Related
I have some C++11 code generated via a code generator that contains a large array of floats, and I want to make sure that the compiled values are precisely the same as the compiled values in the generator (assuming that both depend on the same float ISO norm)
So I figured the best way to do it is to store the values as hex representations and interpret them as float in the code.
Edit for Clarification: The code generator takes the float values and converts them to their corresponding hex representations. The target code is supposed to convert back to float.
It looks something like this:
const unsigned int data[3] = { 0x3d13f407U, 0x3ea27884U, 0xbe072dddU};
float const* ptr = reinterpret_cast<float const*>(&data[0]);
This works and gives me access to all the data element as floats, but I recently stumbled upon the fact that this is actually undefined behavior and only works because my compiler resolves it the way I intended:
https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8
https://en.cppreference.com/w/cpp/language/reinterpret_cast.
The standard basically says that reinterpret_cast is not defined between POD pointers of different type.
So basically I have three options:
Use memcopy and hope that the compiler will be able to optimize this
Store the data not as hex-values but in a different way.
Use std::bit_cast from C++20.
I cannot use 3) because I'm stuck with C++11.
I don't have the resources to store the data array twice, so I would have to rely on the compiler to optimize this. Due to this, I don't particularly like 1) because it could stop working if I changed compilers or compiler settings.
So that leaves me with 2):
Is there a standardized way to express float values in source code so that they map to the exact float value when compiled? Does the ISO float standard define this in a way that guarantees that any compiler will follow the interpretation? I imagine if I deviate from the way the compiler expects, I could run the risk that the float "neighbor" of the number I actually want is used.
I would also take alternative ideas if there is an option 4 I forgot.
How to express float constants precisely in source code
Use hexadecimal floating point literals. Assuming some endianess for the hexes you presented:
float floats[] = { 0x1.27e80ep-5, 0x1.44f108p-2, -0x1.0e5bbap-3 };
If you have the generated code produce the full representation of the floating-point value—all of the decimal digits needed to show its exact value—then a C++ 11 compiler is required to parse the number exactly.
C++ 11 draft N3092 2.14.4 1 says, of a floating literal:
… The exponent, if present, indicates the power of 10 by which the significant [likely typo, should be “significand”] part is to be scaled. If the scaled value is in the range of representable values for its type, the result is the scaled value if representable, else the larger or smaller representable value nearest the scaled value, chosen in an implementation-defined manner…
Thus, if the floating literal does not have all the digits needed to show the exact value, the implementation may round it either upward or downward, as the implementation defines. But if it does have all the digits, then the value represented by the floating literal is representable in the floating-point format, and so its value must be the result of the parsing.
I have read some very valuable information here and would like to throw in an option that does not strictly answer the question, but could be a solution.
It might be problematic, but if so, I would like to discuss it.
The simple solution would be: Leave it as it is.
A short rundown of why I am hesitant about the suggested options:
memcpy relies on the compiler to optimize away the actual copy and understand that I only want to read the values. Since I am having large arrays of data I would want to avoid a surprise event in which a compiler setting would be changed that suddenly introduces increased runtime and would require a fix on short notice.
bit_cast is only available from C++20. There are reference implementations but they basically use memcpy under the hood (see above).
hex float literals are only available from C++17
Directly writing the floats precisely... I don't know, it seems to be somewhat dangerous, because if I make a slight mistake I may end up with a data block that is slightly off and could have an impact on my classification results. A mistake like that would be a nightmare to spot.
So why do I think I can get away with an implementation that is strictly speaking undefined? The rationale is that the standard may not define it, but compiler manufacturers likely do, at least the ones I have worked with so far gave me exact results. The code has been running without major problems for a fairly long time, across dozens of code generator run and I would expect that a failed reinterpret_cast would break the conversion so severely that I would spot the result in my classification results right away.
Still not robust enough though. So my idea was to write a unit test that contains a significant number of hex-floats, do the reinterpret_cast and compare to reference float values for exact correspondence to tell me if a setting or compiler failed in this regard.
