.isConditional() && "Can only unswitch a conditional branch!");
|
|
|
|
DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
|
|
|
|
|
|
|
|
Value *LoopCond = BI.getCondition();
|
|
|
|
|
|
|
|
// Need a trivial loop condition to unswitch.
|
|
|
|
if (!L.isLoopInvariant(LoopCond))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// FIXME: We should compute this once at the start and update it!
|
|
|
|
SmallVector<BasicBlock *, 16> ExitBlocks;
|
|
|
|
L.getExitBlocks(ExitBlocks);
|
|
|
|
SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(),
|
|
|
|
ExitBlocks.end());
|
|
|
|
|
|
|
|
// Check to see if a successor of the branch is guaranteed to
|
|
|
|
// exit through a unique exit block without having any
|
|
|
|
// side-effects. If so, determine the value of Cond that causes
|
|
|
|
// it to do this.
|
|
|
|
ConstantInt *CondVal = ConstantInt::getTrue(BI.getContext());
|
|
|
|
ConstantInt *Replacement = ConstantInt::getFalse(BI.getContext());
|
|
|
|
int LoopExitSuccIdx = 0;
|
|
|
|
auto *LoopExitBB = BI.getSuccessor(0);
|
|
|
|
if (!ExitBlockSet.count(LoopExitBB)) {
|
|
|
|
std::swap(CondVal, Replacement);
|
|
|
|
LoopExitSuccIdx = 1;
|
|
|
|
LoopExitBB = BI.getSuccessor(1);
|
|
|
|
if (!ExitBlockSet.count(LoopExitBB))
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
|
|
|
|
assert(L.contains(ContinueBB) &&
|
|
|
|
"Cannot have both successors exit and still be in the loop!");
|
|
|
|
|
2017-05-12 10:19:59 +08:00
|
|
|
auto *ParentBB = BI.getParent();
|
|
|
|
if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB))
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
return false;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << " unswitching trivial branch when: " << CondVal
|
|
|
|
<< " == " << LoopCond << "\n");
|
|
|
|
|
|
|
|
// Split the preheader, so that we know that there is a safe place to insert
|
|
|
|
// the conditional branch. We will change the preheader to have a conditional
|
|
|
|
// branch on LoopCond.
|
|
|
|
BasicBlock *OldPH = L.getLoopPreheader();
|
|
|
|
BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI);
|
|
|
|
|
|
|
|
// Now that we have a place to insert the conditional branch, create a place
|
|
|
|
// to branch to: this is the exit block out of the loop that we are
|
|
|
|
// unswitching. We need to split this if there are other loop predecessors.
|
|
|
|
// Because the loop is in simplified form, *any* other predecessor is enough.
|
|
|
|
BasicBlock *UnswitchedBB;
|
|
|
|
if (BasicBlock *PredBB = LoopExitBB->getUniquePredecessor()) {
|
|
|
|
(void)PredBB;
|
2017-05-12 10:19:59 +08:00
|
|
|
assert(PredBB == BI.getParent() &&
|
|
|
|
"A branch's parent isn't a predecessor!");
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
UnswitchedBB = LoopExitBB;
|
|
|
|
} else {
|
|
|
|
UnswitchedBB = SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now splice the branch to gate reaching the new preheader and re-point its
|
|
|
|
// successors.
|
|
|
|
OldPH->getInstList().splice(std::prev(OldPH->end()),
|
|
|
|
BI.getParent()->getInstList(), BI);
|
|
|
|
OldPH->getTerminator()->eraseFromParent();
|
|
|
|
BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
|
|
|
|
BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
|
|
|
|
|
|
|
|
// Create a new unconditional branch that will continue the loop as a new
|
|
|
|
// terminator.
|
|
|
|
BranchInst::Create(ContinueBB, ParentBB);
|
|
|
|
|
2017-05-12 10:19:59 +08:00
|
|
|
// Rewrite the relevant PHI nodes.
|
|
|
|
if (UnswitchedBB == LoopExitBB)
|
|
|
|
rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
|
|
|
|
else
|
|
|
|
rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
|
|
|
|
*ParentBB, *OldPH);
|
|
|
|
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
// Now we need to update the dominator tree.
