VLA in Assembler
Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;)
Re: VLA in Assembler
Foo: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) Doing it with alloca is simpler: void main() @nogc { import core.stdc.stdlib: alloca, exit; alias T = int; enum n = 42; auto ptr = cast(T*)alloca(T.sizeof * n); if (ptr == null) exit(1); // Or throw a memory error. auto arr = ptr[0 .. n]; } Bye, bearophile
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 10:59:09 UTC, bearophile wrote: Foo: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) Doing it with alloca is simpler: void main() @nogc { import core.stdc.stdlib: alloca, exit; alias T = int; enum n = 42; auto ptr = cast(T*)alloca(T.sizeof * n); if (ptr == null) exit(1); // Or throw a memory error. auto arr = ptr[0 .. n]; } Bye, bearophile Yes I know, but I really want it in inline assembly. It's for learning purpose. :)
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 11:39:43 UTC, Foo wrote: On Wednesday, 17 December 2014 at 10:59:09 UTC, bearophile wrote: Foo: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) Doing it with alloca is simpler: void main() @nogc { import core.stdc.stdlib: alloca, exit; alias T = int; enum n = 42; auto ptr = cast(T*)alloca(T.sizeof * n); if (ptr == null) exit(1); // Or throw a memory error. auto arr = ptr[0 .. n]; } Bye, bearophile Yes I know, but I really want it in inline assembly. It's for learning purpose. :) You could look at the disassembly.
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 12:15:23 UTC, uri wrote: On Wednesday, 17 December 2014 at 11:39:43 UTC, Foo wrote: On Wednesday, 17 December 2014 at 10:59:09 UTC, bearophile wrote: Foo: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) Doing it with alloca is simpler: void main() @nogc { import core.stdc.stdlib: alloca, exit; alias T = int; enum n = 42; auto ptr = cast(T*)alloca(T.sizeof * n); if (ptr == null) exit(1); // Or throw a memory error. auto arr = ptr[0 .. n]; } Bye, bearophile Yes I know, but I really want it in inline assembly. It's for learning purpose. :) You could look at the disassembly. And how? I'm on Windows.
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 10:35:39 UTC, Foo wrote: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) It's probably something like that: module runnable; import std.stdio; import std.c.stdlib; ubyte[] newArr(size_t aLength) { asm { naked; mov ECX, EAX; // saves aLength in ECX push ECX; call malloc;// .ptr = malloc(aLength); mov ECX,[EAX]; // saved the .ptr of our array mov EAX, 8; // an array is a struct with length and ptr // so 8 bytes in 32 bit call malloc;// EAX points to the first byte of the struct mov [EAX + 4], ECX; // .ptr pop ECX; mov [EAX], ECX; // .length mov EAX, [EAX]; // curretnly EAX is a ref, so need to dig... ret; } } try and see ;) Actually it may be wrong
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 12:54:44 UTC, btdc wrote: On Wednesday, 17 December 2014 at 10:35:39 UTC, Foo wrote: Hi, Could someone explain me, if and how it is possible to allocate a variable length array with inline assembly? Somewhat like int[] arr; int n = 42; asm { // allocate n stack space for arr } I know it is dangerous and all that, but I just want it know. ;) It's probably something like that: module runnable; import std.stdio; import std.c.stdlib; ubyte[] newArr(size_t aLength) { asm { naked; mov ECX, EAX; // saves aLength in ECX push ECX; call malloc;// .ptr = malloc(aLength); mov ECX,[EAX]; // saved the .ptr of our array mov EAX, 8; // an array is a struct with length and ptr // so 8 bytes in 32 bit call malloc;// EAX points to the first byte of the struct mov [EAX + 4], ECX; // .ptr pop ECX; mov [EAX], ECX; // .length mov EAX, [EAX]; // curretnly EAX is a ref, so need to dig... ret; } } try and see ;) Actually it may be wrong fuck...the comments are once again cut...
Re: VLA in Assembler
And it is using malloc... ;) I wanted something that increases the stack pointer ESP. e.g. void main() { int[] arr; int n = 42; writeln(arr.length); writeln(arr.ptr); asm { mov EAX, n; mov [arr + 8], ESP; sub [ESP], EAX; mov [arr + 0], EAX; } writeln(arr.length); //writeln(arr[0]); } but that does not work...
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 12:29:53 UTC, Foo wrote: And how? I'm on Windows. Digital Mars sells an obj2asm function that will disassemble dmd generated code. I think it is in the $15 basic utility package. But VLA/alloca is more complex than a regular function - the compiler needs to know about it to adjust for the changed stack. It'll take more length to write this up, I'll do it in a separate post.
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 14:11:32 UTC, Foo wrote: And it is using malloc... ;) I wanted something that increases the stack pointer ESP. e.g. void main() { int[] arr; int n = 42; writeln(arr.length); writeln(arr.ptr); asm { mov EAX, n; mov [arr + 8], ESP; sub [ESP], EAX; mov [arr + 0], EAX; } writeln(arr.length); //writeln(arr[0]); } but that does not work... You cant always get what you want. try more, speak less.
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 14:11:32 UTC, Foo wrote: asm { mov EAX, n; mov [arr + 8], ESP; sub [ESP], EAX; mov [arr + 0], EAX; } but that does not work... That wouldn't work even with malloc remember, an integer more than one byte long, so your subtract is 1/4 the size it needs to be! Also, since the stack grows downward, you're storing the pointer to the end of the array instead of the beginning of it. NOTE: I've never actually done this before, so I'm figuring it out as I go too. This might be buggy or otherwise mistaken at points. (Personally, I prefer to use a static array sized to the max thing I'll probably need that I slice instead of alloca...) Here's some code that runs successfully (in 32 bit!): void vla(int n) { int[] arr; asm { mov EAX, [n]; // the first word in an array is the length, store that mov [arr], EAX; shl EAX, 2; // number of bytes == n * int.sizeof sub ESP, EAX; // allocate the bytes mov [arr + size_t.sizeof], ESP; // store the beginning of it in the arr.ptr } import std.stdio; writeln(arr.length); writeln(arr.ptr); // initialize the data... foreach(i, ref a; arr) a = i; writeln(arr); // and print it back out } void main() { vla(8); } This looks right but isn't, we changed the stack and didn't put it back. That's usually a no-no. If we disassemble the function, we can take a look at the end and see something scary: 8084ec6: e8 9d 6a 00 00 call 808b968 _D3std5stdio15__T7writelnTAiZ7writelnFAiZv // our final writeln call 8084ecb: 5e popesi // uh oh 8084ecc: 5b popebx 8084ecd: c9 leave 8084ece: c3 ret Before the call to leave, which puts the stack back how it was at the beginning of the function - which saves us from a random EIP being restored upon the ret instruction - the compiler put in a few pop instructions. main() will have different values in esi and ebx than it expects! Running it in the debugger shows these values changed too: before (gdb) info registers [...] ebx0xd4f4 -11020 [...] esi0x80916e8134813416 after ebx0x1 1 esi0x0 0 It popped the values of our array. According to the ABI: EBX, ESI, EDI, EBP must be preserved across function calls. http://dlang.org/abi.html They are pushed for a reason - the compiler assumes they remain the same. In this little test program, nothing went wrong because no more code was run after vla returned. But, if we were using, say a struct, it'd probably fault when it tried to access `this`. It'd probably mess up other local variables too. No good! So, we'll need to store and restore the stack pointer... can we use the stack's push and pop instructions? Nope, we're changing the stack! Our own pop would grab the wrong data too. We could save it in a local variable. How do we restore it though? scope(exit) won't work, it won't happen at the right time and will corrupt the stack even worse. Gotta do it ourselves - which means we can't do the alloca even as a single mixin, since it needs code added before any return point too! (There might be other, better ways to do this... and indeed, there is, as we'll see later on. I peeked at the druntime source code and it does it differently. Continue reading...) Here's code that we can verify in the debugger leaves everything how it should be and doesn't crash: void vla(int n) { int[] arr; void* saved_esp; asm { mov EAX, [n]; mov [arr], EAX; shl EAX, 2; // number of bytes == n * int.sizeof // NEW LINE mov [saved_esp], ESP; // save it for later sub ESP, EAX; mov [arr + size_t.sizeof], ESP; } import std.stdio; writeln(arr.length); writeln(arr.ptr); foreach(i, ref a; arr) a = i; writeln(arr); // NEW LINE asm { mov ESP, [saved_esp]; } // restore it before we return } Note that this still isn't quite right - the allocated size should be aligned too. It works for the simple case of 8 ints since that's coincidentally aligned, but if we were doing like 3 bytes, it would mess things up. Gotta be rounded up to a multiple of 4 or 16 on some systems. hmm, I'm looking at the alloca source and there's a touch of a guard page on Windows too. Check out the file: dmd2/src/druntime/src/rt/alloca.d, it is written in mostly inline asm. Note the comment though: * This is a 'magic' function that needs help from the compiler to
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 15:20:28 UTC, btdc wrote: On Wednesday, 17 December 2014 at 14:11:32 UTC, Foo wrote: And it is using malloc... ;) I wanted something that increases the stack pointer ESP. e.g. void main() { int[] arr; int n = 42; writeln(arr.length); writeln(arr.ptr); asm { mov EAX, n; mov [arr + 8], ESP; sub [ESP], EAX; mov [arr + 0], EAX; } writeln(arr.length); //writeln(arr[0]); } but that does not work... You cant always get what you want. try more, speak less. Very helpful. And soo friendly! ;)
Re: VLA in Assembler
On Wednesday, 17 December 2014 at 16:10:40 UTC, Adam D. Ruppe wrote: On Wednesday, 17 December 2014 at 14:11:32 UTC, Foo wrote: asm { mov EAX, n; mov [arr + 8], ESP; sub [ESP], EAX; mov [arr + 0], EAX; } but that does not work... That wouldn't work even with malloc remember, an integer more than one byte long, so your subtract is 1/4 the size it needs to be! Also, since the stack grows downward, you're storing the pointer to the end of the array instead of the beginning of it. NOTE: I've never actually done this before, so I'm figuring it out as I go too. This might be buggy or otherwise mistaken at points. (Personally, I prefer to use a static array sized to the max thing I'll probably need that I slice instead of alloca...) Here's some code that runs successfully (in 32 bit!): void vla(int n) { int[] arr; asm { mov EAX, [n]; // the first word in an array is the length, store that mov [arr], EAX; shl EAX, 2; // number of bytes == n * int.sizeof sub ESP, EAX; // allocate the bytes mov [arr + size_t.sizeof], ESP; // store the beginning of it in the arr.ptr } import std.stdio; writeln(arr.length); writeln(arr.ptr); // initialize the data... foreach(i, ref a; arr) a = i; writeln(arr); // and print it back out } void main() { vla(8); } This looks right but isn't, we changed the stack and didn't put it back. That's usually a no-no. If we disassemble the function, we can take a look at the end and see something scary: 8084ec6: e8 9d 6a 00 00 call 808b968 _D3std5stdio15__T7writelnTAiZ7writelnFAiZv // our final writeln call 8084ecb: 5e popesi // uh oh 8084ecc: 5b popebx 8084ecd: c9 leave 8084ece: c3 ret Before the call to leave, which puts the stack back how it was at the beginning of the function - which saves us from a random EIP being restored upon the ret instruction - the compiler put in a few pop instructions. main() will have different values in esi and ebx than it expects! Running it in the debugger shows these values changed too: before (gdb) info registers [...] ebx0xd4f4 -11020 [...] esi0x80916e8134813416 after ebx0x1 1 esi0x0 0 It popped the values of our array. According to the ABI: EBX, ESI, EDI, EBP must be preserved across function calls. http://dlang.org/abi.html They are pushed for a reason - the compiler assumes they remain the same. In this little test program, nothing went wrong because no more code was run after vla returned. But, if we were using, say a struct, it'd probably fault when it tried to access `this`. It'd probably mess up other local variables too. No good! So, we'll need to store and restore the stack pointer... can we use the stack's push and pop instructions? Nope, we're changing the stack! Our own pop would grab the wrong data too. We could save it in a local variable. How do we restore it though? scope(exit) won't work, it won't happen at the right time and will corrupt the stack even worse. Gotta do it ourselves - which means we can't do the alloca even as a single mixin, since it needs code added before any return point too! (There might be other, better ways to do this... and indeed, there is, as we'll see later on. I peeked at the druntime source code and it does it differently. Continue reading...) Here's code that we can verify in the debugger leaves everything how it should be and doesn't crash: void vla(int n) { int[] arr; void* saved_esp; asm { mov EAX, [n]; mov [arr], EAX; shl EAX, 2; // number of bytes == n * int.sizeof // NEW LINE mov [saved_esp], ESP; // save it for later sub ESP, EAX; mov [arr + size_t.sizeof], ESP; } import std.stdio; writeln(arr.length); writeln(arr.ptr); foreach(i, ref a; arr) a = i; writeln(arr); // NEW LINE asm { mov ESP, [saved_esp]; } // restore it before we return } Note that this still isn't quite right - the allocated size should be aligned too. It works for the simple case of 8 ints since that's coincidentally aligned, but if we were doing like 3 bytes, it would mess things up. Gotta be rounded up to a multiple of 4 or 16 on some systems. hmm, I'm looking at the alloca source and there's a touch of a guard page on Windows too. Check out the file: dmd2/src/druntime/src/rt/alloca.d, it is written in mostly inline asm. Note the comment