I have one doubt though: Is the assumption somewhat reasonable that a failed reinterpret_cast would break things spectacularly, or are the bets totally off when it comes to undefined behavior?
I am a bit worried that if the compiler implementation defines the undefined behavior in a way that it would pick a float that is close the hex value instead of the precise one (although I would wonder why), and that it happens only sporadically so that my unit test misses the problems.
So the endgame would be to unit test every single data entry against the corresponding reference float. Since the code is generated, I can generate the test as well. I think that should put all my worries to rest and make sure that I can get this to work across all possible compilers and compiler settings or be notified if anything breaks.
Note that this is purely an academic question, from a language lawyer perspective. It's about the theoretically safest way to accomplish the conversion.
Suppose I have a void* and I need to convert it to a 64-bit integer. The reason is that this pointer holds the address of a faulting instruction; I wish to report this to my backend to be logged, and I use a fixed-size protocol - so I have precisely 64 bits to use for the address.
The cast will of course be implementation defined. I know my platform (64-bit Windows) allows this conversion, so in practice it's fine to just reinterpret_cast<uint64_t>(address).
But I'm wondering: from a theoretical standpoint, is it any safer to first convert to uintptr_t? That is: static_cast<uint64_t>(reinterpret_cast<uintptr_t>(address)). https://en.cppreference.com/w/cpp/language/reinterpret_cast says (emphasis mine):
Unlike static_cast, but like const_cast, the reinterpret_cast expression does not compile to any CPU instructions (except when converting between integers and pointers or on obscure architectures where pointer representation depends on its type).
So, in theory, pointer representation is not defined to be anything in particular; going from pointer to uintptr_t might theoretically perform a conversion of some kind to make the pointer representable as an integer. After that, I forcibly extract the lower 64 bits. Whereas just directly casting to uint64_t would not trigger the conversion mentioned above, and so I'd get a different result.
Is my interpretation correct, or is there no difference whatsoever between the two casts in theory as well?
FWIW, on a 32-bit system, apparently the widening conversion to unsigned 64-bit could sign-extend, as in this case. But on 64-bit I shouldn't have that issue.
You’re parsing that (shockingly informal, for cppreference) paragraph too closely. The thing it’s trying to get at is simply that other casts potentially involve conversion operations (float/int stuff, sign extension, pointer adjustment), whereas reinterpret_cast has the flavor of direct reuse of the bits.
If you reinterpret a pointer as an integer and the integer type is not large enough, you get a compile-time error. If it is large enough, you’re fine. There’s nothing magical about uintptr_t other than the guarantee that (if it exists) it’s large enough, and if you then re-cast to a smaller type you lose that anyway. Either 64 bits is enough, in which case you get the same guarantees with either type, or it’s not, and you’re screwed no matter what you do. And if your implementation is willing to do something weird inside reinterpret_cast, which might give different results than (say) bit_cast, neither method will guarantee nor prevent that.
That’s not to say the two are guaranteed identical, of course. Consider a DS9k-ish architecture with 32-bit pointers, where reinterpret_cast of a pointer to a uint64_t resulted in the pointer bits being duplicated in the low and high words. There you’d get both copies if you went directly to a uint64_t, and zeros in the top half if you went through a 32-bit uintptr_t. In that case, which one was “right” would be a matter of personal opinion.
The question is clear.
I wonder why they even thought this would be handy, as clearly negative indices are unusable in the containers that would be used with them (see for example QList's docs).
I thought they wanted to allow that for some crazy form of indexing, but it seems unsupported?
It also generates a ton of (correct) compiler warnings about casting to and comparing of signed/unsigned types (on MSVC).
It just seems incompatible with the STL by design for some reason...
Although I am deeply sympathetic to Chris's line of reasoning, I will disagree here (at least in part, I am playing devil's advocate). There is nothing wrong with using unsigned types for sizes, and it can even be beneficial in some circumstances.
Chris's justification for signed size types is that they are naturally used as array indices, and you may want to do arithmetic on array indices, and that arithmetic may create temporary values that are negative.