|
|
|
|
updateDTAfterUnswitch(UnswitchedBB, OldPH, DT);
|
|
|
|
// But if we split something off of the loop exit block then we also removed
|
|
|
|
// one of the predecessors for the loop exit block and may need to update its
|
|
|
|
// idom.
|
|
|
|
if (UnswitchedBB != LoopExitBB)
|
|
|
|
updateLoopExitIDom(LoopExitBB, L, DT);
|
|
|
|
|
|
|
|
// Since this is an i1 condition we can also trivially replace uses of it
|
|
|
|
// within the loop with a constant.
|
|
|
|
replaceLoopUsesWithConstant(L, *LoopCond, *Replacement);
|
|
|
|
|
|
|
|
++NumTrivial;
|
|
|
|
++NumBranches;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Unswitch a trivial switch if the condition is loop invariant.
|
|
|
|
///
|
|
|
|
/// This routine should only be called when loop code leading to the switch has
|
|
|
|
/// been validated as trivial (no side effects). This routine checks if the
|
|
|
|
/// condition is invariant and that at least one of the successors is a loop
|
|
|
|
/// exit. This allows us to unswitch without duplicating the loop, making it
|
|
|
|
/// trivial.
|
|
|
|
///
|
|
|
|
/// If this routine fails to unswitch the switch it returns false.
|
|
|
|
///
|
|
|
|
/// If the switch can be unswitched, this routine splits the preheader and
|
|
|
|
/// copies the switch above that split. If the default case is one of the
|
|
|
|
/// exiting cases, it copies the non-exiting cases and points them at the new
|
|
|
|
/// preheader. If the default case is not exiting, it copies the exiting cases
|
|
|
|
/// and points the default at the preheader. It preserves loop simplified form
|
|
|
|
/// (splitting the exit blocks as necessary). It simplifies the switch within
|
|
|
|
/// the loop by removing now-dead cases. If the default case is one of those
|
|
|
|
/// unswitched, it replaces its destination with a new basic block containing
|
|
|
|
/// only unreachable. Such basic blocks, while technically loop exits, are not
|
|
|
|
/// considered for unswitching so this is a stable transform and the same
|
|
|
|
/// switch will not be revisited. If after unswitching there is only a single
|
|
|
|
/// in-loop successor, the switch is further simplified to an unconditional
|
|
|
|
/// branch. Still more cleanup can be done with some simplify-cfg like pass.
|
|
|
|
static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
|
|
|
|
LoopInfo &LI) {
|
|
|
|
DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
|
|
|
|
Value *LoopCond = SI.getCondition();
|
|
|
|
|
|
|
|
// If this isn't switching on an invariant condition, we can't unswitch it.
|
|
|
|
if (!L.isLoopInvariant(LoopCond))
|
|
|
|
return false;
|
|
|
|
|
2017-05-12 10:19:59 +08:00
|
|
|
auto *ParentBB = SI.getParent();
|
|
|
|
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
// FIXME: We should compute this once at the start and update it!
|
|
|
|
SmallVector<BasicBlock *, 16> ExitBlocks;
|
|
|
|
L.getExitBlocks(ExitBlocks);
|
|
|
|
SmallPtrSet<BasicBlock *, 16> ExitBlockSet(ExitBlocks.begin(),
|
|
|
|
ExitBlocks.end());
|
|
|
|
|
|
|
|
SmallVector<int, 4> ExitCaseIndices;
|
|
|
|
for (auto Case : SI.cases()) {
|
|
|
|
auto *SuccBB = Case.getCaseSuccessor();
|
2017-05-12 10:19:59 +08:00
|
|
|
if (ExitBlockSet.count(SuccBB) &&
|
|
|
|
areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB))
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
ExitCaseIndices.push_back(Case.getCaseIndex());
|
|
|
|
}
|
|
|
|
BasicBlock *DefaultExitBB = nullptr;
|
|
|
|
if (ExitBlockSet.count(SI.getDefaultDest()) &&
|
2017-05-12 10:19:59 +08:00
|
|
|
areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) &&
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
!isa<UnreachableInst>(SI.getDefaultDest()->getTerminator()))
|
|
|
|
DefaultExitBB = SI.getDefaultDest();
|
|
|
|
else if (ExitCaseIndices.empty())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << " unswitching trivial cases...\n");
|
|
|
|
|
|
|
|
SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases;
|
|
|
|
ExitCases.reserve(ExitCaseIndices.size());
|
|
|
|
// We walk the case indices backwards so that we remove the last case first
|
|
|
|
// and don't disrupt the earlier indices.
