Again co-developed with @bollu.
Based on top of: #3225
While hunting down the performance discrepancy on qsort.lean between C
and LLVM we noticed there was a single, trivially optimizeable, alloca
(LLVM's stack memory allocation instruction) that had load/stores in the
hot code path. We then found:
https://groups.google.com/g/llvm-dev/c/e90HiFcFF7Y.
TLDR: `mem2reg`, the pass responsible for getting rid of allocas if
possible, only triggers on an alloca if it is in the first BB. The
allocas of the current implementation get put right at the location
where they are needed -> they are ignored by mem2reg.
Thus we decided to add functionality that allows us to push all allocas
up into the first BB.
We initially wanted to write `buildPrologueAlloca` in a `withReader`
style so:
1. get the current position of the builder
2. jump to first BB and do the thing
3. revert position to the original
However the LLVM C API does not expose an option to obtain the current
position of an IR builder. Thus we ended up at the current
implementation which resets the builder position to the end of the BB
that the function was called from. This is valid because we never
operate anywhere but the end of the current BB in the LLVM emitter.
The numbers on the qsort benchmark got improved by the change as
expected, however we are not fully there yet:
```
C:
Benchmark 1: ./qsort.lean.out 400
Time (mean ± σ): 2.005 s ± 0.013 s [User: 1.996 s, System: 0.003 s]
Range (min … max): 1.993 s … 2.036 s 10 runs
LLVM before aligning the types
Benchmark 1: ./qsort.lean.out 400
Time (mean ± σ): 2.151 s ± 0.007 s [User: 2.146 s, System: 0.001 s]
Range (min … max): 2.142 s … 2.161 s 10 runs
LLVM after aligning the types
Benchmark 1: ./qsort.lean.out 400
Time (mean ± σ): 2.073 s ± 0.011 s [User: 2.067 s, System: 0.002 s]
Range (min … max): 2.060 s … 2.097 s 10 runs
LLVM after this
Benchmark 1: ./qsort.lean.out 400
Time (mean ± σ): 2.038 s ± 0.009 s [User: 2.032 s, System: 0.001 s]
Range (min … max): 2.027 s … 2.052 s 10 runs
```
Note: If you wish to merge this PR independently from its predecessor,
there is no technical dependency between the two, I'm merely stacking
them so we can see the performance impacts of each more clearly.
Debugged and authored in collaboration with @bollu.
This PR fixes several performance regressions of the LLVM backend
compared to the C backend
as described in #3192. We are now at the point where some benchmarks
from `tests/bench` achieve consistently equal and sometimes ever so
slightly better performance when using LLVM instead of C. However there
are still a few testcases where we are lacking behind ever so slightly.
The PR contains two changes:
1. Using the same types for `lean.h` runtime functions in the LLVM
backend as in `lean.h` it turns out that:
a) LLVM does not throw an error if we declare a function with a
different type than it actually has. This happened on multiple occasions
here, in particular when the function used `unsigned`, as it was
wrongfully assumed to be `size_t` sized.
b) Refuses to inline a function to the call site if such a type mismatch
occurs. This means that we did not inline important functionality such
as `lean_ctor_set` and were thus slowed down compared to the C backend
which did this correctly.
2. While developing this change we noticed that LLVM does treat the
following as invalid: Having a function declared with a certain type but
called with integers of a different type. However this will manifest in
completely nonsensical errors upon optimizing the bitcode file through
`leanc` such as:
```
error: Invalid record (Producer: 'LLVM15.0.7' Reader: 'LLVM 15.0.7')
```
Presumably because the generate .bc file is invalid in the first place.
Thus we added a call to `LLVMVerifyModule` before serializing the module
into a bitcode file. This ended producing the expected type errors from
LLVM an aborting the bitcode file generation as expected.
We manually checked each function in `lean.h` that is mentioned in
`EmitLLVM.lean` to make sure that all of their types align correctly
now.
Quick overview of the fast benchmarks as measured on my machine, 2 runs
of LLVM and 2 runs of C to get a feeling for how far the averages move:
- binarytrees: basically equal performance
- binarytrees.st: basically equal performance
- const_fold: equal if not slightly better for LLVM
- deriv: LLVM has 8% more instructions than C but same wall clock time
- liasolver: basically equal performance
- qsort: LLVM is slower by 7% instructions, 4% time. We have identified
why the generated code is slower (there is a store/load in a hot loop in
LLVM that is not in C) but not figured out why that happens/how to
address it.
- rbmap: LLVM has 3% less instructions and 13% less wall-clock time than
C (woop woop)
- rbmap_1 and rbmap_10 show similar behavior
- rbmap_fbip: LLVM has 2% more instructions but 2% better wall time
- rbmap_library: equal if not slightly better for LLVM
- unionfind: LLVM has 5% more instructions but 4% better wall time
Leaving out benchmarks related to the compiler itself as I was too lazy
to keep recompiling it from scratch until we are on a level with C.
Summing things up, it appears that LLVM has now caught up or surpassed
the C backend in the microbenchmarks for the most part. Next steps from
our side are:
- trying to win the qsort benchmark
- figuring out why/how LLVM runs more instructions for less wall-clock
time. My current guesses would be measurement noise and/or better use of
micro architecture?
- measuring the larger benchmarks as well
Switches from encoding `let_fun` using an annotated `(fun x : t => b) v`
expression to a function application `letFun v (fun x : t => b)`.
---------
Co-authored-by: Sebastian Ullrich <sebasti@nullri.ch>
This obviates the need to play weak linkage games when
we build `lean.h.bc` from `lean.h`. We perform the following steps:
1. We remove the `static` modifier from all definitions in `lean.h`.
This makes all definitions have `extern` linkage. Thus, when we build
a `lean.h.bc` using `clang`, we will actually get definitions
(instead of an empty file)
2. We build `lean.h.bc` from `lean.h` using `clang`.
3. When it comes time to link, we link
`current_module.bc := LLVMLinkModules2(current_module.bc, lean.h.bc)`.
4. We loop over every symbol that arrived from `lean.h.bc`
in `current_module.bc` and we then set this symbol to have
`internal` linkage. This simulates the effect of
`#include <lean.h>` where every definition in `lean.h`
has internal linkage.
This yajna, one hopes, pleases the linker gods.
* feat: add --target flag for LLVM backend to build objects of a different architecture
* chore: remove dead comment
* Update src/Lean/Compiler/IR/EmitLLVM.lean
Co-authored-by: Sebastian Ullrich <sebasti@nullri.ch>
* chore: normalize indentation in src/util/shell.cpp
* chore: strip trailing whitespace
Co-authored-by: Sebastian Ullrich <sebasti@nullri.ch>
Closes#2004.
In porting the bugfix from
6eb852e28f,
I noticed that the LLVM backend was incorrectly generating declaration
initializers (in `callIODeclInitFn`), by assuming the return type of the
initializer is the return type of the declaration. Rather, it must be be
`lean_object`, since the initializer returns an `IO a` value which must be unpacked.
TODO: stop using the `getOrCreateFunction` pattern pervasively.
perform the `create` at the right location, and the `get`
at the correct location.
