digitalmars.D - can we un-deprecate .ptr on arrays in safe code? cf issue 18529
- Timothee Cour (1/1) Feb 27 2018 see rationale in https://issues.dlang.org/show_bug.cgi?id=18529
- Mike Franklin (6/7) Feb 27 2018 It looks like the actual deprecation was made with this PR:
- bauss (5/6) Feb 27 2018 All I see is a rationale for how it can't be replaced, but not a
- Stefan Koch (2/9) Feb 27 2018 Checking if an array is the slice of another.
- bauss (4/17) Feb 27 2018 Like?
- Simen =?UTF-8?B?S2rDpnLDpXM=?= (13/31) Feb 27 2018 That only checks if the first element is the same. For a full 'is
- Jonathan M Davis (55/58) Feb 27 2018 As I understand it, the way that @safety checks generally work is they c...
- Timothee Cour (15/73) Feb 27 2018 this would be more bearable if there was a standard @trusted method to
- Steven Schveighoffer (10/22) Feb 27 2018 That completely defeats the purpose of the restriction! Of course we
- Seb (2/14) Feb 27 2018 aka isSliceOf -> https://github.com/dlang/phobos/pull/6147
- Dukc (9/10) Feb 28 2018 I don't think a just iterated array is automatically set to null,
- Cym13 (5/15) Feb 28 2018 If it cannot be proven safe by the compiler but is safe anyway it
- Jonathan M Davis (26/39) Feb 27 2018 Except that that's really not how @trusted is supposed to be used. The
- Atila Neves (9/31) Feb 27 2018 There's a common case where it's not equivalent - when the
- ag0aep6g (28/42) Feb 27 2018 As Jonathan says, you have to manually verify that it's safe, because it...
- Steven Schveighoffer (5/13) Feb 27 2018 fun(x.length ? &x[0] : null);
- Steven Schveighoffer (9/21) Feb 27 2018 Hm... borrowing from Timothee's suggestion:
- Timothee Cour (3/28) Feb 27 2018 how about this: https://github.com/dlang/phobos/pull/6231
- Walter Bright (2/15) Feb 27 2018 Yes, this bears repeating.
see rationale in https://issues.dlang.org/show_bug.cgi?id=18529
Feb 27 2018
On Tuesday, 27 February 2018 at 08:43:32 UTC, Timothee Cour wrote:see rationale in https://issues.dlang.org/show_bug.cgi?id=18529It looks like the actual deprecation was made with this PR: https://github.com/dlang/dmd/pull/5860 Meaning it's been deprecated for more than a year and a half. I think you're going to have to take it up with Walter. Mike
Feb 27 2018
On Tuesday, 27 February 2018 at 08:43:32 UTC, Timothee Cour wrote:see rationale in https://issues.dlang.org/show_bug.cgi?id=18529All I see is a rationale for how it can't be replaced, but not a rationale for actually doing it. I can't imagine a "safe" situation where you want to do it. Tbh. just put your code in trusted
Feb 27 2018
On Tuesday, 27 February 2018 at 09:23:19 UTC, bauss wrote:On Tuesday, 27 February 2018 at 08:43:32 UTC, Timothee Cour wrote:Checking if an array is the slice of another.see rationale in https://issues.dlang.org/show_bug.cgi?id=18529All I see is a rationale for how it can't be replaced, but not a rationale for actually doing it. I can't imagine a "safe" situation where you want to do it. Tbh. just put your code in trusted
Feb 27 2018
On Tuesday, 27 February 2018 at 09:47:51 UTC, Stefan Koch wrote:On Tuesday, 27 February 2018 at 09:23:19 UTC, bauss wrote:Like? assert(a.ptr != b.ptr); Which already works in safe.On Tuesday, 27 February 2018 at 08:43:32 UTC, Timothee Cour wrote:Checking if an array is the slice of another.see rationale in https://issues.dlang.org/show_bug.cgi?id=18529All I see is a rationale for how it can't be replaced, but not a rationale for actually doing it. I can't imagine a "safe" situation where you want to do it. Tbh. just put your code in trusted
Feb 27 2018