That's fine, and unsigned arithmetic introduces no problem in doing so, as long as you make sure to interpret your values correctly when you do comparisons. Because the overflow behavior of unsigned integers is fully specified, temporary overflows into the negative range (or into huge positive numbers) do not introduce any error as long as they are corrected before a comparison is performed.
Sometimes, the overflow behavior is even desirable, as the overflow behavior of unsigned arithmetic makes certain range checks expressible as a single comparison that would require two comparisons otherwise. If I want to check if x is in the range [a,b] and all the values are unsigned, I can simply do:
if (x - a < b - a) {
}
That doesn't work with signed variables; such range checks are pretty common with sizes and array offsets.
I mentioned before that a benefit is that overflow arithmetic has defined results. If your index arithmetic overflows a signed type, the behavior is implementation defined; there is no way to make your program portable. Use an unsigned type and this problem goes away. Admittedly this only applies to huge offsets, but it is a concern for some uses.
Basically, the objections to unsigned types are frequently overstated. The real problem is that most programmers don't really think about the exact semantics of the code they write, and for small integer values, signed types behave more nearly in line with their intuition. However, data sizes grow pretty fast. When we deal with buffers or databases, we're frequently way outside of the range of "small", and signed overflow is far more problematic to handle correctly than is unsigned overflow. The solution is not "don't use unsigned types", it is "think carefully about the code you are writing, and make sure you understand it".
Because, realistically, you usually want to perform arithmetic on indices, which means that you might want to create temporaries that are negative.
This is clearly painful when the underlying indexing type is unsigned.
The only appropriate time to use unsigned numbers is with modulus arithmetic.
Using "unsgined" as some kind of contract specifier "a number in the range [0..." is just clumsy, and too coarse to be useful.
Consider: What type should I use to represent the idea that the number should be a positive integer between 1 and 10? Why is 0...2^x a more special range?
(I guess this question could apply to many typed languages, but I chose to use C++ as an example.)
Why is there no way to just write:
struct foo {
little int x; // little-endian
big long int y; // big-endian
short z; // native endianness
};
to specify the endianness for specific members, variables and parameters?
Comparison to signedness
I understand that the type of a variable not only determines how many bytes are used to store a value but also how those bytes are interpreted when performing computations.
For example, these two declarations each allocate one byte, and for both bytes, every possible 8-bit sequence is a valid value:
signed char s;
unsigned char u;
but the same binary sequence might be interpreted differently, e.g. 11111111 would mean -1 when assigned to s but 255 when assigned to u. When signed and unsigned variables are involved in the same computation, the compiler (mostly) takes care of proper conversions.
In my understanding, endianness is just a variation of the same principle: a different interpretation of a binary pattern based on compile-time information about the memory in which it will be stored.
It seems obvious to have that feature in a typed language that allows low-level programming. However, this is not a part of C, C++ or any other language I know, and I did not find any discussion about this online.
Update
I'll try to summarize some takeaways from the many comments that I got in the first hour after asking:
signedness is strictly binary (either signed or unsigned) and will always be, in contrast to endianness, which also has two well-known variants (big and little), but also lesser-known variants such as mixed/middle endian. New variants might be invented in the future.
endianness matters when accessing multiple-byte values byte-wise. There are many aspects beyond just endianness that affect the memory layout of multi-byte structures, so this kind of access is mostly discouraged.
C++ aims to target an abstract machine and minimize the number of assumptions about the implementation. This abstract machine does not have any endianness.
Also, now I realize that signedness and endianness are not a perfect analogy, because:
endianness only defines how something is represented as a binary sequence, but now what can be represented. Both big int and little int would have the exact same value range.
signedness defines how bits and actual values map to each other, but also affects what can be represented, e.g. -3 can't be represented by an unsigned char and (assuming that char has 8 bits) 130 can't be represented by a signed char.
So that changing the endianness of some variables would never change the behavior of the program (except for byte-wise access), whereas a change of signedness usually would.
What the standard says
[intro.abstract]/1:
The semantic descriptions in this document define a parameterized nondeterministic abstract machine.
