DEVNOTES.md
In brodieG/vetr: Trust, but Verify
gcc-10
See SO issue.
Issues related to compiling under gcc 10.0.1 causing segfaults on BDR machines.
To reproduce (though this will no longer work when debian:testing moves away
from from gcc-10.0.1):
docker pull rocker/r-base
docker run --rm -ti --security-opt seccomp=unconfined \
rocker/r-base /bin/bash
apt-get update
apt-get install gcc-10 gdb
gcc-10 --version # confirm 10.0.1
# gcc-10 (Debian 10-20200222-1) 10.0.1 20200222 (experimental)
# [master revision 01af7e0a0c2:487fe13f218:e99b18cf7101f205bfdd9f0f29ed51caaec52779]
mkdir ~/.R
touch ~/.R/Makevars
echo "CC = gcc-10
CFLAGS = -g -O2 -Wall -pedantic -mtune=native -Werror=format-security -Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector-strong -fstack-clash-protection -fcf-protection
" >> ~/.R/Makevars
R -d gdb --vanilla
Then in the R console, after typing run to get gdb to run the program:
f.dl <- tempfile()
f.uz <- tempfile()
github.url <- 'https://github.com/brodieG/vetr/archive/v0.2.8.zip'
download.file(github.url, f.dl)
unzip(f.dl, exdir=f.uz)
install.packages(file.path(f.uz, 'vetr-0.2.8'), repos=NULL, type='source')
unlink(c(f.dl, f.uz))
# minimal set of commands I could come up that would cause segfault follow
library(vetr)
alike(pairlist(a=1, b="character"), pairlist(a=1, b=letters))
alike(pairlist(1, "character"), pairlist(1, letters))
alike(NULL, 1:3) # not a wild card at top level
alike(list(NULL), list(1:3)) # but yes when nested
alike(list(NULL, NULL), list(list(list(1, 2, 3)), 1:25))
alike(list(NULL), list(1, 2))
alike(list(), list(1, 2))
alike(matrix(integer(), ncol=7), matrix(1:21, nrow=3))
alike(matrix(character(), nrow=3), matrix(1:21, nrow=3))
alike(
matrix(integer(), ncol=3, dimnames=list(NULL, c("R", "G", "B"))),
matrix(1:21, ncol=3, dimnames=list(NULL, c("R", "G", "B")))
)
# Adding tests from docs
mx.tpl <- matrix(
integer(), ncol=3, dimnames=list(row.id=NULL, c("R", "G", "B"))
)
mx.cur <- matrix(
sample(0:255, 12), ncol=3, dimnames=list(row.id=1:4, rgb=c("R", "G", "B"))
)
mx.cur2 <-
matrix(sample(0:255, 12), ncol=3, dimnames=list(1:4, c("R", "G", "B")))
alike(mx.tpl, mx.cur2)
Inspecting in gdb pretty quickly shows (if I understand correctly) that
CSR_strmlen_x is trying to access the string that was not initialized.
Notes from 8/29
Dig into the syntactic names issue:
unitizer> vetr:::syntactic_names(quote(mtcars[1:10, ]))
[1] FALSE
Notes from 8/22
Most of the alike side of things is dealt with, now we need to figure out how we
want to store stuff in lists. Used to be that we would just use LISTSXPs, but
now we need a growing collection of VALC_res objects? It should contain:
- An array of
ALIKEC_res objects
- The last inserted index
- The current allocated size
- By extension, a value in settings that has the default initial allocated size
We will need a mechanism to grow the array each time we hit the size limit.
Need to remember that we will have protected as many times as there are elements
in that array.
Also, need to figure out what functions need the added argument so we can pass
around our array pointer.
Notes from 8/21
So main thing to do is cleanly separate the "message" SEXP into the message
character values as part of the ALIKEC_res_interim object or some such, and
just the wrap object.
So we need to restructure the ALIKEC_res and ALIKEC_res_sub objects to
contain all the string structure. Additionally, it seems like the interim
object really doesn't need the call object? I guess it can be there explicitly
as a call object to distinguish it from the wrap object? The only real use
case is that after we've wrapped the call and are ready to return it to vetr
we need to put it someplace.
So we need objects that:
- Track the strings
- Track the strings + recursion index + wrap lang
- Track the strings + wrap lang + attr type
- Track the strings + final call
Also need to track success status, df status, attr level
ALIKEC_res_strings: just the stringsALIKEC_res_min: strings + successALIKEC_res_wrap: strings + success + wrap + [rec index] + [df] + [attr]ALIKEC_res: strings + success + call
One thing left to do: figure out what ALIKEC_wrap should be returning.
Notes from 8/20
Currently working through trying to delay the construction of error messages to
the very last minute.
Seems like the main issue is that that in VALC_evaluate, we need to change from
a simple LISTSXP to some homegrown object containing all the errors. The major
issue we're going to have here is that we need to make sure that the protection
stack works out since as we've heading right now the VALC_res needs to be
protected explicitly since it contains non-SEXP crap in it. Most likely we'll
need to change that to be a SEXP that contains the vector messages in component
pieces plus the ancillary stuff.
All this is to delay having to call pad_or_quote or smprintf business. Have
to wonder if we truly optimized those, maybe we could just skip worrying about
these? Seems like no matter what we'll have to create some CHARSXPS...
We'll need to figure out how to update VALC_process_error to deal with the new
format.
Looks like creating a 5 long character vector with new strings (i.e. requiring
new CHARSXPs) takes about 4 microseconds, which seems way too long. So maybe an
alternative solution is to pass around a results object that contains the result
data as C objects, and one LISTSXP with all the SEXP objects. This means that
every function that returns an ALIKEC_res_fin or VALC_res object needs to be
modified to accept the result object as an input.
Actually, not quite, we just need the VALC_res ones. We can just harvest and
copy over the ALIKEC_res_fin separately.
There are two types of SEXPs we need to handle:
- The call info from
alike that has all the names(x)[1] stuff
- The result of evaluating standard tokens
For the non-SEXP stuff, which is most likely to be the alike strings, do we want
to implement a pair list type structure, or a resizing array? Probably pair
list. For standard tokens seems like recording the value of the eval_c
variable might be sufficient.
ALIKE structs
ALIKEC_res_interim
Inputs
ALIKEC_res_strings_to_SEXP doesnt use object slotALIKEC_res_interim_as_strings doesn't use object slot
For the next two, while used, call could just as easily be passed separately
ALIKEC_strsxp_or_true deparses object into callALIKEC_string_or_true deparses object into call
Output
ALIKEC_alike_wrap generates the final call object that can then be deparsed,
so can be saved as a simple callALIKEC_type_alike_internal doesn't even use the objectALIKEC_fun_alike_internal doesn't even use the object
ALIKEC_res
Inputs
ALIKEC_wrap
Output
ALIKEC_alike_objALIKEC_alike_recALIKEC_alike_internal
NSE?
Is it possible to allow the functions to accept arbitrary language objects and
expressions?
For example, currently something like validate(INT.1, x) works fine because
INT.1 is a language object, but validate(INT.1 || LGL.1) fails because R
will try to evaluate the expression which obviously doesn't work.
