I'm generating LLVM IR for JIT purposes, and I notice that LLVM's calling conventions don't seem to match the C calling conventions when aggregate values are involved. For instance, when I declare a function as taking a {i32, i32} (that is, a struct {int a, b;} in C terms) parameter, it appears to pass each of the struct elements in its own x86-64 GPR to the function, even though the x86-64 ABI specifies (sec. 3.2.3) that such a struct should be packed in a single 64-bit GPR.
This is in spite of LLVM's documentation claiming to match the C calling convention by default:
“ccc” - The C calling convention
This calling convention (the default if no other calling convention is specified) matches the target C calling conventions. This calling convention supports varargs function calls and tolerates some mismatch in the declared prototype and implemented declaration of the function (as does normal C).
My question, then, is: Am I doing something wrong to cause LLVM to not match the C calling convention, or is this known behavior? (At the very least, the documentation seems to be wrong, no?)
I can find only very few references to the issue at all on the web, such as this bug report from 2007, which claims to be fixed. It also claims that "First, LLVM has no way to deal with aggregates as singular Value*'s", which I don't know if it was true in 2007, but it doesn't seem to be true now, given the extractvalue/insertvalue instructions. I also found this SO question whose second (non-accepted) answer simply seems to accept implicitly that argument coercion has to be done manually.
I'm currently building code for doing argument coercion in my IR generator, but it is complicating my design considerably (not to mention making it architecture-specific), so if I'm simply doing something wrong, I'd rather know about that. :)
LLVM's support for C-language compatible calling convention is extremely limited I'm afraid. Several folks have wished for more direct calling convention support in LLVM (or a related library), but so far this has not emerged. That logic is currently encoded in the C-language frontend (Clang for example).
What LLVM provides is a mapping from specific LLVM IR types to specific C ABI lowerings for a specific CPU backend. You can see which IR types to use for a given C function by using Clang to emit LLVM IR, much as the comment above suggests:
https://c.compiler-explorer.com/z/8jWExWPYq
struct S { int x, y; };
void f(struct S s);
void test(int x, int y) {
struct S s = {x, y};
f(s);
}
Turns into:
define dso_local void #test(i32 noundef %0, i32 noundef %1) #0 {
%3 = alloca i32, align 4
%4 = alloca i32, align 4
%5 = alloca %struct.S, align 4
store i32 %0, ptr %3, align 4
store i32 %1, ptr %4, align 4
%6 = getelementptr inbounds %struct.S, ptr %5, i32 0, i32 0
%7 = load i32, ptr %3, align 4
store i32 %7, ptr %6, align 4
%8 = getelementptr inbounds %struct.S, ptr %5, i32 0, i32 1
%9 = load i32, ptr %4, align 4
store i32 %9, ptr %8, align 4
%10 = load i64, ptr %5, align 4
call void #f(i64 %10)
ret void
}
declare void #f(i64) #1
There is sadly some non-trivial logic to map specific C types into the LLVM IR that will match the ABI when lowered for a platform. Outside of extremely simple types (basic C integer types, pointers, float, double, maybe a few others), these aren't even portable between the different architecture ABIs/calling-conventions.
FWIW, the situation is even worse for C++ which has much more complexity here I'm afraid.
So your choices are to:
Use a very small set of types in a limited range of signatures that you build custom logic to lower correctly into LLVM IR, checking that it matches what Clang (or another C frontend) produces in every case.
Directly use Clang or another C frontend to emit the LLVM IR.
Take on the major project of extracting this ABI/calling-convention logic from Clang into a re-usable library. There has in the past been appetite for this in the LLVM/Clang communities, but it is a very large and complex undertaking from my understanding. There are some partial efforts (specifically for C and JITs) that you may be able to find and re-use, but I don't have a good memory of where all those are.
Related
I have a function that takes an array as argument and I need to get the size of the array first thing in the function. I need to do this in LLVM IR. Is this possible? I can access the array but I don't know the size.
void test(int[] a) {
}
is right now translating to
define void #test(i32* %__p__a) {
entry:
%a = alloca i32*, align 4
store i32* %__p__a , i32** %a, align 4
ret void
}
I need to get the size of the array first thing in the function. I need to do this in LLVM IR. Is this possible?
If all you have is an i32* with no additional information about what it points to, then no, it's not possible. In order to get an array's size, you'll need to store that information somewhere where the test function can access it.
Since this is your own language and you control what the generated LLVM IR looks like, you could for example represent arrays as structs that contain the array's size and a pointer to the data.
