Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
On 6/21/21 9:25 PM, Bill Dunlap wrote: NULL cannot be in an integer or numeric vector so it would not be a good fit for substring's 'first' or 'last' argument (or substr's 'start' and 'stop'). Yes, that would only work if used as a scalar, such as in the default for 'last' where 100L is used now. In other cases, users already had to provide their own values for 'last' explicitly, and hence they would know if they provided a value too small given their data. Also, it is conceivable that string lengths may be 64 bit integers in the future, so why not use Inf as the default? Then the following would give 4 identical results with no warning: Yes, that would work also in vector use, but integers over 2^53 won't be representable as doubles exactly, so we would have to revisit/change the interface when moving to 64 bit integers. Yet another option would be say using -1, that would also work with vector use and integers. But, negative indexes (and zero) are now treated as start of the string (1), and while not documented, perhaps this is good/intuitive behavior. Tomas substring("abcde", 3, c(10, 2^31-1, 2^31, Inf)) [1] "cde" "cde" NANA Warning message: In substring("abcde", 3, c(10, 2^31 - 1, 2^31, Inf)) : NAs introduced by coercion to integer range -Bill On Mon, Jun 21, 2021 at 10:22 AM Michael Chirico wrote: Thanks all, great points well taken. Indeed it seems the default of 100 predates SVN tracking in 1997. I think a NULL default behaving as "end of string" regardless of encoding makes sense and avoids the overheads of a $ call and a much heavier nchar() calculation. Mike C On Mon, Jun 21, 2021 at 1:32 AM Martin Maechler wrote: Tomas Kalibera on Mon, 21 Jun 2021 10:08:37 +0200 writes: > On 6/21/21 9:35 AM, Martin Maechler wrote: >>> Michael Chirico >>> on Sun, 20 Jun 2021 15:20:26 -0700 writes: >> > Currently, substring defaults to last=100L, which >> > strongly suggests the intent is to default to "nchar(x)" >> > without having to compute/allocate that up front. >> >> > Unfortunately, this default makes no sense for "very >> > large" strings which may exceed 100L in "width". >> >> Yes; and I tend to agree with you that this default is outdated >> (Remember : R was written to work and run on 2 (or 4?) MB of RAM on the >> student lab Macs in Auckland in ca 1994). >> >> > The max width of a string is .Machine$integer.max-1: >> >> (which Brodie showed was only almost true) >> >> > So it seems to me either .Machine$integer.max or >> > .Machine$integer.max-1L would be a more sensible default. Am I missing >> > something? >> >> The "drawback" is of course that .Machine$integer.max is still >> a function call (as R beginners may forget) contrary to L, >> but that may even be inlined by the byte compiler (? how would we check ?) >> and even if it's not, it does more clearly convey the concept >> and idea *and* would probably even port automatically if ever >> integer would be increased in R. > We still have the problem that we need to count characters, not bytes, > if we want the default semantics of "until the end of the string". > I think we would have to fix this either by really using > "nchar(type="c"))" or by using e.g. NULL and then treating this as a > special case, that would be probably faster. > Tomas You are right, as always, Tomas. I agree that would be better and we should do it if/when we change the default there. Martin __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel [[alternative HTML version deleted]] __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
NULL cannot be in an integer or numeric vector so it would not be a good fit for substring's 'first' or 'last' argument (or substr's 'start' and 'stop'). Also, it is conceivable that string lengths may be 64 bit integers in the future, so why not use Inf as the default? Then the following would give 4 identical results with no warning: > substring("abcde", 3, c(10, 2^31-1, 2^31, Inf)) [1] "cde" "cde" NANA Warning message: In substring("abcde", 3, c(10, 2^31 - 1, 2^31, Inf)) : NAs introduced by coercion to integer range -Bill On Mon, Jun 21, 2021 at 10:22 AM Michael Chirico wrote: > Thanks all, great points well taken. Indeed it seems the default of > 100 predates SVN tracking in 1997. > > I think a NULL default behaving as "end of string" regardless of > encoding makes sense and avoids the overheads of a $ call and a much > heavier nchar() calculation. > > Mike C > > On Mon, Jun 21, 2021 at 1:32 AM Martin Maechler > wrote: > > > > > Tomas Kalibera > > > on Mon, 