MSSA-based LICM has been enabled by default for a few years now.
This drops the old AST-based implementation. Using loop(licm) will
result in a fatal error, the use of loop-mssa(licm) is required
(or just licm, which defaults to loop-mssa).
Note that the core canSinkOrHoistInst() logic has to retain AST
support for now, because it is shared with LoopSink.
Differential Revision: https://reviews.llvm.org/D108244
This was a diagnostic option used to demonstrate a weakness in
the AST-based LICM implementation. This problem does not exist
in the MSSA-based LICM implementation, which has been enabled
for a long time now. As such, this option is no longer relevant.
This option has been enabled by default for quite a while now.
The practical impact of removing the option is that MSSA use
cannot be disabled in default pipelines (both LPM and NPM) and
in manual LPM invocations. NPM can still choose to enable/disable
MSSA using loop vs loop-mssa.
The next step will be to require MSSA for LICM and drop the
AST-based implementation entirely.
Differential Revision: https://reviews.llvm.org/D108075
Currently, LNICM pass does not support sinking instructions out of loop nest.
This patch enables LNICM to sink down as many instructions to the exit block of outermost loop as possible.
Reviewed By: Whitney
Differential Revision: https://reviews.llvm.org/D107219
When hoisting/moving calls to locations, we strip unknown metadata. Such calls are usually marked `speculatable`, i.e. they are guaranteed to not cause undefined behaviour when run anywhere. So, we should strip attributes that can cause immediate undefined behaviour if those attributes are not valid in the context where the call is moved to.
This patch introduces such an API and uses it in relevant passes. See
updated tests.
Fix for PR50744.
Reviewed By: nikic, jdoerfert, lebedev.ri
Differential Revision: https://reviews.llvm.org/D104641
This patch adds a new pass called LNICM which is a LoopNest version of LICM and a test case to show how LNICM works.
Basically, LNICM only hoists invariants out of loop nest (not a loop) to keep/make perfect loop nest. This enables later optimizations that require perfect loop nest.
Reviewed By: Whitney
Differential Revision: https://reviews.llvm.org/D104180
The MaybePromotable set keeps track of loads/stores for which
promotion was not attempted yet. Normally, any load/stores that
are promoted in the current iteration will be removed from this
set, because they naturally MustAlias with the promoted value.
However, if the source program has UB with metadata claiming that
a store is NoAlias, while it is actually MustAlias, and multiple
different pointers are promoted in the same iteration, it can
happen that a store is removed that is still in the MaybePromotable
set, causing a use-after-free.
While this could be fixed by explicitly invalidating values in
MaybePromotable in the LoopPromoter, I'm going with the more
radical option of dropping the set entirely here and check all
load/stores on each promotion iteration. As promotion, and especially
repeated promotion, are quite rare, this doesn't seem to have any
impact on compile-time.
Fixes https://bugs.llvm.org/show_bug.cgi?id=50367.
This appears to miscompile google benchmark's GetCacheSizesFromKVFS()
when compiling with -fstrict-vtable-pointers.
Runnable reproducer: https://godbolt.org/z/f9ovKqTzb
The "f.fail()" crashes with BUS error, it is compiled into testb,
and the adress it is testing is non-sensical.
This reverts commit 4c89bcadf6.
During store promotion, we check whether the pointer was captured
to exclude potential reads from other threads. However, we're only
interested in captures before or inside the loop. Check this using
PointerMayBeCapturedBefore against the loop header.
Differential Revision: https://reviews.llvm.org/D100706
Previously loading the vtable used in calling a virtual method in a loop
was not hoisted out of the loop. This fixes that.
canSinkOrHoistInst() itself doesn't check that the load operands are
loop invariant, callers also check that separately.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D99784
LICM can sink instructions that have uses inside the loop, as
long as these uses are considered "free". However, if there were
only free uses inside the loop, and no uses outside the loop at
all, the instruction would still count towards the NumSunk
statistic. This resulted in a wild inflation of the NumSunk metric.
After this patch it drops down from 1141787 to 5852 on test-suite O3.
It is not legal to form a phi node with token type. The generic LCSSA construction code handles this correctly - by not forming LCSSA for such cases - but the adhoc fixup implementation in LICM did not.