On Tuesday, 27 February 2018 at 09:58:00 UTC, bauss wrote:On Tuesday, 27 February 2018 at 09:47:51 UTC, Stefan Koch wrote:That only checks if the first element is the same. For a full 'is slice' check you'd need something like this: assert(a.ptr <= b.ptr && b.ptr + b.length <= a.ptr + a.length); // Or: auto c = b.ptr - a.ptr; assert(c >= 0 && c + b.length <= a.length); And trust me, the compiler complains about both of these. Possibly rightfully in the first example, but the latter never does anything scary with the given pointers. -- SimenOn Tuesday, 27 February 2018 at 09:23:19 UTC, bauss wrote:Like? assert(a.ptr != b.ptr); Which already works in safe.On Tuesday, 27 February 2018 at 08:43:32 UTC, Timothee Cour wrote:Checking if an array is the slice of another.see rationale in https://issues.dlang.org/show_bug.cgi?id=18529All I see is a rationale for how it can't be replaced, but not a rationale for actually doing it. I can't imagine a "safe" situation where you want to do it. Tbh. just put your code in trusted
Feb 27 2018
On Tuesday, February 27, 2018 11:33:04 Simen Kjærås via Digitalmars-d wrote:And trust me, the compiler complains about both of these. Possibly rightfully in the first example, but the latter never does anything scary with the given pointers.As I understand it, the way that safety checks generally work is they check whether a particular operation is safe or not. They don't usually care about what is then done with the result. So, if you do something like take the address of something, that's immediately system regardless of what you do with the result. That changes on some level with DIP 1000 and scope, because then it uses scope to ensure that the lifetime of stuff like pointers doesn't exceed the lifetime of what they point to so that it can then know that taking the address is safe, but without DIP 1000, it takes very little for something to become system. e.g. this is compiles with -dip1000 but otherwise doesn't: void main() safe { int i; assert(&i !is null); } Now, the compiler does seem to be a bit smarter with dynamic arrays and ptr given that this compiles without -dip1000 void main() safe { int[] i; assert(i.ptr !is null); } However, this doesn't compile with -dip1000: void main() safe { int[] i; auto j = i.ptr; assert(j !is null); } and not even this compiles with -dip1000: void main() safe { int[] i; scope j = i.ptr; assert(j !is null); } though I'm inclined to think that that's a bug from what I understand of -dip1000. In any case, safety checks tend to be fairly primitive, so once you start mucking around with pointers, it's not hard to write code that gets treated as system because of a single expression in the code that is clearly safe within the context of the function, but the compiler can't see it. And for better or worse, accessing a dynamic array's ptr member is now system, because it's not safe in all circumstances. If the compiler were smarter, then a number of uses of ptr would probably be safe, but its analysis for stuff like that is usually pretty primitive, in part because making it sophisticated requires stuff like code flow analysis, which the compiler doesn't do a lot of, precisely because it is complicated and easy to get wrong. Walter is particularly leery about making it so that stuff is an error or not based on code flow analysis, and safe falls into that camp. Clearly, some of that is going on with DIP 1000, but that seems to be largely by using the type system to solve the problem rather than doing much in the way of code flow analysis. - Jonathan M Davis
Feb 27 2018
this would be more bearable if there was a standard trusted method to get array `.ptr`, eg: in `object.d` (so that it's indeed standard) ``` trusted nogc pure nothrow auto pointer(T)(T a){ return a.ptr; } ``` again, the deprecation message is misleading because `&a[0]` isn't equivalent to `a.ptr` having something like `pointer` (and making deprecation msg use that) would be a better mitigation On Tue, Feb 27, 2018 at 3:56 AM, Jonathan M Davis via Digitalmars-d <digitalmars-d puremagic.com> wrote:On Tuesday, February 27, 2018 11:33:04 Simen Kjærås via Digitalmars-d wrote:And trust me, the compiler complains about both of these. Possibly rightfully in the first example, but the latter never does anything scary with the given pointers.As I understand it, the way that safety checks generally work is they check whether a particular operation is safe or