This document places no requirement on the structure of conforming implementations.
In particular, they need not copy or emulate the structure of the abstract machine.
Rather, conforming implementations are required to emulate (only) the observable behavior of the abstract machine as explained below.
C++ could not define an endianness qualifier since it has no concept of endianness.
Discussion
About the difference between signness and endianness, OP wrote
In my understanding, endianness is just a variation of the same principle [(signness)]: a different interpretation of a binary pattern based on compile-time information about the memory in which it will be stored.
I'd argue signness both have a semantic and a representative aspect1. What [intro.abstract]/1 implies is that C++ only care about semantic, and never addresses the way a signed number should be represented in memory2. Actually, "sign bit" only appears once in the C++ specs and refer to an implementation-defined value.
On the other hand, endianness only have a representative aspect: endianness conveys no meaning.
With C++20, std::endian appears. It is still implementation-defined, but let us test the endian of the host without depending on old tricks based on undefined behaviour.
1) Semantic aspect: an signed integer can represent values below zero; representative aspect: one need to, for example, reserve a bit to convey the positive/negative sign.
2) In the same vein, C++ never describe how a floating point number should be represented, IEEE-754 is often used, but this is a choice made by the implementation, in any case enforced by the standard: [basic.fundamental]/8 "The value representation of floating-point types is implementation-defined".
In addition to YSC's answer, let's take your sample code, and consider what it might aim to achieve
struct foo {
little int x; // little-endian
big long int y; // big-endian
short z; // native endianness
};
You might hope that this would exactly specify layout for architecture-independent data interchange (file, network, whatever)
But this can't possibly work, because several things are still unspecified:
data type size: you'd have to use little int32_t, big int64_t and int16_t respectively, if that's what you want
padding and alignment, which cannot be controlled strictly within the language: use #pragma or __attribute__((packed)) or some other compiler-specific extension
actual format (1s- or 2s-complement signedness, floating-point type layout, trap representations)
Alternatively, you might simply want to reflect the endianness of some specified hardware - but big and little don't cover all the possibilities here (just the two most common).
So, the proposal is incomplete (it doesn't distinguish all reasonable byte-ordering arrangements), ineffective (it doesn't achieve what it sets out to), and has additional drawbacks:
Performance
Changing the endianness of a variable from the native byte ordering should either disable arithmetic, comparisons etc (since the hardware cannot correctly perform them on this type), or must silently inject more code, creating natively-ordered temporaries to work on.
The argument here isn't that manually converting to/from native byte order is faster, it's that controlling it explicitly makes it easier to minimise the number of unnecessary conversions, and much easier to reason about how code will behave, than if the conversions are implicit.
Complexity
Everything overloaded or specialized for integer types now needs twice as many versions, to cope with the rare event that it gets passed a non-native-endianness value. Even if that's just a forwarding wrapper (with a couple of casts to translate to/from native ordering), it's still a lot of code for no discernible benefit.
The final argument against changing the language to support this is that you can easily do it in code. Changing the language syntax is a big deal, and doesn't offer any obvious benefit over something like a type wrapper:
// store T with reversed byte order
template <typename T>
class Reversed {
T val_;
static T reverse(T); // platform-specific implementation
public:
explicit Reversed(T t) : val_(reverse(t)) {}
Reversed(Reversed const &other) : val_(other.val_) {}
// assignment, move, arithmetic, comparison etc. etc.
operator T () const { return reverse(val_); }
};
Integers (as a mathematical concept) have the concept of positive and negative numbers. This abstract concept of sign has a number of different implementations in hardware.
Endianness is not a mathematical concept. Little-endian is a hardware implementation trick to improve the performance of multi-byte twos-complement integer arithmetic on a microprocessor with 16 or 32 bit registers and an 8-bit memory bus. Its creation required using the term big-endian to describe everything else that had the same byte-order in registers and in memory.
The C abstract machine includes the concept of signed and unsigned integers, without details -- without requiring twos-complement arithmetic, 8-bit bytes or how to store a binary number in memory.