This was noticed in the context of PR49607, but can be demonstrated on ToT with the tweaked test case. This is not specific to gc.relocate btw, it also applies to usage of the preallocated family of intrinsics as well.
Differential Revision: https://reviews.llvm.org/D98728
Relative to the previous implementation, this always uses
aliasesUnknownInst() instead of aliasesPointer() to correctly
handle atomics. The added test case was previously miscompiled.
-----
Even when MemorySSA-based LICM is used, an AST is still populated
for scalar promotion. As the AST has quadratic complexity, a lot
of time is spent in this step despite the existing access count
limit. This patch optimizes the identification of promotable stores.
The idea here is pretty simple: We're only interested in must-alias
mod sets of loop invariant pointers. As such, only populate the AST
with loop-invariant loads and stores (anything else is definitely
not promotable) and then discard any sets which alias with any of
the remaining, definitely non-promotable accesses.
If we promoted something, check whether this has made some other
accesses loop invariant and thus possible promotion candidates.
This is much faster in practice, because we need to perform AA
queries for O(NumPromotable^2 + NumPromotable*NumNonPromotable)
instead of O(NumTotal^2), and NumPromotable tends to be small.
Additionally, promotable accesses have loop invariant pointers,
for which AA is cheaper.
This has a signicant positive compile-time impact. We save ~1.8%
geomean on CTMark at O3, with 6% on lencod in particular and 25%
on individual files.
Conceptually, this change is NFC, but may not be so in practice,
because the AST is only an approximation, and can produce
different results depending on the order in which accesses are
added. However, there is at least no impact on the number of promotions
(licm.NumPromoted) in test-suite O3 configuration with this change.
Differential Revision: https://reviews.llvm.org/D89264
Revert 3d8f842712
Revision triggers a miscompile sinking a store incorrectly outside a
threading loop. Detected by tsan.
Reverting while investigating.
Differential Revision: https://reviews.llvm.org/D89264
See pr46990(https://bugs.llvm.org/show_bug.cgi?id=46990). LICM should not sink store instructions to loop exit blocks which cross coro.suspend intrinsics. This breaks semantic of coro.suspend intrinsic which return to caller directly. Also this leads to use-after-free if the coroutine is freed before control returns to the caller in multithread environment.
This patch disable promotion by check whether loop contains coro.suspend intrinsics.
This is a resubmit of D86190.
Disabling LICM for loops with coroutine suspension is a better option not only for correctness purpose but also for performance purpose.
In most cases LICM sinks memory operations. In the case of coroutine, sinking memory operation out of the loop does not improve performance since coroutien needs to get data from the frame anyway. In fact LICM would hurt coroutine performance since it adds more entries to the frame.
Differential Revision: https://reviews.llvm.org/D96928
Even when MemorySSA-based LICM is used, an AST is still populated
for scalar promotion. As the AST has quadratic complexity, a lot
of time is spent in this step despite the existing access count
limit. This patch optimizes the identification of promotable stores.
The idea here is pretty simple: We're only interested in must-alias
mod sets of loop invariant pointers. As such, only populate the AST
with loop-invariant loads and stores (anything else is definitely
not promotable) and then discard any sets which alias with any of
the remaining, definitely non-promotable accesses.
If we promoted something, check whether this has made some other
accesses loop invariant and thus possible promotion candidates.
This is much faster in practice, because we need to perform AA
queries for O(NumPromotable^2 + NumPromotable*NumNonPromotable)
instead of O(NumTotal^2), and NumPromotable tends to be small.
Additionally, promotable accesses have loop invariant pointers,
for which AA is cheaper.
This has a signicant positive compile-time impact. We save ~1.8%
geomean on CTMark at O3, with 6% on lencod in particular and 25%
on individual files.
Conceptually, this change is NFC, but may not be so in practice,
because the AST is only an approximation, and can produce
different results depending on the order in which accesses are
added. However, there is at least no impact on the number of promotions
(licm.NumPromoted) in test-suite O3 configuration with this change.
Differential Revision: https://reviews.llvm.org/D89264
Added a utility function in Value class to print block name and use
block labels for unnamed blocks.
