Reorganize the code into phases:
* Analyze/normalize
* Create extracted function prototype
* Generate the new function's implementation
* Generate call to new function
* Connect call to original function's CFG
The motivation is #114669 to optionally clone the selected code region
into the new function instead of moving it. The current structure made
it difficult to add such functionality since there was no obvious place
to do so, not made easier by some functions doing more than their name
suggests. For instance, constructFunction modifies code outside the
constructed function, but also function properties such as
setPersonalityFn are derived somewhere else. Another example is
emitCallAndSwitchStatement, which despite its name also inserts stores
for output parameters.
Many operations also implicitly depend on the order they are applied
which this patch tries to reduce. For instance, ExtractedFuncRetVals
becomes the list exit blocks which also defines the return value when
leaving via that block. It is computed early such that the new
function's return instructions and the switch can be generated
independently. Also, ExtractedFuncRetVals is combining the lists
ExitBlocks and OldTargets which were not always kept consistent with
each other or NumExitBlocks. The method recomputeExitBlocks() will
update it when necessary.
The coding style partially contradict the current coding standard. For
instance some local variable start with lower case letters. I updated
some, but not all occurrences to make the diff match at least some lines
as unchanged.
The patch [D96854](https://reviews.llvm.org/D96854) introduced some
confusion of function argument indexes this is fixed here as well, hence
the patch is not NFC anymore. Tested in modified CodeExtractorTest.cpp.
Patch [D121061](https://reviews.llvm.org/D121061) introduced
AllocationBlock, but not all allocas were inserted there.
Efectively includes the following fixes:
1. ce73b1672a
2. 4aaa925786
3. Missing allocas, still unfixed
Originally submitted as https://reviews.llvm.org/D115218
The code extractor tries to apply the names of source input and output
values to function arguments. Not all input and output values get added
as arguments: some are instead placed inside of a struct passed to the
function. The existing renaming code skipped trying to set these
struct-packed arguments names (as there is no corresponding function
argument to rename), but it still incremented the iterator over the
function arguments. This could result in dereferencing an end iterator
if struct-packed inputs/outputs preceded non-struct-packed
inputs/outputs.
This patch rewrites this loop to avoid the end iterator dereference.
This is an experimental address space for strided buffers. These buffers
can have structs as elements and
a stride > 1.
These pointers allow the indexed access in units of stride, i.e., they
point at `buffer[index * stride]`.
Thus, we can use the `idxen` modifier for buffer loads.
We assign address space 9 to 192-bit buffer pointers which contain a
128-bit descriptor, a 32-bit offset and a 32-bit index. Essentially,
they are fat buffer pointers with an additional 32-bit index.
The user of CodeExtractor should be able to specify that
the aggregate argument should be passed as a pointer in zero address
space.
CodeExtractor is used to generate outlined functions required by OpenMP
runtime. The arguments of the outlined functions for OpenMP GPU code
are in 0 address space. 0 address space does not need to be the default
address space for GPU device. That's why there is a need to allow
the user of CodeExtractor to specify, that the allocated aggregate parameter
is passed as pointer in zero address space.
Summary:
Enable CodeExtractor to construct output functions that partially
aggregate inputs/outputs in their argument list. A use case is the
OMPIRBuilder to create outlined functions for parallel regions that
aggregate in a struct the payload variables for the region while passing
as scalars thread and bound identifiers.
Differential Revision: https://reviews.llvm.org/D96854
Previously the CodeExtractor created exit stubs, and the subsequent return value of the outlined function based on the order of out-of-region blocks after splitting any phi nodes, and collecting the blocks to be outlined. This could cause differences in order if there was a difference of exit block phi nodes between the two regions. This patch moves the collection of the output target blocks to be before this occurs, so that the assignment of target block to output value will be the same, regardless of the contents of the output block.
Reviewers: paquette, roelofs
Differential Revision: https://reviews.llvm.org/D108657
The Code Extractor does not provide an easy mechanism for determining the
inputs and outputs after extraction has occurred, this patch gives the
ability to pass in empty SetVectors to be filled with the inputs and
outputs if they need to be analyzed.
Added Tests:
- InputOutputMonitoring in unittests/Transforms/Utils/CodeExtractorTests.cpp
Reviewers: paquette
Differential Revision: https://reviews.llvm.org/D106991
During extraction, stale llvm.assume handles may be retained in the
original function. The setup is:
1) CodeExtractor unregisters assumptions in the blocks that are to be
extracted.
2) Extraction happens. There are now two functions: f1 and f1.extracted.
3) Leftover assumptions in f1 (/not/ removed as they were not in the set of
blocks to be extracted) now have affected-value llvm.assume handles in
f1.extracted.
When assumptions for a value used in f1 are looked up, ValueTracking can assert
as some of the handles are in the wrong function. To fix this, simply erase the
llvm.assume calls in the extracted function.
Alternatives include flushing the assumption cache in the original function, or
walking all values used in the original function to prune stale affected-value
handles. Both seem more expensive.