In section 20 of Scott Meyer's Effective C++, he states:
some compilers refuse to put objects consisting of only a double into a register
When passing built-in types by value, compilers will happily place the data in registers and quickly send ints/doubles/floats/etc. along. However, not all compilers will treat small objects with the same grace. I can easily understand why compilers would treat Objects differently - to pass an Object by value can be a lot more work than copying data members between the vtable and all the constructors.
But still. This seems like an easy problem for modern compilers to solve: "This class is small, maybe I can treat it differently". Meyer's statement seemed to imply that compilers WOULD make this optimization for objects consisting of only an int (or char or short).
Can someone give further insight as to why this optimization sometimes doesn't happen?
I found this document online on "Calling conventions for different C++ compilers and operating systems" (updated on 2018-04-25) which has a table depicting "Methods for passing structure, class and union objects".
From the table you can see that if an object contains long double, copy of entire object is transferred to stack for all compilers shown here.
Also from the same resource (with emphasis added):
There are several different methods to transfer a parameter to a function if the parameter is a structure, class or union object. A copy of the object is always made, and this copy is transferred to the called function either in registers, on the stack, or by a pointer, as specified in table 6. The symbols in the table specify which method to use. S takes precedence over I and R. PI and PS take precedence over all other passing methods.
As table 6 tells, an object cannot be transferred in registers if it is too big or too complex. For example, an object that has a copy constructor cannot be transferred in registers because the copy constructor needs an address of the object. The copy constructor is called by the caller, not the callee.
Objects passed on the stack are aligned by the stack word size, even if higher alignment would be desired. Objects passed by pointers are not aligned by any of the compilers studied, even if alignment is explicitly requested. The 64bit Windows ABI requires that objects passed by pointers be aligned by 16.
An array is not treated as an object but as a pointer, and no copy of the array is made, except if the array is wrapped into a structure, class or union.
The 64 bit compilers for Linux differ from the ABI (version 0.97) in the following respects: Objects with inheritance, member functions, or constructors can be passed in registers. Objects with copy constructor, destructor or virtual are passed by pointers rather than on the stack.
The Intel compilers for Windows are compatible with Microsoft. Intel compilers for Linux are compatible with Gnu.
Here is an example showing that LLVM clang with optimization level O3 treats a class with a single double data member just like it was a double:
$ cat main.cpp
#include <stdio.h>
class MyDouble {
public:
double d;
MyDouble(double _d):d(_d){}
};
void foo(MyDouble d)
{
printf("%lg\n",d.d);
}
int main(int argc, char **argv)
{
if (argc>5)
{
double x=(double)argc;
MyDouble d(x);
foo(d);
}
return 0;
}
When I compile it and view the generated bitcode file, I see that foo behaves
as if it operates on a double type input parameter:
$ clang++ -O3 -c -emit-llvm main.cpp
$ llvm-dis main.bc
Here is the relevant part:
; Function Attrs: nounwind uwtable
define void #_Z3foo8MyDouble(double %d.coerce) #0 {
entry:
%call = tail call i32 (i8*, ...)* #printf(i8* getelementptr inbounds ([5 x i8]* #.str, i64 0, i64 0), double %d.coerce)
ret void
}
See how foo declares its input parameter as double, and moves it around for
printing ``as is". Now let's compile the exact same code with O0:
$ clang++ -O0 -c -emit-llvm main.cpp
$ llvm-dis main.bc
When we look at the relevant part, we see that clang uses a getelementptr instruction to access its first (and only) data member d:
; Function Attrs: uwtable
define void #_Z3foo8MyDouble(double %d.coerce) #0 {
entry:
%d = alloca %class.MyDouble, align 8
%coerce.dive = getelementptr %class.MyDouble* %d, i32 0, i32 0
store double %d.coerce, double* %coerce.dive, align 1
%d1 = getelementptr inbounds %class.MyDouble* %d, i32 0, i32 0
%0 = load double* %d1, align 8
%call = call i32 (i8*, ...)* #printf(i8* getelementptr inbounds ([5 x i8]* #.str, i32 0, i32 0), double %0)
ret void
}
I'm experimenting with LLVM IR as an alternative to assembly for programming AVR chips, but I have encountered a stumbling block.
I have the following working code:
target triple = "avr-atmel-none"
define void #main() {
%1 = load i8, i8* inttoptr (i8 34 to i8*)
%or = or i8 %1, 1
store i8 %or, i8* inttoptr (i8 34 to i8*)
ret void
}
Notice the inttoptr (i8 34 to i8*) - that's a memory-mapped IO location on my chip, which I will interact with a lot, so I'd like to give it a name like #PORTA. However, I can't find a way to initialise a global variable as pointing to a specific memory address. This doesn't work:
#PORTA = constant i8* inttoptr (i8 34 to i8*)
as it initialises the contents of a pointer variable #PORTA, rather than creating it pointing to that address.
