5. PowerPC Backend

5.1 Additional options for this version

This backend provides the following additional options:

`-merge-constants'

Place identical floating point constants at the same memory location. This can reduce program size.

`-const-in-data'

By default constant data will be placed in the .rodata section. Using this option it will be placed in the .data section. Note that on operating systems with memory protection this option will disable write-protection of constant data.

`-fsub-zero'

Use fsub to load a floating-point-register with zero. This is faster but requires all registers to always contain valid values (i.e. no NaNs etc.) which may not be the case depending on startup-code, libraries etc.

`-amiga-align'

Do not require any alignments greater than 2 bytes. This is needed when accessing Amiga system-structures, but can cause a performance penalty.

`-elf'

Do not prefix symbols with '_'. Prefix labels with '.'.

`-poweropen'

Generate code for the PowerOpen ABI like used in AIX. This does not work correctly yet.

`-sc'

Generate code for the modified PowerOpen ABI used in the StormC compiler (aka WarpOS ABI).

`-no-regnames'

Do not use register names but only numbers in the assembly output. This is necessary to avoid name-conflicts when using `-elf'.

`-setccs'

The V.4 ABI requires signalling (in a bit of the condition code register) when arguments to varargs-functions are passed in floating-point registers. vbcc usually does not make use of this and therefore does not set that bit by default. This may lead to problems when linking objects compiled by vbcc to objects/libraries created by other compilers and calling varargs-functions with floating-point arguments. `-setccs' will fix this problem.

`-no-peephole'

Do not perform several peephole optimizations. Currently includes:

• better use of d16(r) addressing
• use of indexed addressing modes
• use of update-flag
• use of record-flag
• use of condition-code-registers to avoid certain branches

`-use-lmw'

Use lmw/stmw-instructions. This can significantly reduce code-size. However these instructions may be slower on certain PPCs.

`-madd'

Use the fmadd/fmsub instructions for combining multiplication with addition/subtraction in one instruction. As these instructions do not round between the operations, they have increased precision over separate addition and multiplication.

While this usually does no harm, it is not ISO conforming and therefore not the default behaviour.

`-eabi'

Use the PowerPC Embedded ABI (eabi).

`-sd'

Place all objects in small data-sections.

`-gas'

Create code suitable for the GNU assembler.

`-no-align-args'

Do not align function arguments on the stack stricter than 4 bytes. Default with `-poweropen'.

`-use-commons'

Use real common symbols instead of bss symbols for non-initialized external variables.

`-baserel32os4'

Use 32bit base-relative addressing as used by AmigaOS 4.

`-baserel32mos'

Use 32bit base-relative addressing as used by MorphOS.

5.2 ABI

This backend supports the following registers:

• r0 through r31 for the general purpose registers,
• f0 through f31 for the floating point registers and
• cr0 through cr7 for the condition-code registers.

Additionally, the register pairs r3/r4, r5/r6, r7/r8, r9/r10, r14/r15, r16/r17, r18/r19, r20/r21, r22/r23, r24/r25, r26/r27, r28/r29 and r30/r31 are available.

r0, r11, r12, f0, f12 and f13 are reserved by the backend.

The current version generates assembly output for use with `vasmppc' or the GNU assembler. The generated code should work on 32bit systems based on a PowerPC CPU using the V.4 ABI or the PowerPC Embedded ABI (eabi).

The registers r0, r3-r12, f0-f13 and cr0-cr1 are used as scratch registers (i.e. they can be destroyed in function calls), all other registers are preserved. r1 is the stack-pointer and r13 is the small-data-pointer if small-data-mode is used.

The first 8 function arguments which have integer or pointer types are passed in registers r3 through r10 and the first 8 floating-point arguments are passed in registers f1 through f8. All other arguments are passed on the stack.

Integers and pointers are returned in r3 (and r4 for long long), floats and doubles in f1. All other types are returned by passing the function the address of the result as a hidden argument - so when you call such a function without a proper declaration in scope you can expect a crash.