Proposal:
- Substitute first
- Parse
- Evaluate language components
- Continue parsing?
Seems like this should work just fine so long as we substitute target first.
Need to watch out for infinite loops with stuff that keeps evaluating to
language? (e.g. y <- quote(get("y")); validate(y, x);)
^^ indeed. In order to protect against this, we will use a hash table to track
re-appearance of a symbol, the problem is we only care about reappearance of a
symbol if it is a child to itself. This means we have to rest the hash tables
after we dive into each branch of the parsed language, or duplicate them each
time we dive into a branch.
So we need a structure that points to:
- Our hash table
- A growable character vector containing all encountered names
- The current index of the character vector
- The current size of the character vector
We also want to extend this to environment tracking as that currently may not be
done correctly? Though with environments maybe less of an issue as there isn't
any and/or logic to contend with, so if an environment matched in one branch it
must be definition match in the next.
One potential challenge on relying on environment memory locations is if
somehow R copies and moves an environment in the middle of our operation. That
seems somewhat unlikely, but should probably be documented someplace.
Parse Rules
fun <- function(argument1, argument2, argument3, argument.opt=FALSE) {
validate(
argument1=integer(1L) && .(!is.na(.)),
argument2=scl.pos.int(5L),
argument3=matrix(integer(), ncol=3, dimnames=list(NULL, c("R", "G", "B"))) ||
(logical(2L) && .(all(is.na(.))),
argument.opt=logical(1L)
)
}
Rules to break down calls:
- if
(, dive in
- if
&& or || and length() == 2 then recurse
- if
.(
- mark for evaluation, but replace all
. with original variable name, and remove dot from anything else that is more than one dot
- else evaluate as
alike(arg, expr)
- but how do we treat something like
matrix(integer(), nrow=.(.))
- main issue is with rule to add dots in order to escape dots
So we need to recurse through all elements hunting for any . and .( or dot only expressions, and:
- if a single dot
- if a call, replace with identical
- if not, replace with arg name
- if more than one dot, trim a dot
UPDATE: now any expression that contains a dot is evaluted as is
Two pass process then
- Descend, capture items to
alike, vs items to plain evaluate
- Re-descend and at marked items, either plain evaluate or
alike
So what data structure allows us to mark during recursive descent? We really should study the standard libraries...
In the meantime, let's consider:
((a && b) || c || (d && e && f)) && g
Maybe the right thing to do is one round of recursive descent / replacement, and then a second round of evaluation under each format. This mostly works except that we still need to know which tests get evaluated normally vs as templates.
Alternate: do the same recursive descent, returning a C "linked list"? Or maybe easier just use an R linked list, and for each CAR use a two element "list" vector, containing:
- A pointer to the call element (the original CAR)
- How the element should be evaluated:
- plain evaluated
- as
alike
- not at all because it is a child element to one of the two above
C Notes
Tracebacks
Tracebacks not set in unitizer when eval(stop()) from C, even though they are set outside of unitizer, and when calling error in C within unitizer.
PROTECT stack and error
It seems like there are no stack imbalance problems when the script finishes with errors; is this okay?
Promise Evaluation
eval in C will cause a promise to be evaluated, even though findVar will keep returning a promise and PRSEEN will still return 0. We tested this by accessing a slow evaluating argument more than once.
Optimization
all_in
Rather than try to do a full on optimization right now, try a simple
implementation that is faster than the simple all(x %in% y):
source('tests/unitizer/_pre/lorem.R')
lorem.words <- unlist(strsplit(lorem, "\\W"))
lorem.unique <- unique(lorem.words)
x <- sample(lorem.unique, 1e4, rep=TRUE)
library(microbenchmark)
microbenchmark(times=10,
all(unique(x) %in% lorem.unique),
all(x %in% lorem.unique),
!anyNA(match(unique(x), lorem.unique))
)
## Unit: microseconds
## expr min lq mean median
## all(unique(x) %in% lorem.unique) 173.327 175.111 221.3290 216.5445
## all(x %in% lorem.unique) 223.698 225.259 228.5640 225.8155
## !anyNA(match(unique(x), lorem.unique)) 188.009 189.564 231.6546 221.5680
## uq max neval
## 264.897 277.004 10
## 232.412 240.201 10
## 258.627 329.226 10
Short of it is that it doesn't seem worth it.
Vs. checkmate
xx <- runif(1e4)
xx[1] <- 1
xx[2] <- 0
microbenchmark(
assertNumeric(xx, any.missing = FALSE, lower = 0, upper = 1),
vet(numeric() && all_bw(., 0, 1), xx),
assertNumeric(xx, any.missing = FALSE, lower = 0),
vet(numeric() && all_bw(., 0, Inf), xx),
vet(numeric() && all(. > 0), xx)
)
microbenchmark(
all_bw(xx, 0, 1),
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq
## assertNumeric(xx, any.missing = FALSE, lower = 0, upper = 1) 26.627 29.9290
## vet(numeric() && all_bw(., 0, 1), xx) 20.198 22.6825
## assertNumeric(xx, any.missing = FALSE, lower = 0) 19.055 20.7130
## vet(numeric() && all_bw(., 0, Inf), xx) 14.885 16.8890
## vet(numeric() && all(. > 0), xx) 73.076 119.4560
## mean median uq max neval
## 42.24239 36.187 55.2480 88.824 100
## 30.46707 25.055 40.4240 86.059 100
## 27.72199 23.029 36.6965 54.854 100
## 23.69168 18.597 30.5225 54.603 100
## 149.69694 133.635 169.2880 301.521 100
ss <- do.call(paste0, expand.grid(letters, letters, letters))
microbenchmark(
all_bw(ss, "a", "zzz"),
all(xx >= "a" & xx <= "zzz")
)
## Unit: microseconds
## expr min lq mean median
## all_bw(ss, "a", "zzz") 336.112 410.057 463.3804 458.985
## all(xx >= "a" & xx <= "zzz") 23731.092 24235.388 26356.3079 24951.518
## uq max neval
## 502.3695 904.064 100
## 27598.0535 34451.201 100
Not entirely sure what's going on here, shouldn't be that much faster.
Loop unswitching
Trusting that gcc does comparable stuff?
xx <- runif(1e4)
microbenchmark(
all_bw(xx, 0, 1),
all_bw2(xx, 0, 1),
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq mean median uq max
## all_bw(xx, 0, 1) 23.117 23.3210 24.65532 23.6515 24.1765 98.217
## all_bw2(xx, 0, 1) 13.665 13.8910 14.62394 14.2865 14.5635 37.207
## all(xx >= 0 & xx <= 1) 82.325 92.8285 139.04823 141.6830 149.6355 1164.560
Switching to 03
> microbenchmark(
+ all_bw(xx, 0, 1),
+ all_bw2(xx, 0, 1),
+ all(xx >= 0 & xx <= 1)
+ )
Unit: microseconds
expr min lq mean median uq max neval
all_bw(xx, 0, 1) 13.659 14.002 15.74092 14.434 14.9120 47.561 100
all_bw2(xx, 0, 1) 18.759 19.114 21.21446 19.823 20.5100 55.189 100
all(xx >= 0 & xx <= 1) 82.436 91.505 145.77960 116.293 146.3375 2128.326 100
Tried seeing if there were some ways to force more loop unswitching as compiler
really should be able to do it, but didn't find obvious setting which suggests
higher probability that different compilers may work differently as well.
