1
0
mirror of https://github.com/KarolS/millfork.git synced 2024-12-25 21:29:25 +00:00
millfork/docs/abi/inlining.md

138 lines
4.1 KiB
Markdown
Raw Normal View History

2019-07-15 12:21:50 +00:00
[< back to index](../doc_index.md)
2018-04-02 22:21:26 +00:00
# Macros and inlining
2018-01-04 00:15:04 +00:00
## Macros
2018-01-04 00:15:04 +00:00
2020-01-03 22:56:37 +00:00
Functions defined with the `macro` keyword are not actual functions, but they are used as syntax replacements.
It implies the following:
* macros must return `void`
* cannot be `inline`, `noinline` or `extern`
* can be `asm` - in this case, they should **not** end with a return instruction
* do not have an address
* their invocations cannot be used as expressions
* in case of `asm` macros, the parameters:
2020-01-03 22:56:37 +00:00
* must be defined as either `const` (compile-time constants), `ref` (variables) or `register(XX)` (registers, where XX is the register you want to use)
* at most one parameter can be defined as a register
* in case of non-`asm` macros, the parameters
* must be defined as either `ref` (variables; default, may be omitted) `const` (compile-time constants), or `call` (expressions, which are evaluated every time they are used)
* `ref` parameters exceptionally can have their type declared as `void`; such parameters accept variables of any type
* `call` parameters exceptionally can have their type declared as `void`;
such parameters accept expressions of any type, including `void`, however, you cannot assign from those expressions
2020-01-03 22:56:37 +00:00
* macros do not have their own scope (they reuse the scope from their invocations) exceptions: the parameters and the local labels defined in assembly
* control-flow statements (`break`, `continue`, `return`, `goto`, `label`) are run as if places in the caller function
When invoking a macro, you need to pass variables as arguments to parameters annotated with `ref` and constants as arguments annotated with `const`.
Invoking a non-`asm` macro requires the types of variables via `ref` parameters to match precisely.
No type conversions are performed.
Exception: parameters of type `void` can accept a variable of any type.
For parameters defined as `const`, `register(XX)` or `call`, the usual type conversions are performed.
2020-01-03 22:56:37 +00:00
You can invoke a macro from assembly, by preceding the invocation with `+`
Examples:
macro void inc_x() {
x += 1
}
byte add_two_1(byte x) {
inc_x()
inc_x()
return x
}
macro void inc(byte b) {
b += 1
}
byte add_two_2(byte x) {
inc(x)
inc(x)
return x
}
macro void perform_twice(void call f) {
f
f
}
byte add_two_3(byte x) {
perform_twice(inc(x))
return x
}
2020-01-03 22:56:37 +00:00
macro void add(byte b, byte v) {
b += v
}
macro void retu(byte result) {
return result
}
byte add_two_4(byte x) {
2020-01-03 22:56:37 +00:00
add(x, 2)
retu(x)
}
macro asm byte add_asm(byte ref b, byte const v) {
LDA b
CLC
ADC #v
STA b
2020-01-03 22:56:37 +00:00
// no RTS!
}
byte add_two_5(byte x) {
2020-01-03 22:56:37 +00:00
add_asm(x, 2)
return x
}
macro asm byte add_asm_2(byte ref b, byte register(x) v) {
TXA
CLC
ADC b
STA b
// no RTS!
}
byte add_two_6(byte x) {
add_asm_2(x, 2)
return x
}
2018-01-04 00:15:04 +00:00
## Automatic inlining
2019-06-11 22:20:24 +00:00
You can control inlining behavior in several ways:
* functions declared with the `const` keyword called with constant arguments will always be inlined,
with the whole invocation being converted into a single constant, regardless of `inline` and `noinline` keywords;
calls with non-constant arguments are subject to the regular rules.
2019-06-11 22:20:24 +00:00
* functions declared with the `inline` keyword will be inlined if possible
* functions declared with the `noinline` keyword will never be inlined
* the remaining functions may be inlined only if the `-finline` command-line option is enabled
and the compiler decides the function is worth doing
2019-12-01 02:16:09 +00:00
## Automatic subroutine extraction
Subroutine extraction is the opposite of inlining.
When given the `-fsubroutine-extraction`, the compiler will attempt to extract common code fragments to new subroutines.
The code will get smaller and slower.
Generally, when using `-fsubroutine-extraction`, it's recommended to also use `-finline`.
This allows the compiler to first inline and optimize code and then extract it back when appropriate.