Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request #351 from MatthewFluet/c-and-llvm-codegen-updates
Updates to C and LLVM codegens. Highlights:
* Add `Machine.Program.rflow` to compute `{returns,raises}To` control
flow (654c557) and use in `functor Chunkify` (1b3b7b8) and in
Machine IR `Raise/Return` transfers (cf8e487).
* Add `chunk-jump-table {false|true}` compile-time option to force
generation of a jump table for the chunk switch (8e0dd2d,
5b6439b, 087a5b1).
* Add `-chunk-{{must,may}-rto-self,must-rto-sing,must-rto-other}-opt`
compile-time options to optimize return/raise transfers (7c10c70,
4d5abde, 4b7c649, c3b9905, 473808f)
* Experiment using LLVM's `cc10` (aka, `ghccc`) calling convention
(2e26ebd).
* Experiment with a new `simple` chunkify strategy (3330cbe,
3d9c499, 138512f, faef164, d1df0de); generally performs
about the same as `coalesce4096`, significantly improves `fib` and
`tak` (for GCC), slightly improves `hamlet`, but slightly worsens
`raytrace`:
config command
C04 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen c -cc gcc-9
C05 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen c -cc gcc-9 -chunkify simple
C09 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen llvm -cc clang
C10 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen llvm -cc clang -chunkify simple
task_clock [email protected] (2-level)
program `C05/C04` `C10/C09`
barnes-hut 0.9978 0.9589
boyer 1.064 1.076
checksum 1.051 0.9775
count-graphs 1.005 0.9876
DLXSimulator 1.000 0.9905
even-odd 1.037 0.9989
fft 0.9616 0.9537
fib 0.6689 0.6260
flat-array 1.000 0.9645
hamlet 0.9547 0.9322
imp-for 1.067 1.014
knuth-bendix 1.092 1.031
lexgen 1.031 1.078
life 1.002 0.9911
logic 1.016 1.015
mandelbrot 0.9776 1.030
matrix-multiply 0.9903 0.9844
md5 1.008 0.9940
merge 0.9927 1.062
mlyacc 0.9810 1.024
model-elimination 0.9877 0.9743
mpuz 1.011 1.010
nucleic 1.036 1.030
output1 0.9943 1.021
peek 1.036 1.027
pidigits 1.000 0.9653
psdes-random 1.009 1.014
ratio-regions 0.9985 0.9881
ray 0.9738 0.9601
raytrace 1.101 1.100
simple 0.9620 0.9272
smith-normal-form 0.9690 0.9806
string-concat 0.9610 0.9772
tailfib 1.006 0.9292
tailmerge 0.9847 1.023
tak 0.8264 1.013
tensor 1.010 0.9998
tsp 0.9981 1.010
tyan 1.045 1.027
vector-rev 1.012 0.9891
vector32-concat 0.9495 1.030
vector64-concat 1.098 0.9744
vliw 0.9413 1.019
wc-input1 0.9301 1.098
wc-scanStream 1.114 0.9234
zebra 1.008 1.001
zern 0.9819 1.014
MIN 0.6689 0.6260
GMEAN 0.9940 0.9912
MAX 1.114 1.100
The `simple` chunkify strategy is not (yet) suitable for a
self-compile; it can generate excessively large chunks, including
one for a self-compile that requires 8min to compile by `gcc`.
* Add `expect: WordX.t option` to RSSA and Machine `Switch.T`
(911b5d4, e2b27ab, 695320d) and add `-gc-expect
{none|false|true}` compile-time option, where `-gc-expect false`
should indicate that performing a GC is cold path (823815a); no
notable performance impact.
* Lots of tweaks to C codegen, ultimately eliminating almost all
`c-chunk.h` macros.
* Eliminate unused `Machine.Operand.Contents` constructor (006269b).
* Make a major refactoring of LLVM codegen (cec30c5).
* Implement `Real<N>_qequal` for C codegen (9b7b2bd) and use
`is{less,lessequal}` for `Real<N>_l{t,e}` for C codegen (7b55819).
* Generalize LLVM type-based alias-analysis (27709ef).
* Add `-llvm-aamd scope` for simple `noalias`/`alias.scope`
alias-analysis metadata in LLVM codegen (b825f56); no notable
performance impact.
* Use C99/C11 `inline` for primitive and Basis Library functions
(311331c, c864492, 4f2d213).