Compare to integer:
xx <- runif(1e4)
yy <- sample(seq.int(1e4))
microbenchmark(
all_bw(xx, 0, 1),
all_bw(yy, 0, 1e4),
all_bw(yy, 0, 9e3), # fail
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq mean median uq max
## all_bw(xx, 0, 1) 12.609 13.2120 13.54127 13.3295 13.6290 23.172
## all_bw(yy, 0, 10000) 8.209 8.6555 8.94603 8.7505 9.1845 11.299
## all_bw(yy, 0, 9000) 4.067 4.3385 5.44750 4.6980 5.2350 55.264
## all(xx >= 0 & xx <= 1) 82.532 89.7550 127.70684 126.4315 148.0600 905.974
New Notes (7/14/17)
It seems overall .Call has gotten a bit faster, although perhaps this is the
new machine being faster:
> microbenchmark(test1(1), test2(1,2), test3(1,2,3))
Unit: nanoseconds
expr min lq mean median uq max neval
test1(1) 488 506.0 953.77 517.5 585 40254 100
test2(1, 2) 592 632.0 724.22 664.5 745 3332 100
test3(1, 2, 3) 699 748.5 853.18 809.0 881 2741 100
These all do trivial work, trying to measure overhead from parameters. Looks
like 150ns per parameter. Probably not enough to worry about too much.
.Call vs. .External:
> microbenchmark(valtest1(1, 2, 3), valtest2(1, 2, 3))
Unit: nanoseconds
expr min lq median uq max neval
valtest1(1, 2, 3) 856 1092.5 1353.5 1523.5 18416 100
valtest2(1, 2, 3) 4316 4601.5 4906.0 5249.5 106648 100
but this was using .External("funname", ...) as the variable version does not seem to work:
Error in .External(VALC_test2, ...) : NULL value passed as symbol address
Actually, if we change both character fun name:
> microbenchmark(valtest1(1, 2, 3), valtest2(1, 2, 3))
Unit: microseconds
expr min lq median uq max neval
valtest1(1, 2, 3) 4.655 4.7890 4.957 5.1860 57.944 100
valtest2(1, 2, 3) 4.296 4.5355 4.660 4.8635 19.669 100
so moral is we need to figure out how to get .External to work with the object?
Recursion vs Loop
Doesn't really seem to make a difference:
SEXP VALC_test1(SEXP a) {
if(TYPEOF(a) == VECSXP) {
return(VALC_test1(VECTOR_ELT(a, 0)));
}
return(a);
}
SEXP VALC_test2(SEXP a) {
while(TYPEOF(a) == VECSXP) {
a = VECTOR_ELT(a, 0);
}
return(a);
}
> microbenchmark(valtest1(a), valtest2(a))
Unit: microseconds
expr min lq median uq max neval
valtest1(a) 1.001 1.022 1.0595 1.1535 9.619 100
valtest2(a) 1.004 1.022 1.0800 1.1575 2.853 100
This is on an about 20-30 deep list
Complex Return Values
We're using SEXPs for complex return values, which often means we will create a SEXP from a C object, and then convert back to C on the other side. Don't know how much overhead stuff like ScalarInteger() has, but probably much better to just make structs and use those minimizing the amount of SEXPs used.
install
Costs about 40ns per, based on simple test.
match_call use from within C code
With two match_call and extracting fun:
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 6.591 7.1275 7.35 7.638 59.792 500
With one match_call and extracting the fun name:
> microbenchmark(fun(1, 2, 3), times=500)
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 4.72 5.0715 5.23 5.418 55.835 500
With just the one match.call:
> microbenchmark(fun(1, 2, 3), times=500)
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 4.722 5.027 5.176 5.3605 38.562 500
So about 1.9us for using match.call to do the matching. If we tried to match
on tag name only could be faster; something to think about.
Initial benchmark
Using version 0.0.1.9000 and the following functions:
fun1 <- function(x, y, z)
validate_args(
x=matrix(integer(), ncol=3) || integer(3L),
y=integer(2L) || NULL || logical(1L),
z=logical(1L)
)
fun1b <- function(x, y, z) {
stopifnot(
(
is.matrix(x) && ncol(x) == 3 && (
is.integer(x) || (is.numeric(x) && all(floor(x) == x))
)
) ||
((is.integer(x) || (is.numeric(x) && all(floor(x) == x))) && length(x) == 3)
)
stopifnot(
((is.integer(y) || (is.numeric(y) && all(floor(y) == y))) && length(y) == 2) ||
is.null(y) || (is.logical(y) && length(y) == 1L)
)
stopifnot(is.logical(z) && length(z) == 1L)
}
we get comparable performance:
> microbenchmark(fun1(1:3, FALSE, FALSE), fun1b(1:3, FALSE, FALSE))
Unit: microseconds
expr min lq median uq max neval
fun1(1:3, FALSE, FALSE) 23.165 24.257 25.050 26.0010 60.102 100
fun1b(1:3, FALSE, FALSE) 19.199 20.684 22.361 23.4795 38.313 100
vs alternatives
library(validate)
library(ensurer)
mean2 <- function(x, ...) {validate_args(numeric()); mean(x, ...)}
mean3 <- function(x, ...) {ensure_that(x, is.numeric(.)); mean(x, ...)}r
mean4 <- function(x, ...) {stopifnot(is.numeric(x)); mean(x, ...)}
mean2a <- function(x, ...) {validate_args(INT.1); mean(x, ...)}
mean4a <- function(x, ...) {
stopifnot(
is.integer(x) || (is.numeric(x) && all(x == round(x))),
length(x) == 1L, !is.na(x));
mean(x, ...)