Is it possible to make a global alias for a certain memory address in LLVM IR? If not, is there some other shortcut that would allow me to alias these names?
Do you really need to create such alias on IR level? LLVM IR isn't supposed to be comfortable to write manually.
If you are generating code from C++ API then just make that inttoptr constant expression global and use it wherever you like:
ConstantExpr* porta = ...
void foo()
{
new LoadInst(..., porta);
...
new StoreInst(..., porta);
}
Sutter says this:
"In the low-level efficiency tradition of C and C++ alike, the
compiler is often not required to initialize variables unless you do
it explicitly (e.g., local variables, forgotten members omitted from
constructor initializer lists)"
I have always wondered why the compiler doesn't initialize primitives like int32 and float to 0. What is the performance hit if the compiler initializes it? It should be better than incorrect code.
This argument is incomplete, actually. Unitialized variables may have two reasons: efficiency and the lack of a suitable default.
1) Efficiency
It is, mostly, a left-over of the old days, when C compilers were simply C to assembly translators and performed no optimization whatsoever.
These days we have smart compilers and Dead Store Elimination which in most cases will eliminate redundant stores. Demo:
int foo(int a) {
int r = 0;
r = a + 3;
return r;
}
Is transformed into:
define i32 #foo(i32 %a) nounwind uwtable readnone {
%1 = add nsw i32 %a, 3
ret i32 %1
}
Still, there are cases where even the smarter compiler cannot eliminate the redundant store and this may have an impact. In the case of a large array that is later initialized piecemeal... the compiler may not realize that all values will end up being initialized and thus not remove the redundant writes:
int foo(int a) {
int* r = new int[10]();
for (unsigned i = 0; i <= a; ++i) {
r[i] = i;
}
return r[a % 2];
}
Note in the following the call to memset (that I required by suffixing the new call with () which is value initialization). It was not eliminated even though the 0 are unneeded.
define i32 #_Z3fooi(i32 %a) uwtable {
%1 = tail call noalias i8* #_Znam(i64 40)
%2 = bitcast i8* %1 to i32*
tail call void #llvm.memset.p0i8.i64(i8* %1, i8 0, i64 40, i32 4, i1 false)
br label %3
; <label>:3 ; preds = %3, %0
%i.01 = phi i32 [ 0, %0 ], [ %6, %3 ]
%4 = zext i32 %i.01 to i64
%5 = getelementptr inbounds i32* %2, i64 %4
store i32 %i.01, i32* %5, align 4, !tbaa !0
%6 = add i32 %i.01, 1
%7 = icmp ugt i32 %6, %a
br i1 %7, label %8, label %3
; <label>:8 ; preds = %3
%9 = srem i32 %a, 2
%10 = sext i32 %9 to i64
%11 = getelementptr inbounds i32* %2, i64 %10
%12 = load i32* %11, align 4, !tbaa !0
ret i32 %12
}
2) Default ?
The other issue is the lack of a suitable value. While a float could perfectly be initialized to NaN, what of integers ? There is no integer value that represents the absence of value, none at all! 0 is one candidate (among others), but one could argue it's one of the worst candidate: it's a very likely number, and thus likely has a specific meaning for the usecase at hand; are you sure that you are comfortable with this meaning being the default ?
Food for thought
Finally, there is one neat advantage of unitialized variables: they are detectable. The compiler may issue warnings (if it's smart enough), and Valgrind will raise errors. This make logical issues detectable, and only what is detected can be corrected.
Of course a sentinel value, such as NaN, would be as useful. Unfortunately... there is none for integers.
There are two ways in which initialisation might impact performance.
First, initialising a variable takes time. Granted, for a single variable it's probably negligible, but as others have suggested, it can add up with large numbers of variables, arrays, etc.
Second, who's to say that zero is a reasonable default? For every variable for which zero is a useful default, there's probably another one for which it isn't. In that case, if you do initialise to zero you then incur further overhead re-initialising the variable to whatever value you actually want. You essentially pay the initialisation overhead twice, rather than once if the default initialisation does not occur. Note that this is true no matter what you choose as a default value, zero or otherwise.
Given the overhead exists, it's typically more efficient to not initialise and to let the compiler catch any references to uninitialised variables.
Basically, a variable references a place in memory which can be modified to hold data. For an unitialized variable, all that the program needs to know is where this place is, and the compiler usually figures this out ahead of time, so no instructions are required. But when you want it to be initialized (to say, 0), the program needs to use an extra instruction to do so.