21 Jun 2021 10:08:37 +0200 writes: > > > > > On 6/21/21 9:35 AM, Martin Maechler wrote: > > >>> Michael Chirico > > >>> on Sun, 20 Jun 2021 15:20:26 -0700 writes: > > >> > Currently, substring defaults to last=100L, which > > >> > strongly suggests the intent is to default to "nchar(x)" > > >> > without having to compute/allocate that up front. > > >> > > >> > Unfortunately, this default makes no sense for "very > > >> > large" strings which may exceed 100L in "width". > > >> > > >> Yes; and I tend to agree with you that this default is outdated > > >> (Remember : R was written to work and run on 2 (or 4?) MB of RAM > on the > > >> student lab Macs in Auckland in ca 1994). > > >> > > >> > The max width of a string is .Machine$integer.max-1: > > >> > > >> (which Brodie showed was only almost true) > > >> > > >> > So it seems to me either .Machine$integer.max or > > >> > .Machine$integer.max-1L would be a more sensible default. Am I > missing > > >> > something? > > >> > > >> The "drawback" is of course that .Machine$integer.max is still > > >> a function call (as R beginners may forget) contrary to L, > > >> but that may even be inlined by the byte compiler (? how would we > check ?) > > >> and even if it's not, it does more clearly convey the concept > > >> and idea *and* would probably even port automatically if ever > > >> integer would be increased in R. > > > > > We still have the problem that we need to count characters, not > bytes, > > > if we want the default semantics of "until the end of the string". > > > > > I think we would have to fix this either by really using > > > "nchar(type="c"))" or by using e.g. NULL and then treating this as > a > > > special case, that would be probably faster. > > > > > Tomas > > > > You are right, as always, Tomas. > > I agree that would be better and we should do it if/when we change > > the default there. > > > > Martin > > __ > R-devel@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-devel > [[alternative HTML version deleted]] __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
Thanks all, great points well taken. Indeed it seems the default of 100 predates SVN tracking in 1997. I think a NULL default behaving as "end of string" regardless of encoding makes sense and avoids the overheads of a $ call and a much heavier nchar() calculation. Mike C On Mon, Jun 21, 2021 at 1:32 AM Martin Maechler wrote: > > > Tomas Kalibera > > on Mon, 21 Jun 2021 10:08:37 +0200 writes: > > > On 6/21/21 9:35 AM, Martin Maechler wrote: > >>> Michael Chirico > >>> on Sun, 20 Jun 2021 15:20:26 -0700 writes: > >> > Currently, substring defaults to last=100L, which > >> > strongly suggests the intent is to default to "nchar(x)" > >> > without having to compute/allocate that up front. > >> > >> > Unfortunately, this default makes no sense for "very > >> > large" strings which may exceed 100L in "width". > >> > >> Yes; and I tend to agree with you that this default is outdated > >> (Remember : R was written to work and run on 2 (or 4?) MB of RAM on > the > >> student lab Macs in Auckland in ca 1994). > >> > >> > The max width of a string is .Machine$integer.max-1: > >> > >> (which Brodie showed was only almost true) > >> > >> > So it seems to me either .Machine$integer.max or > >> > .Machine$integer.max-1L would be a more sensible default. Am I > missing > >> > something? > >> > >> The "drawback" is of course that .Machine$integer.max is still > >> a function call (as R beginners may forget) contrary to L, > >> but that may even be inlined by the byte compiler (? how would we > check ?) > >> and even if it's not, it does more clearly convey the concept > >> and idea *and* would probably even port automatically if ever > >> integer would be increased in R. > > > We still have the problem that we need to count characters, not bytes, > > if we want the default semantics of "until the end of the string". > > > I think we would have to fix this either by really using > > "nchar(type="c"))" or by using e.g. NULL and then treating this as a > > special case, that would be probably faster. > > > Tomas > > You are right, as always, Tomas. > I agree that would be better and we should do it if/when we change > the default there. > > Martin __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