|
|
|
|
for (unsigned Index : reverse(ExitCaseIndices)) {
|
|
|
|
auto CaseI = SI.case_begin() + Index;
|
|
|
|
// Save the value of this case.
|
|
|
|
ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()});
|
|
|
|
// Delete the unswitched cases.
|
|
|
|
SI.removeCase(CaseI);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check if after this all of the remaining cases point at the same
|
|
|
|
// successor.
|
|
|
|
BasicBlock *CommonSuccBB = nullptr;
|
|
|
|
if (SI.getNumCases() > 0 &&
|
|
|
|
std::all_of(std::next(SI.case_begin()), SI.case_end(),
|
|
|
|
[&SI](const SwitchInst::CaseHandle &Case) {
|
|
|
|
return Case.getCaseSuccessor() ==
|
|
|
|
SI.case_begin()->getCaseSuccessor();
|
|
|
|
}))
|
|
|
|
CommonSuccBB = SI.case_begin()->getCaseSuccessor();
|
|
|
|
|
|
|
|
if (DefaultExitBB) {
|
|
|
|
// We can't remove the default edge so replace it with an edge to either
|
|
|
|
// the single common remaining successor (if we have one) or an unreachable
|
|
|
|
// block.
|
|
|
|
if (CommonSuccBB) {
|
|
|
|
SI.setDefaultDest(CommonSuccBB);
|
|
|
|
} else {
|
|
|
|
BasicBlock *UnreachableBB = BasicBlock::Create(
|
|
|
|
ParentBB->getContext(),
|
|
|
|
Twine(ParentBB->getName()) + ".unreachable_default",
|
|
|
|
ParentBB->getParent());
|
|
|
|
new UnreachableInst(ParentBB->getContext(), UnreachableBB);
|
|
|
|
SI.setDefaultDest(UnreachableBB);
|
|
|
|
DT.addNewBlock(UnreachableBB, ParentBB);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// If we're not unswitching the default, we need it to match any cases to
|
|
|
|
// have a common successor or if we have no cases it is the common
|
|
|
|
// successor.
|
|
|
|
if (SI.getNumCases() == 0)
|
|
|
|
CommonSuccBB = SI.getDefaultDest();
|
|
|
|
else if (SI.getDefaultDest() != CommonSuccBB)
|
|
|
|
CommonSuccBB = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Split the preheader, so that we know that there is a safe place to insert
|
|
|
|
// the switch.
|
|
|
|
BasicBlock *OldPH = L.getLoopPreheader();
|
|
|
|
BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI);
|
|
|
|
OldPH->getTerminator()->eraseFromParent();
|
|
|
|
|
|
|
|
// Now add the unswitched switch.
|
|
|
|
auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
|
|
|
|
|
2017-05-12 10:19:59 +08:00
|
|
|
// Rewrite the IR for the unswitched basic blocks. This requires two steps.
|
|
|
|
// First, we split any exit blocks with remaining in-loop predecessors. Then
|
|
|
|
// we update the PHIs in one of two ways depending on if there was a split.
|
|
|
|
// We walk in reverse so that we split in the same order as the cases
|
|
|
|
// appeared. This is purely for convenience of reading the resulting IR, but
|
|
|
|
// it doesn't cost anything really.
|
|
|
|
SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
|
|
|
|
// Handle the default exit if necessary.
|
|
|
|
// FIXME: It'd be great if we could merge this with the loop below but LLVM's
|
|
|
|
// ranges aren't quite powerful enough yet.