}
library(microbenchmark)
microbenchmark(
mean(1:1e4), mean2(1:1e4), mean3(1:1e4), mean4(1:1e4), mean2a(1L),
mean4a(1L)
)
# Unit: microseconds
# expr min lq mean median uq max neval
# mean(1:10000) 18.081 19.0440 21.45990 20.3400 21.2840 48.570 100
# mean2(1:10000) 31.804 34.9475 40.90450 36.6980 39.2780 121.493 100
# mean3(1:10000) 129.214 134.4045 172.52273 139.9375 158.0205 938.026 100
# mean4(1:10000) 26.739 29.2570 35.75861 30.4975 32.7875 158.021 100`
LISTSXP vs VECSXP
Linked lists seem to be slightly faster, at least for small lists:
SEXP VALC_test(SEXP a, SEXP b) {
SEXP c;
// c = PROTECT(allocVector(VECSXP, 3));
// SET_VECTOR_ELT(c, 0, R_NilValue);
// SET_VECTOR_ELT(c, 1, R_NilValue);
// SET_VECTOR_ELT(c, 2, R_NilValue);
// VECTOR_ELT(c, 0);
// VECTOR_ELT(c, 1);
// VECTOR_ELT(c, 2);
// UNPROTECT(1);
c = PROTECT(list3(R_NilValue, R_NilValue, R_NilValue));
CAR(c);
CADR(c);
CADDR(c);
UNPROTECT(1);
return(R_NilValue);
}
VECSXP
> microbenchmark(valtest(1, 2))
Unit: nanoseconds
expr min lq median uq max neval
valtest(1, 2) 703 734.5 858 1068 35383 100
```
LISTSXP
Unit: nanoseconds
expr min lq median uq max neval
valtest(1, 2) 681 728 837 1024.5 10264 100
## Usability
### Providing Access to Templates
It would be super helpful to actually be able to see the template, and possibly
even get a diff against it. E.g., in:
mdlc <- lm(Sepal.Width ~ Petal.Length + Sepal.Length, iris)
mdlc.a <- abstract(mdlc)
mdld <- lm(Sepal.Width ~ Petal.Length -1, iris)
validate(mdlc.a, mdld)
Error in validate(mdlc.a, mdlc) :
(mdld$terms)[[3]][[1]] should be a call to + (is a call to -)
(mdld$terms)[[3]]
Petal.Length - 1
It would be great to have easy access to `mdlc.a$terms[[3]]`
## OR error messages
One of the more challenging ones are OR error messages, where our current naive
display of the multiple possibilities is repetitive an unclear. Here is an
improvement:
Argument x does not meet any of the allowable forms:
- names(letters[1:3]) is "NULL", but should be "character"
- letters[1:3] is "character", but should be "matrix" or "list"
```
This one is pretty good, but boy is it going to be a pain to figure out
all the elements that could be collapsed into it... Or is it? can we
just sort and then remove the common pieces, and append with ", or"?
More importantly, is it worth the trouble for a fairly rare use case?
Need to make sure we match up to the is, otherwise we'll get
nonsensical errors. Blergh
Additional complication: what about multi-line expressions? Or
expressions that are not template based (e.g. "is not TRUE (FALSE)")
Part of the problem is that we have errors that are generated by alike, and
errors that are generated by validate for the non-alike methods, and we're
trying to use string output from those these processes in a unified logical way.
It may make more sense to split up the errors into lists with the following
components:
- A deparsed expression of the problem element, as a
character(1L)
- A
character(1L) of what it should be, without the 'should be', or more
specifically without the should, as we need to contend with 'should be' and
'should contain', etc.
- A
character(1L) of what it is
Then when we get a list of these, we identify which elements are equal in the
first and last elements, and collapse the "should bes" into one comma/or
delimited string.
Implementation may be more straightforward than anticipated. The only thing we
seem to need to do for alike is to break up the res.message business into
res.target and res.actual. Some special treatment needed for the language
results since those have the annoying chr->SEXP->chr translation business going
on via ALIKEC_lang_alike_internal() (actually, that's probably not going away
any time soon as there are additional complexities therein).
Random output
all_bw(zzz, 0, 1) is not TRUE (is "character" instead of a "logical")
all_bw(zzz, 0, 1) is not TRUE (string: "contains values outside of only values in range [0,1] (2.783366 at index 3)")
Error in fun2a(x = letters[1:3]) :
For argument `x` at least one of these should pass:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
Error in fun2a(x = letters[1:3]) :
Argument `x` does not validate; at least one of the following should be
true:
In order to pass validation, argument `x` should meet one or more of the
following requirements, but it meets none:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
Argument `x` fails validation because it should meet one or more of the
following requirements, but it meets none:
Argument `x` is invalid as no allowable conditions are met:
Argument `x` does not meet any of the allowable forms:
Argument `x` does not meet any of the allowable validators:
Argument `x` does not meet any of the permitted validators:
Argument `x` fails all permitted validators:
Argument `x` fails every permitted vetting expression:
- `names(letters[1:3])` could be type "character" (is "NULL")
- `letters[1:3]` could be matrix (is character)
- `letters[1:3]` could be type "list" (is "character")
Argument `x` fails because none of the following are true:
For argument `x` at least one of these should pass:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
# Problem with this one is it doesn't tell you what it should be:
Error in fun2a(x = letters[1:3]) :
For argument `x` at least one of the following should be true but none are:
- `typeof(names(letters[1:3])) == "character"`
- `class(letters[1:3]) == "matrix"`
- `typeof(letters[1:3]) == "list"`
#
Error in fun2a(x = letters[1:3]) :
Argument `x` must meet at least one of the following to validate:
- `typeof(names(letters[1:3]))` is NULL (expected "character")
- `class(letters[1:3])` is character (expected "matrix")
- `typeof(letters[1:3])` is character (expected "list")
# Below is best, but doesn't work when we're subsetting / using accesor funs
Error in fun2a(x = letters[1:3]) :
Argument `x` validation failure: `letters[1:3]` is character, but
should be matrix, or type list
Validation error in fun2a(x = letters[1:3]) :
Argument `x`: `letters[1:3]` is character, but should be matrix or
list
Argument `x`:
letters[1:3]
is character, but should be matrix or type list
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must meet at least one of the following, but does not:
- Expected length 1, but got 3
- Expected type "NULL", but got "integer"
Error in validate(INT.1 || NULL, 1:3) :
At least one of the following should be true:
- `length(1:3)` should be 1, but is 3
- `1:3` should be type NULL, but is integer
Error in fun2(x = 1:3, y = TRUE) :
Argument `x` fails validation because none of the following are true:
- `length(1:3)` should be 1, but is 3
- `1:3` should be type NULL, but is integer
Error in fun2(x = 1:3, y = TRUE) :
Input error in fun2(x = 1:3, y = TRUE) :
For argument `x` at least one of the following must be TRUE:
For argument `x` at least one of these must be TRUE:
At least one of the following should be TRUE for argument `x`, but none
are:
Bad argument `x`; does not pass any of the following:
- `length(1:3)` should be 1, not 3
- `1:3` should be type NULL, not integer
- `all(-1:1 > 0)` should evaluate to TRUE (is FALSE)
- Expression:
all(-1:1 > 0)
should evaluate to TRUE (is FALSE)
Error in fun2(x = matrix(c(1:9), nrow = 3), y = -1:1, z = NULL) :
Argument have `all(-1:1 > 0)` evaluate to TRUE (is FALSE)
For argument `y`, `all(-1:1 > 0)` should evaluate to TRUE (is FALSE)
Error in fun2(x = 1:3, y = TRUE) :
Argument y wrong because TRUE should be type integer, not logical
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must:
- have length 1 (is 3), OR
- have type "NULL" (is "integer"), OR
- have class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must have:
- length 1 (is 3), OR
- type "NULL" (is "integer"), OR
- class "matrix" (is "integer"), OR
- `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inheritance from class "foo"
Argument `current` should:
- be length 1 (is 3), OR
- be type "NULL" (is "integer"), OR
- be class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
Argument `current` should:
- be length 1 (is 3), OR
- be type "NULL" (is "integer"), OR
- be class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
brodieG/vetr documentation built on Jan. 31, 2023, 5:56 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
gcc-10
See SO issue.