> Tomas Kalibera > on Mon, 21 Jun 2021 10:08:37 +0200 writes: > On 6/21/21 9:35 AM, Martin Maechler wrote: >>> Michael Chirico >>> on Sun, 20 Jun 2021 15:20:26 -0700 writes: >> > Currently, substring defaults to last=100L, which >> > strongly suggests the intent is to default to "nchar(x)" >> > without having to compute/allocate that up front. >> >> > Unfortunately, this default makes no sense for "very >> > large" strings which may exceed 100L in "width". >> >> Yes; and I tend to agree with you that this default is outdated >> (Remember : R was written to work and run on 2 (or 4?) MB of RAM on the >> student lab Macs in Auckland in ca 1994). >> >> > The max width of a string is .Machine$integer.max-1: >> >> (which Brodie showed was only almost true) >> >> > So it seems to me either .Machine$integer.max or >> > .Machine$integer.max-1L would be a more sensible default. Am I missing >> > something? >> >> The "drawback" is of course that .Machine$integer.max is still >> a function call (as R beginners may forget) contrary to L, >> but that may even be inlined by the byte compiler (? how would we check ?) >> and even if it's not, it does more clearly convey the concept >> and idea *and* would probably even port automatically if ever >> integer would be increased in R. > We still have the problem that we need to count characters, not bytes, > if we want the default semantics of "until the end of the string". > I think we would have to fix this either by really using > "nchar(type="c"))" or by using e.g. NULL and then treating this as a > special case, that would be probably faster. > Tomas You are right, as always, Tomas. I agree that would be better and we should do it if/when we change the default there. Martin __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
On 6/21/21 9:35 AM, Martin Maechler wrote: Michael Chirico on Sun, 20 Jun 2021 15:20:26 -0700 writes: > Currently, substring defaults to last=100L, which > strongly suggests the intent is to default to "nchar(x)" > without having to compute/allocate that up front. > Unfortunately, this default makes no sense for "very > large" strings which may exceed 100L in "width". Yes; and I tend to agree with you that this default is outdated (Remember : R was written to work and run on 2 (or 4?) MB of RAM on the student lab Macs in Auckland in ca 1994). > The max width of a string is .Machine$integer.max-1: (which Brodie showed was only almost true) > So it seems to me either .Machine$integer.max or > .Machine$integer.max-1L would be a more sensible default. Am I missing > something? The "drawback" is of course that .Machine$integer.max is still a function call (as R beginners may forget) contrary to L, but that may even be inlined by the byte compiler (? how would we check ?) and even if it's not, it does more clearly convey the concept and idea *and* would probably even port automatically if ever integer would be increased in R. We still have the problem that we need to count characters, not bytes, if we want the default semantics of "until the end of the string". I think we would have to fix this either by really using "nchar(type="c"))" or by using e.g. NULL and then treating this as a special case, that would be probably faster. Tomas Martin __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
> Michael Chirico > on Sun, 20 Jun 2021 15:20:26 -0700 writes: > Currently, substring defaults to last=100L, which > strongly suggests the intent is to default to "nchar(x)" > without having to compute/allocate that up front. > Unfortunately, this default makes no sense for "very > large" strings which may exceed 100L in "width". Yes; and I tend to agree with you that this default is outdated (Remember : R was written to work and run on 2 (or 4?) MB of RAM on the student lab Macs in Auckland in ca 1994). > The max width of a string is .Machine$integer.max-1: (which Brodie showed was only almost true) > So it seems to me either .Machine$integer.max or > .Machine$integer.max-1L would be a more sensible default. Am I missing > something? The "drawback" is of course that .Machine$integer.max is still a function call (as R beginners may forget) contrary to L, but that may even be inlined by the byte compiler (? how would we check ?) and even if it's not, it does more clearly convey the concept and idea *and* would probably even port automatically if ever integer would be increased in R. Martin __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