|
2017-05-12 10:19:59 +08:00
|
|
|
if (DefaultExitBB) {
|
|
|
|
if (pred_empty(DefaultExitBB)) {
|
|
|
|
UnswitchedExitBBs.insert(DefaultExitBB);
|
|
|
|
rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
|
|
|
|
} else {
|
|
|
|
auto *SplitBB =
|
|
|
|
SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI);
|
|
|
|
rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
|
|
|
|
*ParentBB, *OldPH);
|
|
|
|
updateLoopExitIDom(DefaultExitBB, L, DT);
|
|
|
|
DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
|
|
|
|
}
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
}
|
|
|
|
// Note that we must use a reference in the for loop so that we update the
|
|
|
|
// container.
|
|
|
|
for (auto &CasePair : reverse(ExitCases)) {
|
|
|
|
// Grab a reference to the exit block in the pair so that we can update it.
|
2017-05-12 10:19:59 +08:00
|
|
|
BasicBlock *ExitBB = CasePair.second;
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
|
|
|
|
// If this case is the last edge into the exit block, we can simply reuse it
|
|
|
|
// as it will no longer be a loop exit. No mapping necessary.
|
2017-05-12 10:19:59 +08:00
|
|
|
if (pred_empty(ExitBB)) {
|
|
|
|
// Only rewrite once.
|
|
|
|
if (UnswitchedExitBBs.insert(ExitBB).second)
|
|
|
|
rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
continue;
|
2017-05-12 10:19:59 +08:00
|
|
|
}
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
|
|
|
|
// Otherwise we need to split the exit block so that we retain an exit
|
|
|
|
// block from the loop and a target for the unswitched condition.
|
|
|
|
BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
|
|
|
|
if (!SplitExitBB) {
|
|
|
|
// If this is the first time we see this, do the split and remember it.
|
|
|
|
SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI);
|
2017-05-12 10:19:59 +08:00
|
|
|
rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
|
|
|
|
*ParentBB, *OldPH);
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
updateLoopExitIDom(ExitBB, L, DT);
|
|
|
|
}
|
2017-05-12 10:19:59 +08:00
|
|
|
// Update the case pair to point to the split block.
|
|
|
|
CasePair.second = SplitExitBB;
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// Now add the unswitched cases. We do this in reverse order as we built them
|
|
|
|
// in reverse order.
|
|
|
|
for (auto CasePair : reverse(ExitCases)) {
|
|
|
|
ConstantInt *CaseVal = CasePair.first;
|
|
|
|
BasicBlock *UnswitchedBB = CasePair.second;
|
|
|
|
|
|
|
|
NewSI->addCase(CaseVal, UnswitchedBB);
|
|
|
|
updateDTAfterUnswitch(UnswitchedBB, OldPH, DT);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If the default was unswitched, re-point it and add explicit cases for
|
|
|
|
// entering the loop.
|
|
|
|
if (DefaultExitBB) {
|
|
|
|
NewSI->setDefaultDest(DefaultExitBB);
|
|
|
|
updateDTAfterUnswitch(DefaultExitBB, OldPH, DT);
|
|
|
|
|
|
|
|
// We removed all the exit cases, so we just copy the cases to the
|
|
|
|
// unswitched switch.
|
|
|
|
for (auto Case : SI.cases())
|
|
|
|
NewSI->addCase(Case.getCaseValue(), NewPH);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we ended up with a common successor for every path through the switch
|
|
|
|
// after unswitching, rewrite it to an unconditional branch to make it easy
|
|
|
|
// to recognize. Otherwise we potentially have to recognize the default case
|
|
|
|
// pointing at unreachable and other complexity.
|
|
|
|
if (CommonSuccBB) {
|
|
|
|
BasicBlock *BB = SI.getParent();
|
|
|
|
SI.eraseFromParent();
|
|
|
|
BranchInst::Create(CommonSuccBB, BB);
|
|
|
|
}
|
|
|
|
|
|
|
|
DT.verifyDomTree();
|
|
|
|
++NumTrivial;
|
|
|
|
++NumSwitches;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// This routine scans the loop to find a branch or switch which occurs before
|
|
|
|
/// any side effects occur. These can potentially be unswitched without
|
|
|
|
/// duplicating the loop. If a branch or switch is successfully unswitched the
|
|
|
|
/// scanning continues to see if subsequent branches or switches have become
|
|
|
|
/// trivial. Once all trivial candidates have been unswitched, this routine
|
|
|
|
/// returns.
