The ARM processor (Thumb-2), part 16: The calling convention

Raymond

For non-variadic functions, the Windows calling convention for ARM matches the Procedure Call Standard for the Arm Architecture, so this will largely match what you see on most other operating systems.

The fine points of the calling convention are spelled out in the standard document, but here’s the short and simple version.

First, parameters smaller than 32 bits are extended to 32-bit values in a manner consistent with their type: Signed types are sign-extended and unsigned types are zero-extended.

Next, parameters are laid out as if inside a giant structure:

struct Parameters
{
#if function returns a structure larger than 64 bits
    Ret* result;
#endif
#if function has a "this" pointer
    T* this;
#endif
    T1 param1;
    T2 param2;
    T3 param3;
    T4 param4;
    T5 param5;
    ... etc ...
};

Padding is inserted as necessary to preserve natural alignment.

The first 4 × 4 = 16 bytes of the resulting structure are loaded into registers r0 through r3, and the rest are put onto the stack.

For example, suppose we have this function:

void f(int8_t a, int64_t b, int16_t c);

The parameter a is sign-extended to int32_t and placed in r0.

The next parameter requires 8-byte alignment, so we skip r1 so that we can reach 8-byte alignment starting at r2.

Parameter b goes into registers r2 and r3.

Parameter c goes onto the stack.

On function return, register r0 contains the integer value. If the return value is smaller than 32 bits, it is extended to a 32-bit value in a manner consistent with its type. If the return value is a 64-bit integer, then r1 holds the upper 32 bits of the result. If the return type is a floating point variable, it is returned in s0 or d0, as appropriate.

Things get messier once you introduce floating point parameters.

Each floating point parameter goes into the next available s# or d# register, starting with s0 and d0. Remember that the s# and d# registers overlap, so if you use s0 and then need a double-precision register, you have to move up to d1.

On the other hand, parameters can backfill: If you need a single-precision register, you can use an odd-numbered s# register that had previously been skipped in order to get to an earlier double-precision parameter.

Floating point registers used for passing parameters are s0 through s15, and d0 through d7. If you run out of floating point registers, then that parameter and all subsequent floating point parameters go onto the stack. (No backfilling after spilling.)

void f(int i1, float f1, int i2, double d1, float f2);

The parameter assignment for this function goes like this:

  • i1 goes into r0.
  • f1 goes into s0.
  • i2 goes into r1.
  • d1 goes into d1. Skip over s1.
  • f2 goes into s1. Go backward to backfill s1.

Integer parameters and floating point parameters are allocated independently, and floating point parameters can backfill. Together, this means that a lot of distinct function signatures end up using the same registers:

void f(int i1, float f1, int i2, double d1, float f2);
void f(int i1, int i2, float f1, double d1, float f2);
void f(int i1, int i2, float f1, float f2, double d1);
void f(float f1, float f2, double d1, int i1, int i2);

All of these functions pass the parameters in the same registers, even though they are listed differently in the source code.

There is no parameter home space on the stack. At function entry, the first stack-based parameter is stored directly at the top of the stack.

If any parameters were put onto the stack, they are the responsibility of the caller to clean up. In practice, instead of cleaning the stack after every call, the caller preadjusts the stack pointer at function entry to reserve space for all outbound stack-based parameters and just reuses the space for each function call, doing the cleanup at the end of the function.

Variadic functions follow a different set of register assignment rules: All floating point parameters are passed as if they were integer parameters: A single-precision floating point parameter is passed as if it were a 32-bit integer, and a double-precision floating point parameter is passed as if it were a 64-bit integer. This rule applies even to the non-variadic parameters.

Next time, we’ll look at how these parameter passing rules are implemented in code.

3 comments

Comments are closed. Login to edit/delete your existing comments

  • Antonio Rodríguez

    “Remember that the s# and d# registers overlap, so if you use s0 and then need a double-precision register, you have to move up to d1.”

    Even through the FPU hasn’t been covered, its register model has been seen in the first part of the series, published three weeks ago. In it, Raymond states that the single and double precision registers overlap: s0 and s1 share space with d0, s2 and s3 with d1, and so on. So, if you are using s0, the first half of d0 is already in use, and the first fully available double precision register is d1. That leaves s1 (the second half of d0) free, which explains the backfill.

    It is a bit strange if you have in mind the 80×87 register model, where each register can hold either a single, double or extended precision value, simply leaving the unused bits alone. But, hey, this is a RISC processor, so you have to assume things work differently. And these things aren’t unseen even on CISC processors: the 8080/Z80 can operate pairs of 8-bit registers as 16-bit registers, and even the 8086/88 can pair two 16-bit registers to contain 32-bit results and operands in the MUL and DIV instructions.

    • Dmitry

      Well, in fact, x87 registers are 80 bits wide and always hold Extended values, and fld/fild instructions perform appropriate conversion while loading if necessary (fst/fstp/fist/fistp converting back if necessary). But yes, the RISCy way is quite different.