not. They don't usually care about what is then done with the result. So, if you do something like take the address of something, that's immediately system regardless of what you do with the result. That changes on some level with DIP 1000 and scope, because then it uses scope to ensure that the lifetime of stuff like pointers doesn't exceed the lifetime of what they point to so that it can then know that taking the address is safe, but without DIP 1000, it takes very little for something to become system. e.g. this is compiles with -dip1000 but otherwise doesn't: void main() safe { int i; assert(&i !is null); } Now, the compiler does seem to be a bit smarter with dynamic arrays and ptr given that this compiles without -dip1000 void main() safe { int[] i; assert(i.ptr !is null); } However, this doesn't compile with -dip1000: void main() safe { int[] i; auto j = i.ptr; assert(j !is null); } and not even this compiles with -dip1000: void main() safe { int[] i; scope j = i.ptr; assert(j !is null); } though I'm inclined to think that that's a bug from what I understand of -dip1000. In any case, safety checks tend to be fairly primitive, so once you start mucking around with pointers, it's not hard to write code that gets treated as system because of a single expression in the code that is clearly safe within the context of the function, but the compiler can't see it. And for better or worse, accessing a dynamic array's ptr member is now system, because it's not safe in all circumstances. If the compiler were smarter, then a number of uses of ptr would probably be safe, but its analysis for stuff like that is usually pretty primitive, in part because making it sophisticated requires stuff like code flow analysis, which the compiler doesn't do a lot of, precisely because it is complicated and easy to get wrong. Walter is particularly leery about making it so that stuff is an error or not based on code flow analysis, and safe falls into that camp. Clearly, some of that is going on with DIP 1000, but that seems to be largely by using the type system to solve the problem rather than doing much in the way of code flow analysis. - Jonathan M Davis
Feb 27 2018
On 2/27/18 7:20 AM, Timothee Cour wrote:this would be more bearable if there was a standard trusted method to get array `.ptr`, eg: in `object.d` (so that it's indeed standard) ``` trusted nogc pure nothrow auto pointer(T)(T a){ return a.ptr; } ```That completely defeats the purpose of the restriction! Of course we aren't going to do that :) This is the standard way (and it's on you to make sure you don't mess it up): auto x = (() trusted => a.ptr)();again, the deprecation message is misleading because `&a[0]` isn't equivalent to `a.ptr`It isn't equivalent, but it is a safe mechanism to do it. If you want to do unsafe things, then you need to use trusted escapes, druntime/phobos is not going to help you there. -Steve
Feb 27 2018
On Tuesday, 27 February 2018 at 11:33:04 UTC, Simen Kjærås wrote:That only checks if the first element is the same. For a full 'is slice' check you'd need something like this: assert(a.ptr <= b.ptr && b.ptr + b.length <= a.ptr + a.length); // Or: auto c = b.ptr - a.ptr; assert(c >= 0 && c + b.length <= a.length); And trust me, the compiler complains about both of these. Possibly rightfully in the first example, but the latter never does anything scary with the given pointers. -- Simenaka isSliceOf -> https://github.com/dlang/phobos/pull/6147
Feb 27 2018
I don't think a just iterated array is automatically set to null, so taking it's pointer won't hit a memory-proteted area. So undeprectating arr.ptr in safe would break safety and be a step backward. If this is an issue, one can define a trusted function which takes a starting pointer from array and casts it to size_t before returning it so memory corruption cannot happen via it. On Tuesday, 27 February 2018 at 09:47:51 UTC, Stefan Koch wrote:Checking if an array is the slice of another.For that there is also std.array.overlap().