Issues related to compiling under gcc 10.0.1 causing segfaults on BDR machines. To reproduce (though this will no longer work when debian:testing moves away from from gcc-10.0.1):
docker pull rocker/r-base
docker run --rm -ti --security-opt seccomp=unconfined \
rocker/r-base /bin/bash
apt-get update
apt-get install gcc-10 gdb
gcc-10 --version # confirm 10.0.1
# gcc-10 (Debian 10-20200222-1) 10.0.1 20200222 (experimental)
# [master revision 01af7e0a0c2:487fe13f218:e99b18cf7101f205bfdd9f0f29ed51caaec52779]
mkdir ~/.R
touch ~/.R/Makevars
echo "CC = gcc-10
CFLAGS = -g -O2 -Wall -pedantic -mtune=native -Werror=format-security -Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector-strong -fstack-clash-protection -fcf-protection
" >> ~/.R/Makevars
R -d gdb --vanilla
Then in the R console, after typing run to get gdb to run the program:
f.dl <- tempfile()
f.uz <- tempfile()
github.url <- 'https://github.com/brodieG/vetr/archive/v0.2.8.zip'
download.file(github.url, f.dl)
unzip(f.dl, exdir=f.uz)
install.packages(file.path(f.uz, 'vetr-0.2.8'), repos=NULL, type='source')
unlink(c(f.dl, f.uz))
# minimal set of commands I could come up that would cause segfault follow
library(vetr)
alike(pairlist(a=1, b="character"), pairlist(a=1, b=letters))
alike(pairlist(1, "character"), pairlist(1, letters))
alike(NULL, 1:3) # not a wild card at top level
alike(list(NULL), list(1:3)) # but yes when nested
alike(list(NULL, NULL), list(list(list(1, 2, 3)), 1:25))
alike(list(NULL), list(1, 2))
alike(list(), list(1, 2))
alike(matrix(integer(), ncol=7), matrix(1:21, nrow=3))
alike(matrix(character(), nrow=3), matrix(1:21, nrow=3))
alike(
matrix(integer(), ncol=3, dimnames=list(NULL, c("R", "G", "B"))),
matrix(1:21, ncol=3, dimnames=list(NULL, c("R", "G", "B")))
)
# Adding tests from docs
mx.tpl <- matrix(
integer(), ncol=3, dimnames=list(row.id=NULL, c("R", "G", "B"))
)
mx.cur <- matrix(
sample(0:255, 12), ncol=3, dimnames=list(row.id=1:4, rgb=c("R", "G", "B"))
)
mx.cur2 <-
matrix(sample(0:255, 12), ncol=3, dimnames=list(1:4, c("R", "G", "B")))
alike(mx.tpl, mx.cur2)
Inspecting in gdb pretty quickly shows (if I understand correctly) that
CSR_strmlen_x is trying to access the string that was not initialized.
Notes from 8/29
Dig into the syntactic names issue:
unitizer> vetr:::syntactic_names(quote(mtcars[1:10, ]))
[1] FALSE
Notes from 8/22
Most of the alike side of things is dealt with, now we need to figure out how we
want to store stuff in lists. Used to be that we would just use LISTSXPs, but
now we need a growing collection of VALC_res objects? It should contain:
- An array of
ALIKEC_resobjects - The last inserted index
- The current allocated size
- By extension, a value in settings that has the default initial allocated size
We will need a mechanism to grow the array each time we hit the size limit.
Need to remember that we will have protected as many times as there are elements in that array.
Also, need to figure out what functions need the added argument so we can pass around our array pointer.
Notes from 8/21
So main thing to do is cleanly separate the "message" SEXP into the message
character values as part of the ALIKEC_res_interim object or some such, and
just the wrap object.
So we need to restructure the ALIKEC_res and ALIKEC_res_sub objects to
contain all the string structure. Additionally, it seems like the interim
object really doesn't need the call object? I guess it can be there explicitly
as a call object to distinguish it from the wrap object? The only real use
case is that after we've wrapped the call and are ready to return it to vetr
we need to put it someplace.
So we need objects that:
- Track the strings
- Track the strings + recursion index + wrap lang
- Track the strings + wrap lang + attr type
- Track the strings + final call
Also need to track success status, df status, attr level
ALIKEC_res_strings: just the stringsALIKEC_res_min: strings + successALIKEC_res_wrap: strings + success + wrap + [rec index] + [df] + [attr]ALIKEC_res: strings + success + call
One thing left to do: figure out what ALIKEC_wrap should be returning.
Notes from 8/20
Currently working through trying to delay the construction of error messages to the very last minute.
Seems like the main issue is that that in VALC_evaluate, we need to change from a simple LISTSXP to some homegrown object containing all the errors. The major issue we're going to have here is that we need to make sure that the protection stack works out since as we've heading right now the VALC_res needs to be protected explicitly since it contains non-SEXP crap in it. Most likely we'll need to change that to be a SEXP that contains the vector messages in component pieces plus the ancillary stuff.
All this is to delay having to call pad_or_quote or smprintf business. Have
to wonder if we truly optimized those, maybe we could just skip worrying about
these? Seems like no matter what we'll have to create some CHARSXPS...
We'll need to figure out how to update VALC_process_error to deal with the new
format.
Looks like creating a 5 long character vector with new strings (i.e. requiring
new CHARSXPs) takes about 4 microseconds, which seems way too long. So maybe an
alternative solution is to pass around a results object that contains the result
data as C objects, and one LISTSXP with all the SEXP objects. This means that
every function that returns an ALIKEC_res_fin or VALC_res object needs to be
modified to accept the result object as an input.
Actually, not quite, we just need the VALC_res ones. We can just harvest and
copy over the ALIKEC_res_fin separately.
There are two types of SEXPs we need to handle:
- The call info from
alikethat has all thenames(x)[1]stuff - The result of evaluating standard tokens
For the non-SEXP stuff, which is most likely to be the alike strings, do we want
to implement a pair list type structure, or a resizing array? Probably pair
list. For standard tokens seems like recording the value of the eval_c
variable might be sufficient.
ALIKE structs
ALIKEC_res_interim
Inputs
ALIKEC_res_strings_to_SEXPdoesnt use object slotALIKEC_res_interim_as_stringsdoesn't use object slot
For the next two, while used, call could just as easily be passed separately
ALIKEC_strsxp_or_truedeparses object into callALIKEC_string_or_truedeparses object into call
Output
ALIKEC_alike_wrapgenerates the final call object that can then be deparsed, so can be saved as a simple callALIKEC_type_alike_internaldoesn't even use the objectALIKEC_fun_alike_internaldoesn't even use the object
ALIKEC_res
Inputs
ALIKEC_wrap
Output
ALIKEC_alike_objALIKEC_alike_recALIKEC_alike_internal
NSE?
Is it possible to allow the functions to accept arbitrary language objects and expressions?
For example, currently something like validate(INT.1, x) works fine because
INT.1 is a language object, but validate(INT.1 || LGL.1) fails because R
will try to evaluate the expression which obviously doesn't work.
Proposal:
- Substitute first
- Parse
- Evaluate language components
- Continue parsing?
Seems like this should work just fine so long as we substitute target first.