> On Sunday, June 20, 2021, 9:29:28 PM EDT, brodie gaslam via R-devel > wrote: > >> On Sunday, June 20, 2021, 6:21:22 PM EDT, Michael Chirico >> wrote: >> >> The max width of a string is .Machine$integer.max-1: > > I think the max width is .Machine$integer.max. What happened below is a > bug due to buffer overflow in `strrep`: Sorry, integer overflow. >> # works >> x = strrep(" ", .Machine$integer.max-1L) >> # fails >> x = strrep(" ", .Machine$integer.max) >> Error in strrep(" ", .Machine$integer.max) : >> 'Calloc' could not allocate memory (18446744071562067968 of 1 bytes) >> (see also the comment in src/main/character.c: "Character strings in R >> are less than 2^31-1 bytes, so we use int not size_t.") > > FWIW WRE states: > >> Note that R character strings are restricted to 2^31 - 1 bytes > > This is INT_MAX or .Machine$integer.max, at least on machines for which > `int` is 32 bits, which I think typical for machines R builds on. From > having looked at the code a while ago I think WRE is right (so maybe the > comment in the code is wrong), but it was a while ago and I haven't tried > to allocate an INT_MAX long string. So I tried it on a machine with more memory, and it works: > x <- strrep(" ", .Machine$integer.max-1L) > x <- paste0(x, " ") > nchar(x) [1] 2147483647 > nchar(x) == .Machine$integer.max [1] TRUE B. __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Should last default to .Machine$integer.max-1 for substring()
> On Sunday, June 20, 2021, 6:21:22 PM EDT, Michael Chirico > wrote: > > Currently, substring defaults to last=100L, which strongly > suggests the intent is to default to "nchar(x)" without having to > compute/allocate that up front. > > Unfortunately, this default makes no sense for "very large" strings > which may exceed 100L in "width". > > The max width of a string is .Machine$integer.max-1: I think the max width is .Machine$integer.max. What happened below is a bug due to buffer overflow in `strrep`: > # works > x = strrep(" ", .Machine$integer.max-1L) > # fails > x = strrep(" ", .Machine$integer.max) > Error in strrep(" ", .Machine$integer.max) : > 'Calloc' could not allocate memory (18446744071562067968 of 1 bytes) Notice the very large number that was tried to be Calloc'ed. That's (size_t) -1. The problem is (src/include/R_ext/RS.h@85): #define CallocCharBuf(n) (char *) R_chk_calloc((R_SIZE_T) ((n)+1), sizeof(char)) The `((n) + 1)` overflows `int` and produces -1 (well, undefined behavior so who knows), which when cast to size_t produces that very large number which can't be allocated. I think this should be: #define CallocCharBuf(n) (char *) R_chk_calloc(((R_SIZE_T)(n))+1, sizeof(char)) I can reproduce the failure before the change. After the change I get: > x = strrep(" ", .Machine$integer.max) Error in strrep(" ", .Machine$integer.max) : 'Calloc' could not allocate memory (2147483648 of 1 bytes) I believe this to be the expected result on a machine that doesn't have enough memory to allocate INT_MAX + 1 bytes, as happens to be the case on my R build system (it's a VM that gets 2GB total as the host machine can barely spare that to begin with). > (see also the comment in src/main/character.c: "Character strings in R > are less than 2^31-1 bytes, so we use int not size_t.") FWIW WRE states: > Note that R character strings are restricted to 2^31 - 1 bytes This is INT_MAX or .Machine$integer.max, at least on machines for which `int` is 32 bits, which I think typical for machines R builds on. From having looked at the code a while ago I think WRE is right (so maybe the comment in the code is wrong), but it was a while ago and I haven't tried to allocate an INT_MAX long string. Sorry this doesn't answer your original question. Best, Brodie. > > > So it seems to me either .Machine$integer.max or > .Machine$integer.max-1L would be a more sensible default. Am I missing > something? > > Mike C > > __ > R-devel@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-devel > __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
[Rd] Should last default to .Machine$integer.max-1 for substring()
Currently, substring defaults to last=100L, which strongly suggests the intent is to default to "nchar(x)" without having to compute/allocate that up front. Unfortunately, this default makes no sense for "very large" strings which may exceed 100L in "width". The max width of a string is .Machine$integer.max-1: # works x = strrep(" ", .Machine$integer.max-1L) # fails x = strrep(" ", .Machine$integer.max) Error in strrep(" ", .Machine$integer.max) : 'Calloc' could not allocate memory (18446744071562067968 of 1 bytes) (see also the comment in src/main/character.c: "Character strings in R are less than 2^31-1 bytes, so we use int not size_t.") So it seems to me either .Machine$integer.max or .Machine$integer.max-1L would be a more sensible default. Am I missing something? Mike C __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel