Thread overview
D on lm32-CPU: string argument on stack instead of register
Jul 31
Chad Joan
Aug 01
Chad Joan
Jul 31
Johan
July 31
Hi all,

at work we put embedded lm32 soft-core CPUs in FPGAs and write the firmware in C.
At home I enjoy writing small projects in D from time to time, but I don't consider myself a D expert.

Now, I'm trying to run some toy examples in D on the lm32 cpu. I'm using a recent gcc-elf-lm32. I succeeded in compiling and running some code and it works fine.

But I noticed, when calling a function with a string argument, the string is not stored in registers, but on the stack.
Consider a simple function (below) that writes bytes to a peripheral (that forwards the data to the host computer via USB). I've two versions, an ideomatic D one, and another version where pointer and length are two distinct function parameters.
I also show the generated assembly code. The string version is 4 instructions longer, just because of the stack manipulation. In addition, it is also slower because it need to access the ram, and it needs more stack space.

My question: Is there a way I can tell the D compiler to use registers instead of stack for string arguments, or any other trick to reduce code size while maintaining an ideomatic D codestyle?

Best regards
Michael


// ideomatic D version
void write_to_host(in string msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	foreach(ch; msg) {
		*usb_slave = ch;
	}
}
// resulting assembly code (compiled with -Os) 12 instructions
_D10firmware_d13write_to_hostFxAyaZv:
	addi     sp, sp, -8
	addi     r3, r0, 4096
	sw       (sp+4), r1
	sw       (sp+8), r2
	add      r1, r2, r1
.L3:
	be     r2,r1,.L1
	lbu      r4, (r2+0)
	addi     r2, r2, 1
	sb       (r3+0), r4
	bi       .L3
.L1:
	addi     sp, sp, 8
	b        ra

// C-like version
void write_to_hostC(const char *msg, int len) {
	char *ptr = cast(char*)msg;
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	while (len--) {
		*usb_slave = *ptr++;
	}
}
// resulting assembly code (compiled with -Os) 8 instructions
_D10firmware_d14write_to_hostCFxPaiZv:
	add      r2, r1, r2
	addi     r3, r0, 4096
.L7:
	be     r1,r2,.L5
	lbu      r4, (r1+0)
	addi     r1, r1, 1
	sb       (r3+0), r4
	bi       .L7
.L5:
	b        ra


July 31
On Friday, 31 July 2020 at 10:22:20 UTC, Michael Reese wrote:
> Hi all,
>
> at work we put embedded lm32 soft-core CPUs in FPGAs and write the firmware in C.
> At home I enjoy writing small projects in D from time to time, but I don't consider myself a D expert.
>
> Now, I'm trying to run some toy examples in D on the lm32 cpu. I'm using a recent gcc-elf-lm32. I succeeded in compiling and running some code and it works fine.
>
> But I noticed, when calling a function with a string argument, the string is not stored in registers, but on the stack.
> Consider a simple function (below) that writes bytes to a peripheral (that forwards the data to the host computer via USB). I've two versions, an ideomatic D one, and another version where pointer and length are two distinct function parameters.
> I also show the generated assembly code. The string version is 4 instructions longer, just because of the stack manipulation. In addition, it is also slower because it need to access the ram, and it needs more stack space.
>
> My question: Is there a way I can tell the D compiler to use registers instead of stack for string arguments, or any other trick to reduce code size while maintaining an ideomatic D codestyle?
>
> Best regards
> Michael
>
>
> // ideomatic D version
> void write_to_host(in string msg) {
> 	// a fixed address to get bytes to the host via usb
> 	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
> 	foreach(ch; msg) {
> 		*usb_slave = ch;
> 	}
> }
> // resulting assembly code (compiled with -Os) 12 instructions
> _D10firmware_d13write_to_hostFxAyaZv:
> 	addi     sp, sp, -8
> 	addi     r3, r0, 4096
> 	sw       (sp+4), r1
> 	sw       (sp+8), r2
> 	add      r1, r2, r1
> .L3:
> 	be     r2,r1,.L1
> 	lbu      r4, (r2+0)
> 	addi     r2, r2, 1
> 	sb       (r3+0), r4
> 	bi       .L3
> .L1:
> 	addi     sp, sp, 8
> 	b        ra
>
> // C-like version
> void write_to_hostC(const char *msg, int len) {
> 	char *ptr = cast(char*)msg;
> 	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
> 	while (len--) {
> 		*usb_slave = *ptr++;
> 	}
> }
> // resulting assembly code (compiled with -Os) 8 instructions
> _D10firmware_d14write_to_hostCFxPaiZv:
> 	add      r2, r1, r2
> 	addi     r3, r0, 4096
> .L7:
> 	be     r1,r2,.L5
> 	lbu      r4, (r1+0)
> 	addi     r1, r1, 1
> 	sb       (r3+0), r4
> 	bi       .L7
> .L5:
> 	b        ra

Hi Michael!

Last time I checked, D doesn't have any specific type attributes or special ways to force variables to enregister. But I could be poorly informed. Maybe there are GDC-specific hints or something. I hope that if anyone else knows better, they will toss in an answer.

THAT SAID, I think there are things to try and I hope we can get you what you want.

If you're willing to entertain more experimentation, here are my thoughts:

---------------------------------------
(1) Try writing "in string" as "in const(char)[]" instead:

// ideomatic D version
void write_to_host(in const(char)[] msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	foreach(ch; msg) {
		*usb_slave = ch;
	}
}

Explanation:

The "string" type is an alias for "immutable(char)[]".

In D, "immutable" is a stronger guarantee than "const". The "const" modifier, like in C, tells the compiler that this function shall not modify the data referenced by this pointer/array/whatever. The "immutable" modifier is a bit different, as it says that NO ONE will modify the data referenced by this pointer/array/whatever, including other functions that may or may not be concurrently executing alongside the one you're in. So "const" constraints the callee, while "immutable" constrains both the callee AND the caller. This makes it more useful for some multithreaded code, because if you can accept the potential inefficiency of needing to do more copying of data (if you can't modify, usually you must copy instead), then you can have more deterministic behavior and sometimes even much better total efficiency by way of parallelization. This might not be a guarantee you care about though, at which point you can just toss it out completely and see if the compiler generates better code now that it sees the same type qualifier as in the other example.

I'd actually be surprised if using "immutable" causes /less/ efficient code in this case, because it should be even /safer/ to use the argument as-is. But it IS a difference between the two examples, and one that might not be benefiting your cause (though that's totally up to you).

---------------------------------------
(2) Try keeping the string argument, but make the function more closely identical in semantics:

// ideomatic D version
void write_to_host(string msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	while(msg.length > 0) {
		*usb_slave = msg[0];
		msg = msg[1 .. $];
	}
}

Explanation:

First of all, I wouldn't expect you to keep this, especially if you need utf-8 autodecoding behavior (more on that later). But it might be revealing if this leads to different assembly output.

The idea behind this one is to see if the regression is actually caused by the foreach construct, rather than the parameter type. I did have to change the parameter slightly by removing the "in" qualifier. It shouldn't make much difference though, because the 'string' type's pointer and length are copied from the caller, so any modifications to "msg" (that don't affect "msg"'s array elements) will be contained within the function and will not be observable anywhere else. In other words, the "in" qualifier is largely redundant with "string"'s immutability guarantees plus function argument copying semantics.

---------------------------------------
(3) Try a different type of while-loop in the D-style version:

// ideomatic D version
void write_to_host(in string msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	size_t i = 0;
	while(i < msg.length) {
		*usb_slave = msg[i++];
	}
}

Explanation:

This is a variant of #2. It does ask for an extra size_t variable, so I don't have high hopes. But the compiler might optimize that out and make it look like the C-style version. Again, I don't expect you to use this version if it discards one of D's features that you hope to use, but it might at least help you identify where your expenses are coming from.

---------------------------------------
(4) Try having these examples use "const ubyte* msg" and "immutable(ubyte)[] msg" instead of "const char* msg" and "string msg".

// ideomatic D version
void write_to_host(in immutable(ubyte)[] msg) {
	// a fixed address to get bytes to the host via usb
	ubyte *usb_slave = cast(ubyte*)BaseAdr.ft232_slave;
	foreach(ch; msg) {
		*usb_slave = ch;
	}
}

// C-like version
void write_to_hostC(const ubyte *msg, int len) {
	ubyte *ptr = cast(ubyte*)msg;
	ubyte *usb_slave = cast(ubyte*)BaseAdr.ft232_slave;
	while (len--) {
		*usb_slave = *ptr++;
	}
}

Explanation:

The "string" type is an alias for "immutable(char)[]", which seems like it would be very similar to "immutable(ubyte)[]", but the 'char' element type communicates a requirement that the 'ubyte' element type does not: utf-8 awareness. And that can have a cost.

In D, char[] arrays are defined as containing utf-8 text. This is rather different from C, where the 'char' type is more like D's 'byte' or 'ubyte' types and just happens to also be used to store text data in any encoding the author feels like. When I see "foreach(ch; msg)" and msg's element type is "char", then I expect "ch" to be of type 'dchar' (instead of 'char') and I expect the foreach loop to auto-decode the utf-8 text in the string (or immutable(char)[]) type into whole unicode codepoints that are then placed into the 'dchar'. If you are only dealing with ASCII text (or any 8-bit-or-less encoding that isn't utf-8), then you may just want to use the 'byte' or 'ubyte' types instead. In everyday D, this changes the semantics of the foreach loop, because no autodecoding is done on types like byte[] or ubyte[], and it may "behave" (from an implementor perspective) more like the while-loop in your second example.

You probably won't see a lot of text-processing through byte[] or ubyte[] in normal D code, but that's because most programmers will want their programs to be able to process utf-8 text, while in the embedded programming space you might not have to worry about utf-8 at all.

Now, I actually didn't see any autodecoding of utf-8 in the assembly you posted. Maybe I could be wrong though; I am not experienced in lm32 assembly. Nonetheless, I'd expect to seem some sort of conditional call or, at the very least, some kind of masking of the highest bit of every char (to detect utf-8 sequences). Maybe it's a bug in your (cross?) compiler, or even just an intentional configuration choice that I didn't expect. At any rate, I don't think your code is larger or less efficient due to utf-8 decoding, because I don't see the utf-8 decoding.

Still, I'm curious to see if changing up the types causes the compiler to choose different codepaths for its codegen, even for inane reasons. Maybe the autodecoding is turned off, but it still thinks it needs to allocate extra space for the autodecoder's "dchar" or something, and then that exceeds some threshold for passing enregistered arguments. Maybe for similar reasons it thinks it needs to keep a copy of that string around. Compilers are mysterious beasts sometimes. *shrug*

---------------------------------------
(5) And for maximum curiousity, what happens if you write the C-like version this way instead?

// C-like version
// msg parameter change: "const char *msg" -> "const(char)* msg"
void write_to_hostC(const(char)* msg, int len) {
	// cast() statement removed.
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	while (len--) {
		*usb_slave = *msg++;
	}
}

Explanation:

I realize the difference is subtle, but "const char *msg" says that both the pointed-to chars can't be modified and also that the /pointer itself/ cannot be modified. In the other case, with "const(char)* msg", the constraint is looser but still very useful: the pointed-to chars can't be modified, but the pointer can be modified. Because the pointer (but not the referred data) is a copy of the caller's pointer, any modifications to the pointer (increments and such) are only visible within the scope of this function.

The C-like version is already the more optimal one, but if making this change causes it to regress to generating assembly similar to the D-like version, then it might suggest that the additional assignment statement is actually helpful somehow. It'd be unintuitive, but you never know.

---------------------------------------
(6) OK sorry, one more. Because #5 made me think: what if we extended the D-idiomatic-version's immutability guarantee to the whole array value and not just the array elements?

// ideomatic D version
void write_to_host(immutable(char[]) msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	foreach(ch; msg) {
		*usb_slave = ch;
	}
}

And to make it even more like the C-style version without being C-style, it might also be worth stacking it with the immutable->const change:

// ideomatic D version
void write_to_host(const(char[]) msg) {
	// a fixed address to get bytes to the host via usb
	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
	foreach(ch; msg) {
		*usb_slave = ch;
	}
}

After all, if the original C-style version isn't allowed to change its argument's pointer, then we could try making the D-idiomatic version behave that way too, and see if this minor alteration makes the difference.

---------------------------------------

Just to be safe, I also want to point out the difference between "char *ptr" and "char* ptr": in a single-variable declaration, there is none, but if there is more than one, the pointer binds more strongly to the type than to the variable in D.

Consider a declaration like:

char* str0, str1;

In C, this would make str0 a pointer, and str1 a char.
In D, this means that both str0 and str1 are pointers.

Thus, in D, it is more conventional to write the * character next to the type than it is to write it next to the variable/identifier. This reinforces the notion that pointer-ness is (syntactically) part of the type, rather than part of the variables.

There's a similar example in this article:
https://dlang.org/blog/2018/10/17/interfacing-d-with-c-arrays-part-1/

If you already knew that, don't mind me. I realize that a lot of C code gets copied into D without changing this thing, and unless there are multiple variables in the same declaration, it really doesn't matter.



Good luck with your lm32/FPGA coding. That sounds like cool stuff!
July 31
On Friday, 31 July 2020 at 10:22:20 UTC, Michael Reese wrote:
>
> My question: Is there a way I can tell the D compiler to use registers instead of stack for string arguments, or any other trick to reduce code size while maintaining an ideomatic D codestyle?

A D string is a "slice", which is a struct (pointer + length). Depending on the function call ABI, structs are passed in registers or on the stack.
On x86, the D calling convention is to put small POD structs in registers, similar to the C++ calling convention.
I don't know whether GDC has attributes to change the calling convention of functions (besides extern(C/C++/Windows/etc.)), but that's where you'd need to look.
Otherwise, file a bug with GDC. Slice arguments are common enough to require enregistering them in the D calling convention on your lm32 platform.

-Johan

August 01
On Friday, 31 July 2020 at 15:13:29 UTC, Chad Joan wrote:
> THAT SAID, I think there are things to try and I hope we can get you what you want.
>
> If you're willing to entertain more experimentation, here are my thoughts:

Thanks a lot for the suggestions and explanations. I tried them all but the only time I got different assembly code was your suggestion (3) which produced the following.

> (3) Try a different type of while-loop in the D-style version:
>
> // ideomatic D version
> void write_to_host(in string msg) {
> 	// a fixed address to get bytes to the host via usb
> 	char *usb_slave = cast(char*)BaseAdr.ft232_slave;
> 	size_t i = 0;
> 	while(i < msg.length) {
> 		*usb_slave = msg[i++];
> 	}
> }
_D10firmware_d13write_to_hostFAyaZv:
	addi     sp, sp, -8
	addi     r4, r0, 4096
	sw       (sp+8), r2
	sw       (sp+4), r1
	addi     r2, r0, 0
.L3:
	or       r5, r2, r0
	be     r2,r1,.L1
	lw       r3, (sp+8)
	addi     r2, r2, 1
	add      r3, r3, r5
	lbu      r3, (r3+0)
	sb       (r4+0), r3
	bi       .L3
.L1:
	addi     sp, sp, 8
	b        ra

> At any rate, I don't think your code is larger or less efficient due to utf-8 decoding, because I don't see the utf-8 decoding.

Agreed, I think there's no code for autodecoding being generated.

I did some more experiments:
Trying to put pointer and length into a struct and pass that to the function. Same result, the argument ended up on the stack.

Then, I wrote the function in C and compiled it with the C-compiler of lm32-elf-gcc. It also puts the 64-bit POD structure on the stack. It seems the only arguments passed in registers are primitive data types. However, if I pass a uint64_t argument, it is registered using registers r1 and r2. So the compiler knows how to use r1 and r2 for arguments. I checked again in the lm32 manual (https://www.latticesemi.com/view_document?document_id=52077), and it says:

"As illustrated in Table 3 on page 8, the first eight function arguments are
passed in registers. Any remaining arguments are passed on the stack, as
illustrated in Figure 12."

So strings and structs should be passed on the stack and this seems to be more an issue of the gcc lm32 backend than a D issue.

But I just found a workaround using a wrapper function.

void write_to_host(in string msg) {
        write_to_hostC(msg.ptr, msg.length);
}

I checked the assembly code on the caller side, and the call write_to host("Hello, World!\n") is inlined. There is only one call to write_to_hostC. This is still not nice, but I think I can live with that for now. Now I have to figure out how make the cast back from from pointer-length pair into a string. I'm sure I read that somewhere before, but forgot it and was unable to find it now on a quick google search...
And since this is D: is there maybe some CTFE magic that allows to create these wrappers automatically? Somthing like

fix_stack!write_to_host("Hello, World!\n");


> Good luck with your lm32/FPGA coding. That sounds like cool stuff!

I'm doing this mainly to improve my understanding of how embedded processors work, and how to write linker scripts for a given environment. Although I do have actual hardware where I can see if everything runs in the real world, I mainly use simulations. The coolest thing in my opinion is, nowadays it can be done using only open source tools (mainly ghdl and verilator, the lm32 source code is open source, too). The complete system is simulated and you can look at every logic signal in the cpu or in the ram while the program executes.
August 01
On Saturday, 1 August 2020 at 08:58:03 UTC, Michael Reese wrote:
> [...]
> So the compiler knows how to use r1 and r2 for arguments. I checked again in the lm32 manual (https://www.latticesemi.com/view_document?document_id=52077), and it says:
>
> "As illustrated in Table 3 on page 8, the first eight function arguments are
> passed in registers. Any remaining arguments are passed on the stack, as
> illustrated in Figure 12."
>
> So strings and structs should be passed on the stack and this seems to be more an issue of the gcc lm32 backend than a D issue.
>

Nice find!

Though if the compiler is allowed to split a single uint64_t into two registers, I would expect it to split struct/string into two registers as well. At least, the manual doesn't seem to explicitly mention higher-level constructs like structs. It does suggest a one-to-one relationship between arguments and registers (up to a point), but GCC seems to have decided otherwise for certain uint64_t's. (Looking at Table 3...) It even gives you two registers for a return value: enough for a string or an array. And if the backend/ABI weren't up for it, it would be theoretically possible to have the frontend to lower strings (dynamic arrays) and small structs into their components before function calls and then also insert code on the other side to cast them back into their original form. I'm not sure if anyone would want to write it, though. o.O

> But I just found a workaround using a wrapper function.
>
> void write_to_host(in string msg) {
>         write_to_hostC(msg.ptr, msg.length);
> }
>
> I checked the assembly code on the caller side, and the call write_to host("Hello, World!\n") is inlined. There is only one call to write_to_hostC. This is still not nice, but I think I can live with that for now. Now I have to figure out how make the cast back from from pointer-length pair into a string. I'm sure I read that somewhere before, but forgot it and was unable to find it now on a quick google search...

That's pretty clever. I like it.

Getting from pointer-length to string might be pretty easy:
string foo = ptr[0 .. len];

D allows pointers to be indexed, like in C. But unlike C, D has slices, and pointers can be "sliced". The result of a slice operation, at least for primitive arrays+pointers, is always an array (the "dynamic" ptr+length kind). Hope that helps.