Need to watch out for infinite loops with stuff that keeps evaluating to
language? (e.g. y <- quote(get("y")); validate(y, x);)
^^ indeed. In order to protect against this, we will use a hash table to track re-appearance of a symbol, the problem is we only care about reappearance of a symbol if it is a child to itself. This means we have to rest the hash tables after we dive into each branch of the parsed language, or duplicate them each time we dive into a branch.
So we need a structure that points to:
- Our hash table
- A growable character vector containing all encountered names
- The current index of the character vector
- The current size of the character vector
We also want to extend this to environment tracking as that currently may not be done correctly? Though with environments maybe less of an issue as there isn't any and/or logic to contend with, so if an environment matched in one branch it must be definition match in the next.
One potential challenge on relying on environment memory locations is if somehow R copies and moves an environment in the middle of our operation. That seems somewhat unlikely, but should probably be documented someplace.
Parse Rules
fun <- function(argument1, argument2, argument3, argument.opt=FALSE) {
validate(
argument1=integer(1L) && .(!is.na(.)),
argument2=scl.pos.int(5L),
argument3=matrix(integer(), ncol=3, dimnames=list(NULL, c("R", "G", "B"))) ||
(logical(2L) && .(all(is.na(.))),
argument.opt=logical(1L)
)
}
Rules to break down calls:
- if
(, dive in - if
&&or||andlength() == 2then recurse - if
.(- mark for evaluation, but replace all
.with original variable name, and remove dot from anything else that is more than one dot
- mark for evaluation, but replace all
- else evaluate as
alike(arg, expr)- but how do we treat something like
matrix(integer(), nrow=.(.)) - main issue is with rule to add dots in order to escape dots
- but how do we treat something like
So we need to recurse through all elements hunting for any . and .( or dot only expressions, and:
- if a single dot
- if a call, replace with identical
- if not, replace with arg name
- if more than one dot, trim a dot
UPDATE: now any expression that contains a dot is evaluted as is
Two pass process then
- Descend, capture items to
alike, vs items to plain evaluate - Re-descend and at marked items, either plain evaluate or
alike
So what data structure allows us to mark during recursive descent? We really should study the standard libraries...
In the meantime, let's consider:
((a && b) || c || (d && e && f)) && g
Maybe the right thing to do is one round of recursive descent / replacement, and then a second round of evaluation under each format. This mostly works except that we still need to know which tests get evaluated normally vs as templates.
Alternate: do the same recursive descent, returning a C "linked list"? Or maybe easier just use an R linked list, and for each CAR use a two element "list" vector, containing:
- A pointer to the call element (the original CAR)
- How the element should be evaluated:
- plain evaluated
- as
alike - not at all because it is a child element to one of the two above
C Notes
Tracebacks
Tracebacks not set in unitizer when eval(stop()) from C, even though they are set outside of unitizer, and when calling error in C within unitizer.
PROTECT stack and error
It seems like there are no stack imbalance problems when the script finishes with errors; is this okay?
Promise Evaluation
eval in C will cause a promise to be evaluated, even though findVar will keep returning a promise and PRSEEN will still return 0. We tested this by accessing a slow evaluating argument more than once.
Optimization
all_in
Rather than try to do a full on optimization right now, try a simple
implementation that is faster than the simple all(x %in% y):
source('tests/unitizer/_pre/lorem.R')
lorem.words <- unlist(strsplit(lorem, "\\W"))
lorem.unique <- unique(lorem.words)
x <- sample(lorem.unique, 1e4, rep=TRUE)
library(microbenchmark)
microbenchmark(times=10,
all(unique(x) %in% lorem.unique),
all(x %in% lorem.unique),
!anyNA(match(unique(x), lorem.unique))
)
## Unit: microseconds
## expr min lq mean median
## all(unique(x) %in% lorem.unique) 173.327 175.111 221.3290 216.5445
## all(x %in% lorem.unique) 223.698 225.259 228.5640 225.8155
## !anyNA(match(unique(x), lorem.unique)) 188.009 189.564 231.6546 221.5680
## uq max neval
## 264.897 277.004 10
## 232.412 240.201 10
## 258.627 329.226 10
Short of it is that it doesn't seem worth it.
Vs. checkmate
xx <- runif(1e4)
xx[1] <- 1
xx[2] <- 0
microbenchmark(
assertNumeric(xx, any.missing = FALSE, lower = 0, upper = 1),
vet(numeric() && all_bw(., 0, 1), xx),
assertNumeric(xx, any.missing = FALSE, lower = 0),
vet(numeric() && all_bw(., 0, Inf), xx),
vet(numeric() && all(. > 0), xx)
)
microbenchmark(
all_bw(xx, 0, 1),
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq
## assertNumeric(xx, any.missing = FALSE, lower = 0, upper = 1) 26.627 29.9290
## vet(numeric() && all_bw(., 0, 1), xx) 20.198 22.6825
## assertNumeric(xx, any.missing = FALSE, lower = 0) 19.055 20.7130
## vet(numeric() && all_bw(., 0, Inf), xx) 14.885 16.8890
## vet(numeric() && all(. > 0), xx) 73.076 119.4560
## mean median uq max neval
## 42.24239 36.187 55.2480 88.824 100
## 30.46707 25.055 40.4240 86.059 100
## 27.72199 23.029 36.6965 54.854 100
## 23.69168 18.597 30.5225 54.603 100
## 149.69694 133.635 169.2880 301.521 100
ss <- do.call(paste0, expand.grid(letters, letters, letters))
microbenchmark(
all_bw(ss, "a", "zzz"),
all(xx >= "a" & xx <= "zzz")
)
## Unit: microseconds
## expr min lq mean median
## all_bw(ss, "a", "zzz") 336.112 410.057 463.3804 458.985
## all(xx >= "a" & xx <= "zzz") 23731.092 24235.388 26356.3079 24951.518
## uq max neval
## 502.3695 904.064 100
## 27598.0535 34451.201 100
Not entirely sure what's going on here, shouldn't be that much faster.
Loop unswitching
Trusting that gcc does comparable stuff?
xx <- runif(1e4)
microbenchmark(
all_bw(xx, 0, 1),
all_bw2(xx, 0, 1),
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq mean median uq max
## all_bw(xx, 0, 1) 23.117 23.3210 24.65532 23.6515 24.1765 98.217
## all_bw2(xx, 0, 1) 13.665 13.8910 14.62394 14.2865 14.5635 37.207
## all(xx >= 0 & xx <= 1) 82.325 92.8285 139.04823 141.6830 149.6355 1164.560
Switching to 03
> microbenchmark(
+ all_bw(xx, 0, 1),
+ all_bw2(xx, 0, 1),
+ all(xx >= 0 & xx <= 1)
+ )
Unit: microseconds
expr min lq mean median uq max neval
all_bw(xx, 0, 1) 13.659 14.002 15.74092 14.434 14.9120 47.561 100
all_bw2(xx, 0, 1) 18.759 19.114 21.21446 19.823 20.5100 55.189 100
all(xx >= 0 & xx <= 1) 82.436 91.505 145.77960 116.293 146.3375 2128.326 100
Tried seeing if there were some ways to force more loop unswitching as compiler really should be able to do it, but didn't find obvious setting which suggests higher probability that different compilers may work differently as well.