Changed LICM to call this function in its debug output.
Patch by Xiaoqing Wu <xiaoqing_wu@apple.com>
Differential Revision: https://reviews.llvm.org/D93577
Currently, we have some confusion in the codebase regarding the
meaning of LocationSize::unknown(): Some parts (including most of
BasicAA) assume that LocationSize::unknown() only allows accesses
after the base pointer. Some parts (various callers of AA) assume
that LocationSize::unknown() allows accesses both before and after
the base pointer (but within the underlying object).
This patch splits up LocationSize::unknown() into
LocationSize::afterPointer() and LocationSize::beforeOrAfterPointer()
to make this completely unambiguous. I tried my best to determine
which one is appropriate for all the existing uses.
The test changes in cs-cs.ll in particular illustrate a previously
clearly incorrect AA result: We were effectively assuming that
argmemonly functions were only allowed to access their arguments
after the passed pointer, but not before it. I'm pretty sure that
this was not intentional, and it's certainly not specified by
LangRef that way.
Differential Revision: https://reviews.llvm.org/D91649
Summary:
Expand existing loopsink testing to also test loopsinking using new pass
manager. Enable memoryssa for loopsink with new pass manager. This
combination exposed a bug that was previously fixed for loopsink
without memoryssa. When sinking an instruction into a loop, the source
block may not be part of the loop but still needs to be checked for
pointer invalidation. This is the fix for bugzilla #39695 (PR 54659)
expanded to also work with memoryssa.
Respond to review comments. Enable Memory SSA in legacy Loop Sink pass
under EnableMSSALoopDependency option control. Update tests accordingly.
Respond to review comments. Add options controlling whether memoryssa is
used for loop sink, defaulting to off. Expand testing based on these
options.
Respond to review comments. Properly indicated preserved analyses.
This relanding addresses a compile-time performance problem by moving
test for profile data earlier to avoid unnecessary computations.
Author: Jamie Schmeiser <schmeise@ca.ibm.com>
Reviewed By: asbirlea (Alina Sbirlea)
Differential Revision: https://reviews.llvm.org/D90249
This reverts commit 562addba65.
Reverted change too quickly, the failing test cases passed on the next build.
So reverting revert (to include the changes).
Summary:
Expand existing loopsink testing to also test loopsinking using new pass
manager. Enable memoryssa for loopsink with new pass manager. This
combination exposed a bug that was previously fixed for loopsink
without memoryssa. When sinking an instruction into a loop, the source
block may not be part of the loop but still needs to be checked for
pointer invalidation. This is the fix for bugzilla #39695 (PR 54659)
expanded to also work with memoryssa.
Respond to review comments. Enable Memory SSA in legacy Loop Sink pass
under EnableMSSALoopDependency option control. Update tests accordingly.
Respond to review comments. Add options controlling whether memoryssa is
used for loop sink, defaulting to off. Expand testing based on these
options.
Respond to review comments. Properly indicated preserved analyses.
Author: Jamie Schmeiser <schmeise@ca.ibm.com>
Reviewed By: asbirlea (Alina Sbirlea)
Differential Revision: https://reviews.llvm.org/D90249
Summary:
Refactor SinkAdHoistLICMFlags from a struct to a class with accessors and constructors to allow other
classes to construct flags with meaningful defaults while not exposing LICM internal details.
Author: Jamie Schmeiser <schmeise@ca.ibm.com>
Reviewed By: asbirlea (Alina Sbirlea)
Differential Revision: https://reviews.llvm.org/D90482
Results of convergent operations are implicitly affected by the
enclosing control flows and should not be hoisted out of arbitrary
loops.
Patch by Xiaoqing Wu <xiaoqing_wu@apple.com>
Differential Revision: https://reviews.llvm.org/D90361
While promotion currently always has an AST available, it is only
relevant for invalidation purposes in LoopPromoter, so we do not
need to have it as a hard dependency.
Introduce a helper which can be used to update the debug location of an
Instruction after the instruction is hoisted. This can be used to safely
drop a source location as recommended by the docs.
For more context, see the discussion in https://reviews.llvm.org/D60913.