|
|
|
|
///
|
|
|
|
/// The return value indicates whether anything was unswitched (and therefore
|
|
|
|
/// changed).
|
|
|
|
static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
|
|
|
|
LoopInfo &LI) {
|
|
|
|
bool Changed = false;
|
|
|
|
|
|
|
|
// If loop header has only one reachable successor we should keep looking for
|
|
|
|
// trivial condition candidates in the successor as well. An alternative is
|
|
|
|
// to constant fold conditions and merge successors into loop header (then we
|
|
|
|
// only need to check header's terminator). The reason for not doing this in
|
|
|
|
// LoopUnswitch pass is that it could potentially break LoopPassManager's
|
|
|
|
// invariants. Folding dead branches could either eliminate the current loop
|
|
|
|
// or make other loops unreachable. LCSSA form might also not be preserved
|
|
|
|
// after deleting branches. The following code keeps traversing loop header's
|
|
|
|
// successors until it finds the trivial condition candidate (condition that
|
|
|
|
// is not a constant). Since unswitching generates branches with constant
|
|
|
|
// conditions, this scenario could be very common in practice.
|
|
|
|
BasicBlock *CurrentBB = L.getHeader();
|
|
|
|
SmallPtrSet<BasicBlock *, 8> Visited;
|
|
|
|
Visited.insert(CurrentBB);
|
|
|
|
do {
|
|
|
|
// Check if there are any side-effecting instructions (e.g. stores, calls,
|
|
|
|
// volatile loads) in the part of the loop that the code *would* execute
|
|
|
|
// without unswitching.
|
|
|
|
if (llvm::any_of(*CurrentBB,
|
|
|
|
[](Instruction &I) { return I.mayHaveSideEffects(); }))
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
TerminatorInst *CurrentTerm = CurrentBB->getTerminator();
|
|
|
|
|
|
|
|
if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
|
|
|
|
// Don't bother trying to unswitch past a switch with a constant
|
|
|
|
// condition. This should be removed prior to running this pass by
|
|
|
|
// simplify-cfg.
|
|
|
|
if (isa<Constant>(SI->getCondition()))
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
if (!unswitchTrivialSwitch(L, *SI, DT, LI))
|
|
|
|
// Coludn't unswitch this one so we're done.
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// Mark that we managed to unswitch something.
|
|
|
|
Changed = true;
|
|
|
|
|
|
|
|
// If unswitching turned the terminator into an unconditional branch then
|
|
|
|
// we can continue. The unswitching logic specifically works to fold any
|
|
|
|
// cases it can into an unconditional branch to make it easier to
|
|
|
|
// recognize here.
|
|
|
|
auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
|
|
|
|
if (!BI || BI->isConditional())
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
CurrentBB = BI->getSuccessor(0);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
auto *BI = dyn_cast<BranchInst>(CurrentTerm);
|
|
|
|
if (!BI)
|
|
|
|
// We do not understand other terminator instructions.
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// Don't bother trying to unswitch past an unconditional branch or a branch
|
|
|
|
// with a constant value. These should be removed by simplify-cfg prior to
|
|
|
|
// running this pass.
|
|
|
|
if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// Found a trivial condition candidate: non-foldable conditional branch. If
|
|
|
|
// we fail to unswitch this, we can't do anything else that is trivial.
|
|
|
|
if (!unswitchTrivialBranch(L, *BI, DT, LI))
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// Mark that we managed to unswitch something.
|
|
|
|
Changed = true;
|
|
|
|
|
|
|
|
// We unswitched the branch. This should always leave us with an
|
|
|
|
// unconditional branch that we can follow now.
|
|
|
|
BI = cast<BranchInst>(CurrentBB->getTerminator());
|
|
|
|
assert(!BI->isConditional() &&
|
|
|
|
"Cannot form a conditional branch by unswitching1");
|
|
|
|
CurrentBB = BI->getSuccessor(0);
|
|
|
|
|
|
|
|
// When continuing, if we exit the loop or reach a previous visited block,
|
|
|
|
// then we can not reach any trivial condition candidates (unfoldable
|
|
|
|
// branch instructions or switch instructions) and no unswitch can happen.