PS: I agree that binary data compatibility on the net or in memory/storage is a PIA.
That's a good question and I have often thought something like this would be useful. However you need to remember that C aims for platform independence and endianness is only important when a structure like this is converted into some underlying memory layout. This conversion can happen when you cast a uint8_t buffer into an int for example. While an endianness modifier looks neat the programmer still needs to consider other platform differences such as int sizes and structure alignment and packing.
For defensive programming when you want find grain control over how some variables or structures are represented in a memory buffer then it is best to code explicit conversion functions and then let the compiler optimiser generate the most efficient code for each supported platform.
Endianness is not inherently a part of a data type but rather of its storage layout.
As such, it would not be really akin to signed/unsigned but rather more like bit field widths in structs. Similar to those, they could be used for defining binary APIs.
So you'd have something like
int ip : big 32;
which would define both storage layout and integer size, leaving it to the compiler to do the best job of matching use of the field to its access. It's not obvious to me what the allowed declarations should be.
Short Answer: if it should not be possible to use objects in arithmetic expressions (with no overloaded operators) involving ints, then these objects should not be integer types. And there is no point in allowing addition and multiplication of big-endian and little-endian ints in the same expression.
Longer Answer:
As someone mentioned, endianness is processor-specific. Which really means that this is how numbers are represented when they are used as numbers in the machine language (as addresses and as operands/results of arithmetic operations).
The same is "sort of" true of signage. But not to the same degree. Conversion from language-semantic signage to processor-accepted signage is something that needs to be done to use numbers as numbers. Conversion from big-endian to little-endian and reverse is something that needs to be done to use numbers as data (send them over the network or represent metadata about data sent over the network such as payload lengths).
Having said that, this decision appears to be mostly driven by use cases. The flip side is that there is a good pragmatic reason to ignore certain use cases. The pragmatism arises out of the fact that endianness conversion is more expensive than most arithmetic operations.
If a language had semantics for keeping numbers as little-endian, it would allow developers to shoot themselves in the foot by forcing little-endianness of numbers in a program which does a lot of arithmetic. If developed on a little-endian machine, this enforcing of endianness would be a no-op. But when ported to a big-endian machine, there would a lot of unexpected slowdowns. And if the variables in question were used both for arithmetic and as network data, it would make the code completely non-portable.
Not having these endian semantics or forcing them to be explicitly compiler-specific forces the developers to go through the mental step of thinking of the numbers as being "read" or "written" to/from the network format. This would make the code which converts back and forth between network and host byte order, in the middle of arithmetic operations, cumbersome and less likely to be the preferred way of writing by a lazy developer.
And since development is a human endeavor, making bad choices uncomfortable is a Good Thing(TM).
Edit: here's an example of how this can go badly:
Assume that little_endian_int32 and big_endian_int32 types are introduced. Then little_endian_int32(7) % big_endian_int32(5) is a constant expression. What is its result? Do the numbers get implicitly converted to the native format? If not, what is the type of the result? Worse yet, what is the value of the result (which in this case should probably be the same on every machine)?
Again, if multi-byte numbers are used as plain data, then char arrays are just as good. Even if they are "ports" (which are really lookup values into tables or their hashes), they are just sequences of bytes rather than integer types (on which one can do arithmetic).
Now if you limit the allowed arithmetic operations on explicitly-endian numbers to only those operations allowed for pointer types, then you might have a better case for predictability. Then myPort + 5 actually makes sense even if myPort is declared as something like little_endian_int16 on a big endian machine. Same for lastPortInRange - firstPortInRange + 1. If the arithmetic works as it does for pointer types, then this would do what you'd expect, but firstPort * 10000 would be illegal.
Then, of course, you get into the argument of whether the feature bloat is justified by any possible benefit.
From a pragmatic programmer perspective searching Stack Overflow, it's worth noting that the spirit of this question can be answered with a utility library. Boost has such a library:
http://www.boost.org/doc/libs/1_65_1/libs/endian/doc/index.html
The feature of the library most like the language feature under discussion is a set of arithmetic types such as big_int16_t.