The elementary data types are represented like:

type size in bits alignment in bytes (-amiga-align) char 8 1 (1) short 16 2 (2) int 32 4 (2) long 32 4 (2) long long 64 8 (2) all pointers 32 4 (2) float 32 4 (2) double 64 8 (2)

5.3 Target-specific variable-attributes

The PPC-backend offers the following variable-attributes:

__saveds

Load the pointer to the small data segment at function-entry. Applicable only to functions.

__chip

Place variable in chip-memory. Only applicable on AmigaOS to variables with static storage-duration.

__far

Do not place this variable in the small-data segment in small-data-mode. No effect in large-data-mode. Only applicable to variables with static storage- duration.

__near

Currently ignored.

__saveall

Make sure all registers are saved by this function. On lower optimization levels, all volatile registers will be saved additionally. On higher levels, only the ones that may be destroyed, are saved.

__interrupt

Return with en rfi-instruction rather than blr.

__section("name","attr")

Place this function/object in section "name" with attributes "attr".

5.4 Target-specific pragmas

The PPC-backend offers the following #pragmas:

#pragma amiga-align

Set alignment like -amiga-alignment option.

#pragma natural-align

Align every type to its own size.

#pragma default-align

Set alignment according to command-line options.

5.5 Predefined Macros

This backend defines the following macros:

__PPC__

5.6 Stack

If the `-stack-check' option is used, every function-prologue will call the function __stack_check with the stacksize needed by this function in register r12. This function has to consider its own stacksize and must restore all registers.

5.7 Stdarg

A possible <stdarg.h> for V.4 ABI could look like this:

typedef struct { int gpr; int fpr; char *regbase; char *membase; } va_list; char *__va_start(void); char *__va_regbase(void); int __va_fixedgpr(void); int __va_fixedfpr(void); #define va_start(vl,dummy) \ ( \ vl.gpr=__va_fixedgpr(), \ vl.fpr=__va_fixedfpr(), \ vl.regbase=__va_regbase(), \ vl.membase=__va_start() \ ) #define va_end(vl) ((vl).regbase=(vl).membase=0) #define va_copy(new,old) ((new)=(old)) #define __va_align(type) (__alignof(type)>=4?__alignof(type):4) #define __va_do_align(vl,type) ((vl).membase=(char *)((((unsigned int)((vl).membase))+__va_align(type)-1)/__va_align(type)*__va_align(type))) #define __va_mem(vl,type) (__va_do_align((vl),type),(vl).membase+=sizeof(type),((type*)((vl).membase))[-1]) #define va_arg(vl,type) \ ( \ (__typeof(type)&127)>10? \ __va_mem((vl),type) \ : \ ( \ (((__typeof(type)&127)>=6&&(__typeof(type)&127)<=8)) ? \ ( \ ++(vl).fpr<=8 ? \ ((type*)((vl).regbase+32))[(vl).fpr-1] \ : \ __va_mem((vl),type) \ ) \ : \ ( \ ++(vl).gpr<=8 ? \ ((type*)((vl).regbase+0))[(vl).gpr-1] \ : \ __va_mem((vl),type) \ ) \ ) \ )

A possible <stdarg.h> for PowerOpen ABI could look like this:

typedef unsigned char *va_list; #define __va_align(type) (4) #define __va_do_align(vl,type) ((vl)=(char *)((((unsigned int)(vl))+__va_align(type)-1)/__va_align(type)*__va_align(type))) #define __va_mem(vl,type) (__va_do_align((vl),type),(vl)+=sizeof(type),((type*)(vl))[-1]) #define va_start(ap, lastarg) ((ap)=(va_list)(&lastarg+1)) #define va_arg(vl,type) __va_mem(vl,type) #define va_end(vl) ((vl)=0) #define va_copy(new,old) ((new)=(old))

5.8 Known problems

• composite types are put on the stack rather than passed via pointer
• indication of fp-register-args with bit 6 of cr is not done well