|
|
|
|
} while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
|
|
|
|
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Unswitch control flow predicated on loop invariant conditions.
|
|
|
|
///
|
|
|
|
/// This first hoists all branches or switches which are trivial (IE, do not
|
|
|
|
/// require duplicating any part of the loop) out of the loop body. It then
|
|
|
|
/// looks at other loop invariant control flows and tries to unswitch those as
|
|
|
|
/// well by cloning the loop if the result is small enough.
|
|
|
|
static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
|
|
|
|
AssumptionCache &AC) {
|
|
|
|
assert(L.isLCSSAForm(DT) &&
|
|
|
|
"Loops must be in LCSSA form before unswitching.");
|
|
|
|
bool Changed = false;
|
|
|
|
|
|
|
|
// Must be in loop simplified form: we need a preheader and dedicated exits.
|
|
|
|
if (!L.isLoopSimplifyForm())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Try trivial unswitch first before loop over other basic blocks in the loop.
|
|
|
|
Changed |= unswitchAllTrivialConditions(L, DT, LI);
|
|
|
|
|
|
|
|
// FIXME: Add support for non-trivial unswitching by cloning the loop.
|
|
|
|
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
|
|
|
|
LoopStandardAnalysisResults &AR,
|
|
|
|
LPMUpdater &U) {
|
|
|
|
Function &F = *L.getHeader()->getParent();
|
|
|
|
(void)F;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L << "\n");
|
|
|
|
|
|
|
|
if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC))
|
|
|
|
return PreservedAnalyses::all();
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
// Historically this pass has had issues with the dominator tree so verify it
|
|
|
|
// in asserts builds.
|
|
|
|
AR.DT.verifyDomTree();
|
|
|
|
#endif
|
|
|
|
return getLoopPassPreservedAnalyses();
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
2017-05-17 07:10:25 +08:00
|
|
|
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
class SimpleLoopUnswitchLegacyPass : public LoopPass {
|
|
|
|
public:
|
|
|
|
static char ID; // Pass ID, replacement for typeid
|
2017-05-17 07:10:25 +08:00
|
|
|
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
explicit SimpleLoopUnswitchLegacyPass() : LoopPass(ID) {
|
|
|
|
initializeSimpleLoopUnswitchLegacyPassPass(
|
|
|
|
*PassRegistry::getPassRegistry());
|
|
|
|
}
|
|
|
|
|
|
|
|
bool runOnLoop(Loop *L, LPPassManager &LPM) override;
|
|
|
|
|
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override {
|
|
|
|
AU.addRequired<AssumptionCacheTracker>();
|
|
|
|
getLoopAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
};
|
2017-05-17 07:10:25 +08:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
[PM/LoopUnswitch] Introduce a new, simpler loop unswitch pass.
Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
2017-04-28 02:45:20 +08:00
|
|
|
|
|
|
|
bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
|
|
|
|
if (skipLoop(L))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
Function &F = *L->getHeader()->getParent();
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L << "\n");
|
|
|
|
|
|
|
|
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
|
|
|
|
auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
|
|
|
|
auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
|
|
|
|
|
|
|
|
bool Changed = unswitchLoop(*L, DT, LI, AC);
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
// Historically this pass has had issues with the dominator tree so verify it
|
|
|
|
// in asserts builds.
|
|
|
|
DT.verifyDomTree();
|
|
|
|
#endif
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
char SimpleLoopUnswitchLegacyPass::ID = 0;
|
|
|
|
INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
|
|
|
|
"Simple unswitch loops", false, false)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LoopPass)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
|
|
|
|
INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
|
|
|
|
"Simple unswitch loops", false, false)
|
|
|
|
|
|
|
|
Pass *llvm::createSimpleLoopUnswitchLegacyPass() {
|
|
|
|
return new SimpleLoopUnswitchLegacyPass();
|
|
|
|
}
|