One idea might be to zero out the entire heap while the program is starting, with memset, then initialize all of the static stuff, but this isn't needed for anything that is set dynamically before it's read. This would also be a problem for stack-based functions which would need to zero out their stack frame every time a function is called. In short, it's much more efficient to allow variables to default to undefined, particularly when the stack is frequently being overwritten with newly called functions.
Compile with -Wmaybe-uninitialized and find out. Those are the only places the compiler would not be able to optmize out the primitive initialization.
As for the heap ...
I have a source C++ code which I parse using clang, producing llvm bytecode. From this point I want to process the file myself...
However I encoudered a problem. Consider the following scenario:
- I create a class with a nontrivial destructor or copy constructor.
- I define a function, where an object of this class is passed as a parameter, by value (no reference or pointer).
In the produced bytecode, I get a pointer instead. For classes without the destructor, the parameter is annotated as 'byval', but it is not so in this case.
As a result, I cannot distinguish if the parameter is passed by value, or really by a pointer.
Consider the following example:
Input file - cpass.cpp:
class C {
public:
int x;
~C() {}
};
void set(C val, int x) {val.x=x;};
void set(C *ptr, int x) {ptr->x=x;}
Compilation command line:
clang++ -c cpass.cpp -emit-llvm -o cpass.bc; llvm-dis cpass.bc
Produced output file (cpass.ll):
; ModuleID = 'cpass.bc'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-unknown-linux-gnu"
%class.C = type { i32 }
define void #_Z3set1Ci(%class.C* %val, i32 %x) nounwind {
%1 = alloca i32, align 4
store i32 %x, i32* %1, align 4
%2 = load i32* %1, align 4
%3 = getelementptr inbounds %class.C* %val, i32 0, i32 0
store i32 %2, i32* %3, align 4
ret void
}
define void #_Z3setP1Ci(%class.C* %ptr, i32 %x) nounwind {
%1 = alloca %class.C*, align 8
%2 = alloca i32, align 4
store %class.C* %ptr, %class.C** %1, align 8
store i32 %x, i32* %2, align 4
%3 = load i32* %2, align 4
%4 = load %class.C** %1, align 8
%5 = getelementptr inbounds %class.C* %4, i32 0, i32 0
store i32 %3, i32* %5, align 4
ret void
}
As you can see, the parameters of both set functions look exactly the same. So how can I tell that the first function was meant to take the parameter by value, instead of a pointer?
One solution could be to somehow parse the mangled function name, but it may not be always viable. What if somebody puts extern "C" before the function?
Is there a way to tell clang to keep the byval annotation, or to produce an extra annotation for each function parameter passed by a value?
Anton Korobeynikov suggests that I should dig into clang's LLVM IR emission. Unfortunately I know almost nothing about clang internals, the documentation is rather sparse. The Internals Manual of clang does not talk about IR emission. So I don't really know how to start, where to go to get the problem solved, hopefully without actually going through all of clang source code. Any pointers? Hints? Further reading?
In response to Anton Korobeynikov:
I know more-or-less how C++ ABI looks like with respect of parameter passing. Found some good reading here: http://agner.org./optimize/calling_conventions.pdf. But this is very platform dependent! This approach might not be feasable on different architectures or in some special circumstances.
In my case, for example, the function is going to be run on a different device than where it is being called from. The two devices don't share memory, so they don't even share the stack. Unless the user is passing a pointer (in which case we assume he knows what he is doing), an object should always be passed within the function-parameters message. If it has a nontrivial copy constructor, it should be executed by the caller, but the object should be created in the parameter area as well.
So, what I would like to do is to somehow override the ABI in clang, without too much intrusion into their source code. Or maybe add some additional annotation, which would be ignored in a normal compilation pipeline, but I could detect when parsing the .bc/.ll file. Or somehow differently reconstruct the function signature.
Unfortunately, "byval" is not just "annotation", it's parameter attribute which means a alot for optimizers and backends. Basically, the rules how to pass small structs / classes with and without non-trivial functions are government by platform C++ ABI, so you cannot just always use byval here.
In fact, byval here is just a result of minor optimization at frontend level. When you're passing stuff by value, then temporary object should be constructed on stack (via the default copy ctor). When you have a class which is something POD-like, then clang can deduce that copy ctor will be trivial and will optimize the pair of ctor / dtor out, passing just the "contents".
For non-trivial classes (like in your case) clang cannot perform such optimization and have to call both ctor and dtor. Thus you're seeing the pointer to temporary object is created.
Try to call your set() functions and you'll see what's going there.