Feb 28 2018
On Wednesday, 28 February 2018 at 22:34:07 UTC, Dukc wrote:I don't think a just iterated array is automatically set to null, so taking it's pointer won't hit a memory-proteted area. So undeprectating arr.ptr in safe would break safety and be a step backward.If it cannot be proven safe by the compiler but is safe anyway it doesn't break safe, it should just be trusted. That's what trusted is about, not about abusing functions to trick unsafe things into compiling safe-ly.If this is an issue, one can define a trusted function which takes a starting pointer from array and casts it to size_t before returning it so memory corruption cannot happen via it. On Tuesday, 27 February 2018 at 09:47:51 UTC, Stefan Koch wrote:Checking if an array is the slice of another.For that there is also std.array.overlap().
Feb 28 2018
On Tuesday, February 27, 2018 04:20:38 Timothee Cour via Digitalmars-d wrote:this would be more bearable if there was a standard trusted method to get array `.ptr`, eg: in `object.d` (so that it's indeed standard) ``` trusted nogc pure nothrow auto pointer(T)(T a){ return a.ptr; } ```Except that that's really not how trusted is supposed to be used. The programmer needs to verify that the caller is using a.ptr in a manner that is actually safe, because the compiler is not smart enough to determine that for you. Wrapping it in an trusted function means that the caller won't get an error and that the programmer won't necessarily know that they need to verify the calling code. It's the code that's using ptr that needs to be verified, not the actual accessing of ptr. Hiding the access to ptr within an trusted function goes against that entire idea of trusted and makes it easy to use ptr without realizing that you need to be checking the code that's using it, since you just called a wrapper function to silence the compiler instead of listening the compiler and studying the code using ptr to verify its safety.again, the deprecation message is misleading because `&a[0]` isn't equivalent to `a.ptr` having something like `pointer` (and making deprecation msg use that) would be a better mitigationIn almost all cases, &a[0] is equivalent to a.ptr except that it does bounds checking, so it's actually safe and thus doesn't need to be manually verified by the programmer, unlike your pointer function suggestion. If you have a use case where you need a.ptr rather than &a[0], then that just means that you're going to have to verify that your code is safe in spite of using an system operation and mark it trusted if it is. Simply treating a.ptr as safe when it isn't or wrapping it in an trusted function just defeats the purpose of the whole safety system. Now, maybe in some simple cases, the compiler can be improved to detect that what you're doing with a.ptr is actually safe and not give an error, but as long as it can't, any use of a.ptr needs to be verified by the programmer. - Jonathan M Davis
Feb 27 2018
On Tuesday, 27 February 2018 at 12:39:04 UTC, Jonathan M Davis wrote:On Tuesday, February 27, 2018 04:20:38 Timothee Cour via Digitalmars-d wrote:There's a common case where it's not equivalent - when the pointer is null. Imagine I have a C function I want to call: extern(C) void fun(int* things); Imagine also that it's ok to call with null. Well, now I can't use a slice to call this and have it be 1) safe and 2) not throw RangeError. I ran into this the other way. Atila[...]Except that that's really not how trusted is supposed to be used. The programmer needs to verify that the caller is using a.ptr in a manner that is actually safe, because the compiler is not smart enough to determine that for you. Wrapping it in an trusted function means that the caller won't get an error and that the programmer won't necessarily know that they need to verify the calling code. It's the code that's using ptr that needs to be verified, not the actual accessing of ptr. Hiding the access to ptr within an trusted function goes against that entire idea of trusted and makes it easy to use ptr without realizing that you need to be checking the code that's using it, since you just called a wrapper function to silence the compiler instead of listening the compiler and studying the code using ptr to verify its safety.[...]In almost all cases, &a[0] is equivalent to a.ptr except that it does bounds checking, so it's actually safe and thus doesn't need to be manually verified by the programmer, unlike your pointer function suggestion.