Compare to integer:
xx <- runif(1e4)
yy <- sample(seq.int(1e4))
microbenchmark(
all_bw(xx, 0, 1),
all_bw(yy, 0, 1e4),
all_bw(yy, 0, 9e3), # fail
all(xx >= 0 & xx <= 1)
)
## Unit: microseconds
## expr min lq mean median uq max
## all_bw(xx, 0, 1) 12.609 13.2120 13.54127 13.3295 13.6290 23.172
## all_bw(yy, 0, 10000) 8.209 8.6555 8.94603 8.7505 9.1845 11.299
## all_bw(yy, 0, 9000) 4.067 4.3385 5.44750 4.6980 5.2350 55.264
## all(xx >= 0 & xx <= 1) 82.532 89.7550 127.70684 126.4315 148.0600 905.974
New Notes (7/14/17)
It seems overall .Call has gotten a bit faster, although perhaps this is the
new machine being faster:
> microbenchmark(test1(1), test2(1,2), test3(1,2,3))
Unit: nanoseconds
expr min lq mean median uq max neval
test1(1) 488 506.0 953.77 517.5 585 40254 100
test2(1, 2) 592 632.0 724.22 664.5 745 3332 100
test3(1, 2, 3) 699 748.5 853.18 809.0 881 2741 100
These all do trivial work, trying to measure overhead from parameters. Looks like 150ns per parameter. Probably not enough to worry about too much.
.Call vs. .External:
> microbenchmark(valtest1(1, 2, 3), valtest2(1, 2, 3))
Unit: nanoseconds
expr min lq median uq max neval
valtest1(1, 2, 3) 856 1092.5 1353.5 1523.5 18416 100
valtest2(1, 2, 3) 4316 4601.5 4906.0 5249.5 106648 100
but this was using .External("funname", ...) as the variable version does not seem to work:
Error in .External(VALC_test2, ...) : NULL value passed as symbol address
Actually, if we change both character fun name:
> microbenchmark(valtest1(1, 2, 3), valtest2(1, 2, 3))
Unit: microseconds
expr min lq median uq max neval
valtest1(1, 2, 3) 4.655 4.7890 4.957 5.1860 57.944 100
valtest2(1, 2, 3) 4.296 4.5355 4.660 4.8635 19.669 100
so moral is we need to figure out how to get .External to work with the object?
Recursion vs Loop
Doesn't really seem to make a difference:
SEXP VALC_test1(SEXP a) {
if(TYPEOF(a) == VECSXP) {
return(VALC_test1(VECTOR_ELT(a, 0)));
}
return(a);
}
SEXP VALC_test2(SEXP a) {
while(TYPEOF(a) == VECSXP) {
a = VECTOR_ELT(a, 0);
}
return(a);
}
> microbenchmark(valtest1(a), valtest2(a))
Unit: microseconds
expr min lq median uq max neval
valtest1(a) 1.001 1.022 1.0595 1.1535 9.619 100
valtest2(a) 1.004 1.022 1.0800 1.1575 2.853 100
This is on an about 20-30 deep list
Complex Return Values
We're using SEXPs for complex return values, which often means we will create a SEXP from a C object, and then convert back to C on the other side. Don't know how much overhead stuff like ScalarInteger() has, but probably much better to just make structs and use those minimizing the amount of SEXPs used.
install
Costs about 40ns per, based on simple test.
match_call use from within C code
With two match_call and extracting fun:
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 6.591 7.1275 7.35 7.638 59.792 500
With one match_call and extracting the fun name:
> microbenchmark(fun(1, 2, 3), times=500)
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 4.72 5.0715 5.23 5.418 55.835 500
With just the one match.call:
> microbenchmark(fun(1, 2, 3), times=500)
Unit: microseconds
expr min lq median uq max neval
fun(1, 2, 3) 4.722 5.027 5.176 5.3605 38.562 500
So about 1.9us for using match.call to do the matching. If we tried to match
on tag name only could be faster; something to think about.
Initial benchmark
Using version 0.0.1.9000 and the following functions:
fun1 <- function(x, y, z)
validate_args(
x=matrix(integer(), ncol=3) || integer(3L),
y=integer(2L) || NULL || logical(1L),
z=logical(1L)
)
fun1b <- function(x, y, z) {
stopifnot(
(
is.matrix(x) && ncol(x) == 3 && (
is.integer(x) || (is.numeric(x) && all(floor(x) == x))
)
) ||
((is.integer(x) || (is.numeric(x) && all(floor(x) == x))) && length(x) == 3)
)
stopifnot(
((is.integer(y) || (is.numeric(y) && all(floor(y) == y))) && length(y) == 2) ||
is.null(y) || (is.logical(y) && length(y) == 1L)
)
stopifnot(is.logical(z) && length(z) == 1L)
}
we get comparable performance:
> microbenchmark(fun1(1:3, FALSE, FALSE), fun1b(1:3, FALSE, FALSE))
Unit: microseconds
expr min lq median uq max neval
fun1(1:3, FALSE, FALSE) 23.165 24.257 25.050 26.0010 60.102 100
fun1b(1:3, FALSE, FALSE) 19.199 20.684 22.361 23.4795 38.313 100
vs alternatives
library(validate)
library(ensurer)
mean2 <- function(x, ...) {validate_args(numeric()); mean(x, ...)}
mean3 <- function(x, ...) {ensure_that(x, is.numeric(.)); mean(x, ...)}r
mean4 <- function(x, ...) {stopifnot(is.numeric(x)); mean(x, ...)}
mean2a <- function(x, ...) {validate_args(INT.1); mean(x, ...)}
mean4a <- function(x, ...) {
stopifnot(
is.integer(x) || (is.numeric(x) && all(x == round(x))),
length(x) == 1L, !is.na(x));
mean(x, ...)