* Add `-codegen-fuse-op-and-chk {false|true}` compile-time option to
explicitly fuse adjacent `Word<N>_<op>` and
`Word{S,U}<N>_<op>CheckP` primitives in the C and LLVM codegens
(6b738b8, 3d1e89c, 68f8512, 82c019f, 61de560, 5363199,
0d46a85). It appears that GCC (and, to a lesser extent)
Clang/LLVM do not always successfully fuse adjacent adjacent
`Word<N>_<op>` and `Word{S,U}<N>_<op>CheckP` primitives. The
performance results reported at
#273 and
#292 suggest that this does not
always have significant impact, but sometimes
`-codegen-fuse-op-and-chk true` can have a positive. Unfortunately,
it can also have a (significant) negative impact. In
`matrix-multiply` and `vector-rev`, fusing can cause GCC to not
recognize that an explicit sequence index can be replaced by a
stride length; in these benchmarks, it would be nice if MLton
eliminated the overflow checks.
config command
C04 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen c -cc gcc-9
C09 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen llvm -cc clang
C11 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen c -cc gcc-9 -codegen-fuse-op-and-chk true
C15 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen llvm -cc clang -codegen-fuse-op-and-chk true
task_clock [email protected] (2-level)
program `C11/C04` `C15/C09`
barnes-hut 1.005 0.9925
boyer 1.052 1.013
checksum 1.022 1.028
count-graphs 0.9722 1.002
DLXSimulator 1.004 0.9959
even-odd 0.8768 1.003
fft 0.9592 1.016
fib 0.9732 0.9798
flat-array 0.8148 1.019
hamlet 0.9966 1.030
imp-for 0.8993 0.7985
knuth-bendix 1.008 1.013
lexgen 0.9851 1.043
life 0.9954 1.006
logic 0.9994 1.014
mandelbrot 0.9440 1.013
matrix-multiply 1.336 1.009
md5 0.9604 1.007
merge 0.9675 1.037
mlyacc 1.032 1.029
model-elimination 1.010 1.004
mpuz 1.035 0.9599
nucleic 0.9938 0.9983
output1 0.9278 0.9709
peek 0.9850 1.035
pidigits 0.9702 0.9538
psdes-random 1.017 0.9986
ratio-regions 0.9801 0.9887
ray 0.9795 1.009
raytrace 0.9959 1.026
simple 0.9764 1.010
smith-normal-form 1.002 1.049
string-concat 0.7919 0.9035
tailfib 1.030 1.227
tailmerge 1.017 0.9980
tak 0.9790 0.9988
tensor 0.5258 1.000
tsp 0.9845 1.013
tyan 1.019 0.9739
vector-rev 1.178 1.253
vector32-concat 0.8703 0.9230
vector64-concat 0.8906 0.9038
vliw 0.9921 1.044
wc-input1 1.060 0.9809
wc-scanStream 0.9166 1.040
zebra 1.008 1.020
zern 1.051 1.089
MIN 0.5258 0.7985
GMEAN 0.9720 1.007
MAX 1.336 1.253
Note: the issue with `md5` mentioned in the commit messages are with
respect to the `md5` benchmark before 2daaebf.
Overall, this simplifies the C and LLVM codegen slightly, although
there is little significant performance change:
config command
C02 /home/mtf/devel/mlton/builds/g89891a411/bin/mlton -codegen c -cc gcc-9
C04 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen c -cc gcc-9
C08 /home/mtf/devel/mlton/builds/g89891a411/bin/mlton -codegen llvm -cc clang
C09 /home/mtf/devel/mlton/builds/g098009d49/bin/mlton -codegen llvm -cc clang
task_clock [email protected] (2-level)
program `C04/C02` `C09/C08`
barnes-hut 1.036 1.025
boyer 0.9731 1.006
checksum 0.9652 1.002
count-graphs 0.9988 0.9964
DLXSimulator 0.9970 1.023
even-odd 1.002 0.9881
fft 1.026 0.9674
fib 0.9034 0.7846
flat-array 1.014 1.021
hamlet 0.9740 1.010
imp-for 0.9707 0.9908
knuth-bendix 0.9077 0.9777
lexgen 1.048 0.8985
life 1.002 0.9827
logic 1.006 0.9867
mandelbrot 1.000 1.011
matrix-multiply 1.020 0.9957
md5 0.9700 0.9960
merge 0.9974 0.9818
mlyacc 1.003 0.9824
model-elimination 0.9936 0.9817
mpuz 0.9815 0.9466
nucleic 0.9946 1.002
output1 1.007 1.026
peek 0.9832 0.9898
pidigits 0.9950 1.047
psdes-random 1.009 0.9869
ratio-regions 0.9978 0.9725
ray 0.9938 0.9663
raytrace 0.9975 1.032
simple 0.9936 1.000
smith-normal-form 1.038 0.9941
string-concat 1.041 1.014
tailfib 0.9865 0.9741
tailmerge 1.010 1.020
tak 0.9331 0.9041
tensor 0.9938 0.9941
tsp 0.9825 1.004
tyan 0.9960 0.9879
vector-rev 1.014 0.9091
vector32-concat 1.090 0.9016
vector64-concat 0.9994 0.9800
vliw 0.9995 0.9876
wc-input1 0.9685 0.8634
wc-scanStream 1.178 1.105
zebra 0.9857 0.9900
zern 0.9733 0.9890
MIN 0.9034 0.7846
GMEAN 0.9982 0.9815
MAX 1.178 1.105- Loading branch information