Differential Revision: https://reviews.llvm.org/D85670
D65060 was reverted because it introduced non-determinism by using BFI counts from already freed blocks. The parent of this revision fixes that by using a VH callback on blocks to prevent this from happening and makes sure BFI data is passed correctly in LoopStandardAnalysisResults.
This re-introduces the previous optimization of using BFI data to prevent LICM from hoisting/sinking if the instruction will end up moving to a colder block.
Internally at Facebook this change results in a ~7% win in a CPU related metric in one of our big services by preventing hoisting cold code into a hot pre-header like the added test case demonstrates.
Testing:
ninja check
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D87551
~~D65060 uncovered that trying to use BFI in loop passes can lead to non-deterministic behavior when blocks are re-used while retaining old BFI data.~~
~~To make sure BFI is preserved through loop passes a Value Handle (VH) callback is registered on blocks themselves. When a block is freed it now also wipes out the accompanying BFI entry such that stale BFI data can no longer persist resolving the determinism issue. ~~
~~An optimistic approach would be to incrementally update BFI information throughout the loop passes rather than only invalidating them on removed blocks. The issues with that are:~~
~~1. It is not clear how BFI information should be incrementally updated: If a block is duplicated does its BFI information come with? How about if it's split/modified/moved around? ~~
~~2. Assuming we can address these problems the implementation here will be a massive undertaking. ~~
~~There's a known need of BFI in LICM analysis which requires correct but not incrementally updated BFI data. A follow-up change can register BFI in all loop passes so this preserved but potentially lossy data is available to any loop pass that wants it.~~
See: D75341 for an identical implementation of preserving BFI via VH callbacks. The previous statements do still apply but this change no longer has to be in this diff because it's already upstream 😄 .
This diff also moves BFI to be a part of LoopStandardAnalysisResults since the previous method using getCachedResults now (correctly!) statically asserts (D72893) that this data isn't static through the loop passes.
Testing
Ninja check
Reviewed By: asbirlea, nikic
Differential Revision: https://reviews.llvm.org/D86156
I've amended the isLoadInvariantInLoop function to bail out for
scalable vectors for now since the invariant.start intrinsic is only
ever generated by the clang frontend for thread locals or struct
and class constructors, neither of which support sizeless types.
In addition, the intrinsic itself does not currently support the
concept of a scaled size, which makes it impossible to compare
the sizes of different scalable objects, e.g. <vscale x 32 x i8>
and <vscale x 16 x i8>.
Added new tests here:
Transforms/LICM/AArch64/sve-load-hoist.ll
Transforms/LICM/hoisting.ll
Differential Revision: https://reviews.llvm.org/D87227
Introduce a helper on Instruction which can be used to update the debug
location after hoisting.
Use this in GVN and LICM, where we were mistakenly introducing new line
0 locations after hoisting (the docs recommend dropping the location in
this case).
For more context, see the discussion in https://reviews.llvm.org/D60913.
Differential Revision: https://reviews.llvm.org/D85670
- Now all SalvageDebugInfo() calls will mark undef if the salvage
attempt fails.
Reviewed by: vsk, Orlando
Differential Revision: https://reviews.llvm.org/D78369
It's really almost going to be misleading, see the example in
https://bugs.llvm.org/show_bug.cgi?id=45820
Maybe at some point we can do something fancier, but at least
this will fix a bug where we step on dead code while debugging.
There are several different types of cost that TTI tries to provide
explicit information for: throughput, latency, code size along with
a vague 'intersection of code-size cost and execution cost'.
The vectorizer is a keen user of RecipThroughput and there's at least
'getInstructionThroughput' and 'getArithmeticInstrCost' designed to
help with this cost. The latency cost has a single use and a single
implementation. The intersection cost appears to cover most of the
rest of the API.
getUserCost is explicitly called from within TTI when the user has
been explicit in wanting the code size (also only one use) as well
as a few passes which are concerned with a mixture of size and/or
a relative cost. In many cases these costs are closely related, such
as when multiple instructions are required, but one evident diverging
cost in this function is for div/rem.
This patch adds an argument so that the cost required is explicit,
so that we can make the important distinction when necessary.
Differential Revision: https://reviews.llvm.org/D78635
Nikita Popov