}
library(microbenchmark)
microbenchmark(
mean(1:1e4), mean2(1:1e4), mean3(1:1e4), mean4(1:1e4), mean2a(1L),
mean4a(1L)
)
# Unit: microseconds
# expr min lq mean median uq max neval
# mean(1:10000) 18.081 19.0440 21.45990 20.3400 21.2840 48.570 100
# mean2(1:10000) 31.804 34.9475 40.90450 36.6980 39.2780 121.493 100
# mean3(1:10000) 129.214 134.4045 172.52273 139.9375 158.0205 938.026 100
# mean4(1:10000) 26.739 29.2570 35.75861 30.4975 32.7875 158.021 100`
LISTSXP vs VECSXP
Linked lists seem to be slightly faster, at least for small lists:
SEXP VALC_test(SEXP a, SEXP b) {
SEXP c;
// c = PROTECT(allocVector(VECSXP, 3));
// SET_VECTOR_ELT(c, 0, R_NilValue);
// SET_VECTOR_ELT(c, 1, R_NilValue);
// SET_VECTOR_ELT(c, 2, R_NilValue);
// VECTOR_ELT(c, 0);
// VECTOR_ELT(c, 1);
// VECTOR_ELT(c, 2);
// UNPROTECT(1);
c = PROTECT(list3(R_NilValue, R_NilValue, R_NilValue));
CAR(c);
CADR(c);
CADDR(c);
UNPROTECT(1);
return(R_NilValue);
}
VECSXP
> microbenchmark(valtest(1, 2))
Unit: nanoseconds
expr min lq median uq max neval
valtest(1, 2) 703 734.5 858 1068 35383 100
```
LISTSXP
Unit: nanoseconds expr min lq median uq max neval valtest(1, 2) 681 728 837 1024.5 10264 100
## Usability
### Providing Access to Templates
It would be super helpful to actually be able to see the template, and possibly
even get a diff against it. E.g., in:
mdlc <- lm(Sepal.Width ~ Petal.Length + Sepal.Length, iris) mdlc.a <- abstract(mdlc) mdld <- lm(Sepal.Width ~ Petal.Length -1, iris) validate(mdlc.a, mdld)
Error in validate(mdlc.a, mdlc) :
(mdld$terms)[[3]][[1]] should be a call to + (is a call to -)
(mdld$terms)[[3]]
Petal.Length - 1
It would be great to have easy access to `mdlc.a$terms[[3]]`
## OR error messages
One of the more challenging ones are OR error messages, where our current naive
display of the multiple possibilities is repetitive an unclear. Here is an
improvement:
Argument x does not meet any of the allowable forms:
- names(letters[1:3]) is "NULL", but should be "character"
- letters[1:3] is "character", but should be "matrix" or "list"
```
This one is pretty good, but boy is it going to be a pain to figure out
all the elements that could be collapsed into it... Or is it? can we
just sort and then remove the common pieces, and append with ", or"?
More importantly, is it worth the trouble for a fairly rare use case?
Need to make sure we match up to the is, otherwise we'll get
nonsensical errors. Blergh
Additional complication: what about multi-line expressions? Or expressions that are not template based (e.g. "is not TRUE (FALSE)")
Part of the problem is that we have errors that are generated by alike, and
errors that are generated by validate for the non-alike methods, and we're
trying to use string output from those these processes in a unified logical way.
It may make more sense to split up the errors into lists with the following
components:
- A deparsed expression of the problem element, as a
character(1L) - A
character(1L)of what it should be, without the 'should be', or more specifically without theshould, as we need to contend with 'should be' and 'should contain', etc. - A
character(1L)of what it is
Then when we get a list of these, we identify which elements are equal in the first and last elements, and collapse the "should bes" into one comma/or delimited string.
Implementation may be more straightforward than anticipated. The only thing we
seem to need to do for alike is to break up the res.message business into
res.target and res.actual. Some special treatment needed for the language
results since those have the annoying chr->SEXP->chr translation business going
on via ALIKEC_lang_alike_internal() (actually, that's probably not going away
any time soon as there are additional complexities therein).
Random output
all_bw(zzz, 0, 1) is not TRUE (is "character" instead of a "logical")
all_bw(zzz, 0, 1) is not TRUE (string: "contains values outside of only values in range [0,1] (2.783366 at index 3)")
Error in fun2a(x = letters[1:3]) :
For argument `x` at least one of these should pass:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
Error in fun2a(x = letters[1:3]) :
Argument `x` does not validate; at least one of the following should be
true:
In order to pass validation, argument `x` should meet one or more of the
following requirements, but it meets none:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
Argument `x` fails validation because it should meet one or more of the
following requirements, but it meets none:
Argument `x` is invalid as no allowable conditions are met:
Argument `x` does not meet any of the allowable forms:
Argument `x` does not meet any of the allowable validators:
Argument `x` does not meet any of the permitted validators:
Argument `x` fails all permitted validators:
Argument `x` fails every permitted vetting expression:
- `names(letters[1:3])` could be type "character" (is "NULL")
- `letters[1:3]` could be matrix (is character)
- `letters[1:3]` could be type "list" (is "character")
Argument `x` fails because none of the following are true:
For argument `x` at least one of these should pass:
- `names(letters[1:3])` should be type "character" (is "NULL")
- `letters[1:3]` should be matrix (is character)
- `letters[1:3]` should be type "list" (is "character")
# Problem with this one is it doesn't tell you what it should be:
Error in fun2a(x = letters[1:3]) :
For argument `x` at least one of the following should be true but none are:
- `typeof(names(letters[1:3])) == "character"`
- `class(letters[1:3]) == "matrix"`
- `typeof(letters[1:3]) == "list"`
#
Error in fun2a(x = letters[1:3]) :
Argument `x` must meet at least one of the following to validate:
- `typeof(names(letters[1:3]))` is NULL (expected "character")
- `class(letters[1:3])` is character (expected "matrix")
- `typeof(letters[1:3])` is character (expected "list")
# Below is best, but doesn't work when we're subsetting / using accesor funs
Error in fun2a(x = letters[1:3]) :
Argument `x` validation failure: `letters[1:3]` is character, but
should be matrix, or type list
Validation error in fun2a(x = letters[1:3]) :
Argument `x`: `letters[1:3]` is character, but should be matrix or
list
Argument `x`:
letters[1:3]
is character, but should be matrix or type list
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must meet at least one of the following, but does not:
- Expected length 1, but got 3
- Expected type "NULL", but got "integer"
Error in validate(INT.1 || NULL, 1:3) :
At least one of the following should be true:
- `length(1:3)` should be 1, but is 3
- `1:3` should be type NULL, but is integer
Error in fun2(x = 1:3, y = TRUE) :
Argument `x` fails validation because none of the following are true:
- `length(1:3)` should be 1, but is 3
- `1:3` should be type NULL, but is integer
Error in fun2(x = 1:3, y = TRUE) :
Input error in fun2(x = 1:3, y = TRUE) :
For argument `x` at least one of the following must be TRUE:
For argument `x` at least one of these must be TRUE:
At least one of the following should be TRUE for argument `x`, but none
are:
Bad argument `x`; does not pass any of the following:
- `length(1:3)` should be 1, not 3
- `1:3` should be type NULL, not integer
- `all(-1:1 > 0)` should evaluate to TRUE (is FALSE)
- Expression:
all(-1:1 > 0)
should evaluate to TRUE (is FALSE)
Error in fun2(x = matrix(c(1:9), nrow = 3), y = -1:1, z = NULL) :
Argument have `all(-1:1 > 0)` evaluate to TRUE (is FALSE)
For argument `y`, `all(-1:1 > 0)` should evaluate to TRUE (is FALSE)
Error in fun2(x = 1:3, y = TRUE) :
Argument y wrong because TRUE should be type integer, not logical
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must:
- have length 1 (is 3), OR
- have type "NULL" (is "integer"), OR
- have class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
Error in validate(INT.1 || NULL, 1:3) :
Argument `current` must have:
- length 1 (is 3), OR
- type "NULL" (is "integer"), OR
- class "matrix" (is "integer"), OR
- `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inheritance from class "foo"
Argument `current` should:
- be length 1 (is 3), OR
- be type "NULL" (is "integer"), OR
- be class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
Argument `current` should:
- be length 1 (is 3), OR
- be type "NULL" (is "integer"), OR
- be class "matrix" (is "integer"), OR
- have `!is.na(current)` all TRUE (is FALSE, contains FALSE, contains NA)
- inherit from class "foo"
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.