> And since this is D: is there maybe some CTFE magic that allows to create these wrappers automatically? Somthing like
>
> fix_stack!write_to_host("Hello, World!\n");
>

It ended up being a little more complicated than I thought it would be. Hope I didn't ruin the fun. ;)

https://pastebin.com/y6e9mxre


Also, that part where you mentioned a 64-bit integer being passed as a pair of registers made me start to wonder if unions could be (ab)used to juke the ABI:

https://pastebin.com/eGfZN0SL


>
>> Good luck with your lm32/FPGA coding. That sounds like cool stuff!
>
> I'm doing this mainly to improve my understanding of how embedded processors work, and how to write linker scripts for a given environment. Although I do have actual hardware where I can see if everything runs in the real world, I mainly use simulations. The coolest thing in my opinion is, nowadays it can be done using only open source tools (mainly ghdl and verilator, the lm32 source code is open source, too). The complete system is simulated and you can look at every logic signal in the cpu or in the ram while the program executes.

Thanks for the insights; I've done just a little hobby electrical stuff here-and-there, and having some frame of reference for tool and component choice makes me feel good, even if I don't plan on buying any lm32s or FPGAs anytime soon :)  Maybe I can Google some of that later and geek out at images of other people's debugging sessions or something. I'm curious how they manage the complexity that happens when circuits and massive swarms of logic gates do their, uh, complexity thing. o.O


August 04
On Saturday, 1 August 2020 at 23:08:38 UTC, Chad Joan wrote:
> Though if the compiler is allowed to split a single uint64_t into two registers, I would expect it to split struct/string into two registers as well. At least, the manual doesn't seem to explicitly mention higher-level constructs like structs. It does suggest a one-to-one relationship between arguments and registers (up to a point), but GCC seems to have decided otherwise for certain uint64_t's. (Looking at Table 3...) It even gives you two registers for a return value: enough for a string or an array. And if the backend/ABI weren't up for it, it would be theoretically possible to have the frontend to lower strings (dynamic arrays) and small structs into their components before function calls and then also insert code on the other side to cast them back into their original form. I'm not sure if anyone would want to write it, though. o.O

Right, I think at some point one should fix the backend. C programs would also benefit from it when passing structs as arguments. However in C it is more common to just pass pointers and they go into registers. I guess this is why I never noticed before that struct passing is needlessly expensive.

> Getting from pointer-length to string might be pretty easy:
> string foo = ptr[0 .. len];

Ah cool! I did know about array slicing, but wasn't aware that it works on pointers, too.

> It ended up being a little more complicated than I thought it would be. Hope I didn't ruin the fun. ;)
>
> https://pastebin.com/y6e9mxre

Thanks :) I'll have to look into that more closely. But this is the kind of stuff that I hope to make use of in the future on the embedded CPU. But for now I cannot use it yet because I don't have phobos and druntime in my toolchain right now... just naked D.

> Also, that part where you mentioned a 64-bit integer being passed as a pair of registers made me start to wonder if unions could be (ab)used to juke the ABI:
>
> https://pastebin.com/eGfZN0SL

Thanks for suggesting! I tried, and the union works as well, i.e. the function args are registered. But I noticed another thing about all workarounds so far:
Even if calls are inlined and arguments end up on the stack, the linker puts code of the wrapper function in my final binary event if it is never explicitly called. So until I find a way to strip of uncalled functions from the binary (not sure the linker can do it), the workarounds don't solve the size problem. But they still make the code run faster.

August 05
On Tuesday, 4 August 2020 at 17:36:53 UTC, Michael Reese wrote:
> Thanks for suggesting! I tried, and the union works as well, i.e. the function args are registered. But I noticed another thing about all workarounds so far:
> Even if calls are inlined and arguments end up on the stack, the linker puts code of the wrapper function in my final binary event if it is never explicitly called. So until I find a way to strip of uncalled functions from the binary (not sure the linker can do it), the workarounds don't solve the size problem. But they still make the code run faster.

Try -ffunction-sections -Wl,--gc-sections. That should remove all unreferenced functions. It removes all unreferenced sections, and writes every function into a separate section.