Because nobody has proposed to add it to the standard, and/or because compiler implementer have never felt a need for it.
Maybe you could propose it to the committee. I do not think it is difficult to implement it in a compiler: compilers already propose fundamental types that are not fundamental types for the target machine.
The development of C++ is an affair of all C++ coders.
#Schimmel. Do not listen to people who justify the status quo! All the cited arguments to justify this absence are more than fragile. A student logician could find their inconsistence without knowing anything about computer science. Just propose it, and just don't care about pathological conservatives. (Advise: propose new types rather than a qualifier because the unsigned and signed keywords are considered mistakes).
Endianness is compiler specific as a result of being machine specific, not as a support mechanism for platform independence. The standard -- is an abstraction that has no regard for imposing rules that make things "easy" -- its task is to create similarity between compilers that allows the programmer to create "platform independence" for their code -- if they choose to do so.
Initially, there was a lot of competition between platforms for market share and also -- compilers were most often written as proprietary tools by microprocessor manufacturers and to support operating systems on specific hardware platforms. Intel was likely not very concerned about writing compilers that supported Motorola microprocessors.
C was -- after all -- invented by Bell Labs to rewrite Unix.
Is it possible to change the
float *pointer
type that is used in the VS c++ project
to some other type, so that it will still behave as a floating type but with less range?
I know that the floating point values never exceed some fixed value in that project, so I want to optimize the program by memory it uses. It doesn't need 4 bytes for each element of the 'float *pointer', 2 bytes will be enough I think. If I change a float to short and imitate the floating point behaviour, then it will use twice shorter memory. How to do it?
EDIT:
It calculates the probabilities. So there are divisions like
A / B
Where A < B,
And also B (and A) can be from 1 to 10 000.
There is a standard 16-bit floating point format described in IEEE 754-2008 called "binary16". It is specified as a format to store floating point values with reduced precisions. There is almost no compiler support for that yet (I think GCC supports it for certain ARM platforms), but it is quite easy to roll your own routines. This fellow:
http://blog.fpmurphy.com/2008/12/half-precision-floating-point-format_14.html
wrote a bit about it and also presents a routine to convert half-float <-> float.
Also, here seems to be a half-float C++ wrapper class:
half.h:
http://www.koders.com/cpp/fidABD00D95DE84C73BF0218AC621E400E07AA77B53.aspx
half.cpp
http://www.koders.com/cpp/fidF0DD0510FAAED03817A956D251787609BEB5989E.aspx
which supplies "HalfFloat" as a possible drop-in replacement type.
Maybe use fixed-point math? It all depends on value and precision you want to achieve.
http://www.eetimes.com/discussion/other/4024639/Fixed-point-math-in-C
For C there is a lot of code that makes fixed-point easy and I'm pretty sure there are also many C++ classes that make it even easier, but I don't know of any, I'm more into C.
The first, obvious, memory optimization would be to try and get rid of the pointer. If you can store just the float, that may, depending on the larger context, reduce your memory consumption from eight to four bytes already. (On a 64-Bit system, from twelve to four.)
Whether you can get by with a short depends on what your program does with the values. You may be able to use fix point arithmetic using an integral type such as a short, yes but your questions shows way too little context to judge that.
The code you posted and the text in the question do not deal with actual float, but with pointers to float. In all architectures I know of, the size of a pointer is the same regardless of the pointed type, so there would be no improvement in changing that to a short or char pointer.
Now, about the actual pointed elements, what is the range that you expect in your application? What is the precision you need? How many of those elements do you have? What are the memory constraints of your target platform? Unless the range and precision are small and the number of elements huge, just use floats. Also note that if you need floating point operations, storing any other type will require conversions before and after each operation, and you might be impacting performance.
Without greater knowledge of what you are doing, the ranges for short in many architectures are [-32k, 32k), where k stands for 1024. If your data ranges is [-32,32) and you can do with roughly 3 decimal digits you could use fixed point arithmetic with shorts, but there are few such situation.