Feb 27 2018
On 02/27/2018 06:32 PM, Atila Neves wrote:On Tuesday, 27 February 2018 at 12:39:04 UTC, Jonathan M Davis wrote:[...]As Jonathan says, you have to manually verify that it's safe, because it generally isn't. `arr.ptr` can be invalid but not null, even in safe code. Consider: ---- void fun(int* things) safe { int x = things is null ? 0 : *things; import std.stdio; writeln(x); } void main() { int[] arr; fun(arr.ptr); /* Ok, prints "0". */ arr = [1, 2, 3]; fun(arr.ptr); /* Ok, prints "1".*/ arr = arr[$ .. $]; /* This is safe. */ fun(arr.ptr); /* Not ok, prints garbage. */ } ---- The first two calls are actually safe and can be trusted. The third call is invalid and must not be possible in safe code. Maybe it would be possible require `arr.ptr` to be valid or null in safe code. This would outlaw `arr[$ .. $]` and bounds checking would have to catch it. I don't know if that would be any practical than banning `arr.ptr`.In almost all cases, &a[0] is equivalent to a.ptr except that it does bounds checking, so it's actually safe and thus doesn't need to be manually verified by the programmer, unlike your pointer function suggestion.There's a common case where it's not equivalent - when the pointer is null. Imagine I have a C function I want to call: extern(C) void fun(int* things); Imagine also that it's ok to call with null. Well, now I can't use a slice to call this and have it be 1) safe and 2) not throw RangeError. I ran into this the other way.
Feb 27 2018
On 2/27/18 12:32 PM, Atila Neves wrote:There's a common case where it's not equivalent - when the pointer is null. Imagine I have a C function I want to call: extern(C) void fun(int* things); Imagine also that it's ok to call with null. Well, now I can't use a slice to call this and have it be 1) safe and 2) not throw RangeError. I ran into this the other way.fun(x.length ? &x[0] : null); I think even the compiler could elide the bounds check since you already did it, but I'm not sure that actually happens. -Steve
Feb 27 2018
On 2/27/18 3:00 PM, Steven Schveighoffer wrote:On 2/27/18 12:32 PM, Atila Neves wrote:Hm... borrowing from Timothee's suggestion: trusted nogc pure nothrow T* pointer(T)(T[] a){ return a.length > 0 ? a.ptr : null; } This would be fine and safe, but may not be useful for all purposes. However, it would fix your issue. -SteveThere's a common case where it's not equivalent - when the pointer is null. Imagine I have a C function I want to call: extern(C) void fun(int* things); Imagine also that it's ok to call with null. Well, now I can't use a slice to call this and have it be 1) safe and 2) not throw RangeError. I ran into this the other way.fun(x.length ? &x[0] : null);
Feb 27 2018
Hm... borrowing from Timothee's suggestion: This would be fine and safe, but may not be useful for all purposes. However, it would fix your issue.how about this: https://github.com/dlang/phobos/pull/6231 On Tue, Feb 27, 2018 at 12:09 PM, Steven Schveighoffer via Digitalmars-d <digitalmars-d puremagic.com> wrote:On 2/27/18 3:00 PM, Steven Schveighoffer wrote:On 2/27/18 12:32 PM, Atila Neves wrote:Hm... borrowing from Timothee's suggestion: trusted nogc pure nothrow T* pointer(T)(T[] a){ return a.length > 0 ? a.ptr : null; } This would be fine and safe, but may not be useful for all purposes. However, it would fix your issue. -SteveThere's a common case where it's not equivalent - when the pointer is null. Imagine I have a C function I want to call: extern(C) void fun(int* things); Imagine also that it's ok to call with null. Well, now I can't use a slice to call this and have it be 1) safe and 2) not throw RangeError. I ran into this the other way.fun(x.length ? &x[0] : null);
Feb 27 2018
On 2/27/2018 4:39 AM, Jonathan M Davis wrote:Except that that's really not how trusted is supposed to be used. The programmer needs to verify that the caller is using a.ptr in a manner that is actually safe, because the compiler is not smart enough to determine that for you. Wrapping it in an trusted function means that the caller won't get an error and that the programmer won't necessarily know that they need to verify the calling code. It's the code that's using ptr that needs to be verified, not the actual accessing of ptr. Hiding the access to ptr within an trusted function goes against that entire idea of trusted and makes it easy to use ptr without realizing that you need to be checking the code that's using it, since you just called a wrapper function to silence the compiler instead of listening the compiler and studying the code using ptr to verify its safety.Yes, this bears repeating.
Feb 27 2018