Remove CafInfo predictions from the compiler, assign final CafInfos to Ids after codegen
-
Review changes -
-
Download -
Patches
-
Plain diff
(See also: https://gitlab.haskell.org/ghc/ghc/wikis/CafInfo-rework)
(Note: This text will be the commit message in the final version)
This patch removes all CafInfo predictions and various hacks to preserve predicted CafInfos from the compiler and assigns final CafInfos to interface Ids after code generation. SRT analysis is extended to support static data, and Cmm generator is modified to allow generating static_link fields after SRT analysis.
This also fixes -fcatch-bottoms, which introduces error calls in case
expressions in CorePrep, which runs after CoreTidy (which is where we decide
on CafInfos) and turns previously non-CAFFY things into CAFFY. Fixes #17648 (closed).
Evaluation
NoFib
Boot with: make boot mode=fast
Run: make mode=fast EXTRA_RUNTEST_OPTS="-cachegrind" NoFibRuns=1
Results:
--------------------------------------------------------------------------------
Program Size Allocs Instrs Reads Writes
--------------------------------------------------------------------------------
CS -0.0% 0.0% -0.0% -0.0% -0.0%
CSD -0.0% 0.0% -0.0% -0.0% -0.0%
FS -0.0% 0.0% -0.0% -0.0% -0.0%
S -0.0% 0.0% -0.0% -0.0% -0.0%
VS -0.0% 0.0% -0.0% -0.0% -0.0%
VSD -0.0% 0.0% -0.0% -0.0% -0.5%
VSM -0.0% 0.0% -0.0% -0.0% -0.0%
anna -0.1% 0.0% -0.0% -0.0% -0.0%
ansi -0.0% 0.0% -0.0% -0.0% -0.0%
atom -0.0% 0.0% -0.0% -0.0% -0.0%
awards -0.0% 0.0% -0.0% -0.0% -0.0%
banner -0.0% 0.0% -0.0% -0.0% -0.0%
bernouilli -0.0% 0.0% -0.0% -0.0% -0.0%
binary-trees -0.0% 0.0% -0.0% -0.0% -0.0%
boyer -0.0% 0.0% -0.0% -0.0% -0.0%
boyer2 -0.0% 0.0% -0.0% -0.0% -0.0%
bspt -0.0% 0.0% -0.0% -0.0% -0.0%
cacheprof -0.0% 0.0% -0.0% -0.0% -0.0%
calendar -0.0% 0.0% -0.0% -0.0% -0.0%
cichelli -0.0% 0.0% -0.0% -0.0% -0.0%
circsim -0.0% 0.0% -0.0% -0.0% -0.0%
clausify -0.0% 0.0% -0.0% -0.0% -0.0%
comp_lab_zift -0.0% 0.0% -0.0% -0.0% -0.0%
compress -0.0% 0.0% -0.0% -0.0% -0.0%
compress2 -0.0% 0.0% -0.0% -0.0% -0.0%
constraints -0.0% 0.0% -0.0% -0.0% -0.0%
cryptarithm1 -0.0% 0.0% -0.0% -0.0% -0.0%
cryptarithm2 -0.0% 0.0% -0.0% -0.0% -0.0%
cse -0.0% 0.0% -0.0% -0.0% -0.0%
digits-of-e1 -0.0% 0.0% -0.0% -0.0% -0.0%
digits-of-e2 -0.0% 0.0% -0.0% -0.0% -0.0%
dom-lt -0.0% 0.0% -0.0% -0.0% -0.0%
eliza -0.0% 0.0% -0.0% -0.0% -0.0%
event -0.0% 0.0% -0.0% -0.0% -0.0%
exact-reals -0.0% 0.0% -0.0% -0.0% -0.0%
exp3_8 -0.0% 0.0% -0.0% -0.0% -0.0%
expert -0.0% 0.0% -0.0% -0.0% -0.0%
fannkuch-redux -0.0% 0.0% -0.0% -0.0% -0.0%
fasta -0.0% 0.0% -0.0% -0.0% -0.0%
fem -0.0% 0.0% -0.0% -0.0% -0.0%
fft -0.0% 0.0% -0.0% -0.0% -0.0%
fft2 -0.0% 0.0% -0.0% -0.0% -0.0%
fibheaps -0.0% 0.0% -0.0% -0.0% -0.0%
fish -0.0% 0.0% -0.0% -0.0% -0.0%
fluid -0.1% 0.0% -0.0% -0.0% -0.0%
fulsom -0.0% 0.0% -0.0% -0.0% -0.0%
gamteb -0.0% 0.0% -0.0% -0.0% -0.0%
gcd -0.0% 0.0% -0.0% -0.0% -0.0%
gen_regexps -0.0% 0.0% -0.0% -0.0% -0.0%
genfft -0.0% 0.0% -0.0% -0.0% -0.0%
gg -0.0% 0.0% -0.0% -0.0% -0.0%
grep -0.0% 0.0% -0.0% -0.0% -0.0%
hidden -0.0% 0.0% -0.0% -0.0% -0.0%
hpg -0.1% 0.0% -0.0% -0.0% -0.0%
ida -0.0% 0.0% -0.0% -0.0% -0.0%
infer -0.0% 0.0% -0.0% -0.0% -0.0%
integer -0.0% 0.0% -0.0% -0.0% -0.0%
integrate -0.0% 0.0% -0.0% -0.0% -0.0%
k-nucleotide -0.0% 0.0% -0.0% -0.0% -0.0%
kahan -0.0% 0.0% -0.0% -0.0% -0.0%
knights -0.0% 0.0% -0.0% -0.0% -0.0%
lambda -0.0% 0.0% -0.0% -0.0% -0.0%
last-piece -0.0% 0.0% -0.0% -0.0% -0.0%
lcss -0.0% 0.0% -0.0% -0.0% -0.0%
life -0.0% 0.0% -0.0% -0.0% -0.0%
lift -0.0% 0.0% -0.0% -0.0% -0.0%
linear -0.1% 0.0% -0.0% -0.0% -0.0%
listcompr -0.0% 0.0% -0.0% -0.0% -0.0%
listcopy -0.0% 0.0% -0.0% -0.0% -0.0%
maillist -0.0% 0.0% -0.0% -0.0% -0.0%
mandel -0.0% 0.0% -0.0% -0.0% -0.0%
mandel2 -0.0% 0.0% -0.0% -0.0% -0.0%
mate -0.0% 0.0% -0.0% -0.0% -0.0%
minimax -0.0% 0.0% -0.0% -0.0% -0.0%
mkhprog -0.0% 0.0% -0.0% -0.0% -0.0%
multiplier -0.0% 0.0% -0.0% -0.0% -0.0%
n-body -0.0% 0.0% -0.0% -0.0% -0.0%
nucleic2 -0.0% 0.0% -0.0% -0.0% -0.0%
para -0.0% 0.0% -0.0% -0.0% -0.0%
paraffins -0.0% 0.0% -0.0% -0.0% -0.0%
parser -0.1% 0.0% -0.0% -0.0% -0.0%
parstof -0.1% 0.0% -0.0% -0.0% -0.0%
pic -0.0% 0.0% -0.0% -0.0% -0.0%
pidigits -0.0% 0.0% -0.0% -0.0% -0.0%
power -0.0% 0.0% -0.0% -0.0% -0.0%
pretty -0.0% 0.0% -0.3% -0.4% -0.4%
primes -0.0% 0.0% -0.0% -0.0% -0.0%
primetest -0.0% 0.0% -0.0% -0.0% -0.0%
prolog -0.0% 0.0% -0.0% -0.0% -0.0%
puzzle -0.0% 0.0% -0.0% -0.0% -0.0%
queens -0.0% 0.0% -0.0% -0.0% -0.0%
reptile -0.0% 0.0% -0.0% -0.0% -0.0%
reverse-complem -0.0% 0.0% -0.0% -0.0% -0.0%
rewrite -0.0% 0.0% -0.0% -0.0% -0.0%
rfib -0.0% 0.0% -0.0% -0.0% -0.0%
rsa -0.0% 0.0% -0.0% -0.0% -0.0%
scc -0.0% 0.0% -0.3% -0.5% -0.4%
sched -0.0% 0.0% -0.0% -0.0% -0.0%
scs -0.0% 0.0% -0.0% -0.0% -0.0%
simple -0.1% 0.0% -0.0% -0.0% -0.0%
solid -0.0% 0.0% -0.0% -0.0% -0.0%
sorting -0.0% 0.0% -0.0% -0.0% -0.0%
spectral-norm -0.0% 0.0% -0.0% -0.0% -0.0%
sphere -0.0% 0.0% -0.0% -0.0% -0.0%
symalg -0.0% 0.0% -0.0% -0.0% -0.0%
tak -0.0% 0.0% -0.0% -0.0% -0.0%
transform -0.0% 0.0% -0.0% -0.0% -0.0%
treejoin -0.0% 0.0% -0.0% -0.0% -0.0%
typecheck -0.0% 0.0% -0.0% -0.0% -0.0%
veritas -0.0% 0.0% -0.0% -0.0% -0.0%
wang -0.0% 0.0% -0.0% -0.0% -0.0%
wave4main -0.0% 0.0% -0.0% -0.0% -0.0%
wheel-sieve1 -0.0% 0.0% -0.0% -0.0% -0.0%
wheel-sieve2 -0.0% 0.0% -0.0% -0.0% -0.0%
x2n1 -0.0% 0.0% -0.0% -0.0% -0.0%
--------------------------------------------------------------------------------
Min -0.1% 0.0% -0.3% -0.5% -0.5%
Max -0.0% 0.0% -0.0% -0.0% -0.0%
Geometric Mean -0.0% -0.0% -0.0% -0.0% -0.0%
--------------------------------------------------------------------------------
Program Size Allocs Instrs Reads Writes
--------------------------------------------------------------------------------
circsim -0.1% 0.0% -0.0% -0.0% -0.0%
constraints -0.0% 0.0% -0.0% -0.0% -0.0%
fibheaps -0.0% 0.0% -0.0% -0.0% -0.0%
gc_bench -0.0% 0.0% -0.0% -0.0% -0.0%
hash -0.0% 0.0% -0.0% -0.0% -0.0%
lcss -0.0% 0.0% -0.0% -0.0% -0.0%
power -0.0% 0.0% -0.0% -0.0% -0.0%
spellcheck -0.0% 0.0% -0.0% -0.0% -0.0%
--------------------------------------------------------------------------------
Min -0.1% 0.0% -0.0% -0.0% -0.0%
Max -0.0% 0.0% -0.0% -0.0% -0.0%
Geometric Mean -0.0% +0.0% -0.0% -0.0% -0.0%Manual inspection of programs in testsuite/tests/programs
I built these programs with a bunch of dump flags and -O and compared STG,
Cmm, and Asm dumps and file sizes.
(Below the numbers in parenthesis show number of modules in the program)
These programs have identical compiler (same .hi and .o sizes, STG, and Cmm and Asm dumps):
- Queens (1), andre_monad (1), cholewo-eval (2), cvh_unboxing (3), andy_cherry (7), fun_insts (1), hs-boot (4), fast2haskell (2), jl_defaults (1), jq_readsPrec (1), jules_xref (1), jtod_circint (4), jules_xref2 (1), lennart_range (1), lex (1), life_space_leak (1), bargon-mangler-bug (7), record_upd (1), rittri (1), sanders_array (1), strict_anns (1), thurston-module-arith (2), okeefe_neural (1), joao-circular (6), 10queens (1)
Programs with different compiler outputs:
-
jl_defaults (1): For some reason GHC HEAD marks a lot of top-level
[Int]closures as CAFFY for no reason. With this patch we no longer make them CAFFY and generate less SRT entries. For some reason Main.o is slightly larger with this patch (1.3%) and the executable sizes are the same. (I'd expect both to be smaller) -
launchbury (1): Same as jl_defaults: top-level
[Int]closures marked as CAFFY for no reason. SimilarlyMain.ois 1.4% larger but the executable sizes are the same. -
galois_raytrace (13): Differences are in the Parse module. There are a lot, but some of the changes are caused by the fact that for some reason (I think a bug) GHC HEAD marks the dictionary for
Functor Identityas CAFFY. Parse.o is 0.4% larger, the executable size is the same. -
north_array: We now generate less SRT entries because some of array primops used in this program like
NewArrayOpget eliminated during Stg-to-Cmm and turn some CAFFY things into non-CAFFY. Main.o gets 24% larger (9224 bytes from 9000 bytes), executable sizes are the same. -
seward-space-leak: Difference in this program is better shown by this smaller example:
module Lib where data CDS = Case [CDS] [(Int, CDS)] | Call CDS CDS instance Eq CDS where Case sels1 rets1 == Case sels2 rets2 = sels1 == sels2 && rets1 == rets2 Call a1 b1 == Call a2 b2 = a1 == a2 && b1 == b2 _ == _ = FalseIn this program GHC HEAD builds a new SRT for the recursive group of
(==),(/=)and the dictionary closure. Then/=points to==in its SRT field, and==uses the SRT object as its SRT. With this patch we use the closure for/=as the SRT and add==there. Then/=gets an empty SRT field and==points to/=in its SRT field.This change looks fine to me.
Main.o gets 0.07% larger, executable sizes are identical.
head.hackage
head.hackage's CI script builds 428 packages from Hackage using this patch with no failures.
Compiler performance
The compiler perf tests report that the compiler allocates slightly more (worst
case observed so far is 4%). However most programs in the test suite are small,
single file programs. To benchmark compiler performance on something more
realistic I build Cabal (the library, 236 modules) with different optimisation
levels. For the "max residency" row I run GHC with +RTS -s -A100k -i0 -h for
more accurate numbers. Other rows are generated with just -s. (This is because
-i0 causes running GC much more frequently and as a result "bytes copied" gets
inflated by more than 25x in some cases)
Runtime numbers are generated on a Ubuntu 18.04 system with most daemons and other background processes killed.
-O0
| GHC HEAD | This MR | Diff | |
|---|---|---|---|
| Bytes allocated | 54,413,350,872 | 54,701,099,464 | +0.52% |
| Bytes copied | 4,926,037,184 | 4,990,638,760 | +1.31% |
| Max residency | 421,225,624 | 424,324,264 | +0.73% |
| Runtime | 34.42s | 35.38s | +2.48% |
-O1
| GHC HEAD | This MR | Diff | |
|---|---|---|---|
| Bytes allocated | 245,849,209,992 | 246,562,088,672 | +0.28% |
| Bytes copied | 26,943,452,560 | 27,089,972,296 | +0.54% |
| Max residency | 982,643,440 | 991,663,432 | +0.91% |
| Runtime | 156.44s | 157.29s | +0.54% |
-O2
| GHC HEAD | This MR | Diff | |
|---|---|---|---|
| Bytes allocated | 291,044,511,408 | 291,863,910,912 | +0.28% |
| Bytes copied | 37,044,237,616 | 36,121,690,472 | -2.49% |
| Max residency | 1,071,600,328 | 1,086,396,256 | +1.38% |
| Runtime | 187.13s | 188.91s | +0.85% |
Extra compiler allocations
Runtime allocations of programs are as reported above (NoFib section).
The compiler now allocates more than before. Main source of allocation in this patch compared to base commit is the new SRT algorithm (GHC.Cmm.Info.Build). Below is some of the extra work we do with this patch, numbers generated by profiled stage 2 compiler when building a pathological case (the test 'ManyConstructors') with '-O2':
-
We now sort the final STG for a module, which means traversing the entire program, generating free variable set for each top-level binding, doing SCC analysis, and re-ordering the program. In ManyConstructors this step allocates 97,889,952 bytes.
-
We now do SRT analysis on static data, which in a program like ManyConstructors causes analysing 10,000 bindings that we would previously just skip. This step allocates 70,898,352 bytes.
-
We now maintain an SRT map for the entire module as we compile Cmm groups:
data ModuleSRTInfo = ModuleSRTInfo { ... , moduleSRTMap :: SRTMap }(SRTMap is just a strict Map from the 'containers' library)
This map gets an entry for most bindings in a module (exceptions are THUNKs and CAFFY static functions). For ManyConstructors this map gets 50015 entries.
-
Once we're done with code generation we generate a NameSet from SRTMap for the non-CAFFY names in the current module. This set gets the same number of entries as the SRTMap.
-
Finally we update CafInfos in ModDetails for the non-CAFFY Ids, using the NameSet generated in the previous step. This usually does the least amount of allocation among the work listed here.
Only place with this patch where we do less work in the CAF analysis in the
tidying pass (CoreTidy). However that doesn't save us much, as the pass still
needs to traverse the whole program and update IdInfos for other reasons. Only
thing we don't here do is the hasCafRefs pass over the RHS of bindings, which
is a stateless pass that returns a boolean value, so it doesn't allocate much.