Testing: check-llvm, LNT run with -mllvm -hot-cold-split enabled
rdar://58460728
Factor out CodeExtractor's analysis of allocas (for shrinkwrapping
purposes), and allow the analysis to be reused.
This resolves a quadratic compile-time bug observed when compiling
AMDGPUDisassembler.cpp.o.
Pre-patch (Release + LTO clang):
```
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
176.5278 ( 57.8%) 0.4915 ( 18.5%) 177.0192 ( 57.4%) 177.4112 ( 57.3%) Hot Cold Splitting
```
Post-patch (ReleaseAsserts clang):
```
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
1.4051 ( 3.3%) 0.0079 ( 0.3%) 1.4129 ( 3.2%) 1.4129 ( 3.2%) Hot Cold Splitting
```
Testing: check-llvm, and comparing the AMDGPUDisassembler.cpp.o binary
pre- vs. post-patch.
An alternate approach is to hide CodeExtractorAnalysisCache from clients
of CodeExtractor, and to recompute the analysis from scratch inside of
CodeExtractor::extractCodeRegion(). This eliminates some redundant work
in the shrinkwrapping legality check. However, some clients continue to
exhibit O(n^2) compile time behavior as computing the analysis is O(n).
rdar://55912966
Differential Revision: https://reviews.llvm.org/D68616
llvm-svn: 374089
When CodeExtractor saves the result of InvokeInst at the first insertion
point of the 'normal destination' basic block, this block can be omitted
in the outlined region, so store is placed outside of the function. The
suggested solution is to process saving outputs after creating exit
stubs for new function, and stores will be placed in that blocks before
return in this case.
Patch by Sergei Kachkov!
Fixes llvm.org/PR40455.
Differential Revision: https://reviews.llvm.org/D57919
llvm-svn: 353562
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
When CodeExtractor outlines values which are used by the original
function, it must store those values in some in-out parameter. This
store instruction must not be inserted in between a PHI and an EH pad
instruction, as that results in invalid IR.
This fixes the following verifier failure seen while outlining within
ObjC methods with live exit values:
The unwind destination does not have an exception handling instruction!
%call35 = invoke i8* bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to i8* (i8*, i8*)*)(i8* %exn.adjusted, i8* %1)
to label %invoke.cont34 unwind label %lpad33, !dbg !4183
The unwind destination does not have an exception handling instruction!
invoke void @objc_exception_throw(i8* %call35) #12
to label %invoke.cont36 unwind label %lpad33, !dbg !4184
LandingPadInst not the first non-PHI instruction in the block.
%3 = landingpad { i8*, i32 }
catch i8* null, !dbg !1411
rdar://46540815
llvm-svn: 348562
If a PHI node out of extracted region has multiple incoming values from it,
split this PHI on two parts. First PHI has incomings only from region and
extracts with it (they are placed to the separate basic block that added to the
list of outlined), and incoming values in original PHI are replaced by first
PHI. Similar solution is already used in CodeExtractor for PHIs in entry block
(severSplitPHINodes method). It covers PR39433 bug.
Patch by Sergei Kachkov!
Differential Revision: https://reviews.llvm.org/D55018
llvm-svn: 348205
The current splitting algorithm works in three stages:
1) Identify cold blocks, then
2) Use forward/backward propagation to mark hot blocks, then
3) Grow a SESE region of blocks *outside* of the set of hot blocks and
start outlining.
While testing this pass on Apple internal frameworks I noticed that some
kinds of control flow (e.g. loops) are never outlined, even though they
unconditionally lead to / follow cold blocks. I noticed two other issues
related to how cold regions are identified:
- An inconsistency can arise in the internal state of the hotness
propagation stage, as a block may end up in both the ColdBlocks set
and the HotBlocks set. Further inconsistencies can arise as these sets
do not match what's in ProfileSummaryInfo.
- It isn't necessary to limit outlining to single-exit regions.
This patch teaches the splitting algorithm to identify maximal cold
regions and outline them. A maximal cold region is defined as the set of
blocks post-dominated by a cold sink block, or dominated by that sink
block. This approach can successfully outline loops in the cold path. As
a side benefit, it maintains less internal state than the current
approach.
Due to a limitation in CodeExtractor, blocks within the maximal cold
region which aren't dominated by a single entry point (a so-called "max
ancestor") are filtered out.
Results:
- X86 (LNT + -Os + externals): 134KB of TEXT were outlined compared to
47KB pre-patch, or a ~3x improvement. Did not see a performance impact
across two runs.
- AArch64 (LNT + -Os + externals + Apple-internal benchmarks): 149KB
of TEXT were outlined. Ditto re: performance impact.
- Outlining results improve marginally in the internal frameworks I
tested.
Follow-ups:
- Outline more than once per function, outline large single basic
blocks, & try to remove unconditional branches in outlined functions.
Differential Revision: https://reviews.llvm.org/D53627
llvm-svn: 345209
