/*
* %CopyrightBegin%
*
* Copyright Ericsson AB 2004-2009. All Rights Reserved.
*
* The contents of this file are subject to the Erlang Public License,
* Version 1.1, (the "License"); you may not use this file except in
* compliance with the License. You should have received a copy of the
* Erlang Public License along with this software. If not, it can be
* retrieved online at http://www.erlang.org/.
*
* Software distributed under the License is distributed on an "AS IS"
* basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
* the License for the specific language governing rights and limitations
* under the License.
*
* %CopyrightEnd%
*/
/* $Id$
*/
#include <stddef.h> /* offsetof() */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "global.h"
#include "erl_binary.h"
#include <sys/mman.h>
#include "hipe_arch.h"
#include "hipe_native_bif.h" /* nbif_callemu() */
#include "hipe_bif0.h"
#if !defined(__powerpc64__)
const unsigned int fconv_constant[2] = { 0x43300000, 0x80000000 };
#endif
/* Flush dcache and invalidate icache for a range of addresses. */
void hipe_flush_icache_range(void *address, unsigned int nbytes)
{
const unsigned int L1_CACHE_SHIFT = 5;
const unsigned long L1_CACHE_BYTES = 1 << L1_CACHE_SHIFT;
unsigned long start, p;
unsigned int nlines, n;
if (!nbytes)
return;
start = (unsigned long)address & ~(L1_CACHE_BYTES-1);
nlines =
(((unsigned long)address & (L1_CACHE_BYTES-1))
+ nbytes
+ (L1_CACHE_BYTES-1)) >> L1_CACHE_SHIFT;
p = start;
n = nlines;
do {
asm volatile("dcbst 0,%0" : : "r"(p) : "memory");
p += L1_CACHE_BYTES;
} while (--n != 0);
asm volatile("sync");
p = start;
n = nlines;
do {
asm volatile("icbi 0,%0" : : "r"(p) : "memory");
p += L1_CACHE_BYTES;
} while (--n != 0);
asm volatile("sync\n\tisync");
}
/*
* Management of 32MB code segments for regular code and trampolines.
*/
#define SEGMENT_NRBYTES (32*1024*1024) /* named constant, _not_ a tunable */
static struct segment {
unsigned int *base; /* [base,base+32MB[ */
unsigned int *code_pos; /* INV: base <= code_pos <= tramp_pos */
unsigned int *tramp_pos; /* INV: tramp_pos <= base+32MB */
} curseg;
#define in_area(ptr,start,nbytes) \
((unsigned long)((char*)(ptr) - (char*)(start)) < (nbytes))
/* Darwin breakage */
#if !defined(MAP_ANONYMOUS) && defined(MAP_ANON)
#define MAP_ANONYMOUS MAP_ANON
#endif
#if defined(__powerpc64__)
static void *new_code_mapping(void)
{
char *map_hint, *map_start;
/*
* Allocate a new 32MB code segment in the low 2GB of the address space.
*
* This is problematic for several reasons:
* - Linux/ppc64 lacks the MAP_32BIT flag that Linux/x86-64 has.
* - The address space hint to mmap is only respected if that
* area is available. If it isn't, then mmap falls back to its
* defaults, which (according to testing) results in very high
* (and thus useless for us) addresses being returned.
* - Another mapping, presumably the brk, also occupies low addresses.
*
* As initial implementation, simply start allocating at the 0.5GB
* boundary. This leaves plenty of space for the brk before malloc
* needs to switch to mmap, while allowing for 1.5GB of code.
*
* A more robust implementation would be to parse /proc/self/maps,
* reserve all available space between (say) 0.5GB and 2GB with
* PROT_NONE MAP_NORESERVE mappings, and then allocate by releasing
* 32MB segments and re-mapping them properly. This would work on
* Linux/ppc64, I have no idea how things should be done on Darwin64.
*/
if (curseg.base)
map_hint = (char*)curseg.base + SEGMENT_NRBYTES;
else
map_hint = (char*)(512*1024*1024); /* 0.5GB */
map_start = mmap(map_hint, SEGMENT_NRBYTES,
PROT_EXEC|PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANONYMOUS,
-1, 0);
if (map_start != MAP_FAILED &&
(((unsigned long)map_start + (SEGMENT_NRBYTES-1)) & ~0x7FFFFFFFUL)) {
fprintf(stderr, "mmap with hint %p returned code memory %p\r\n", map_hint, map_start);
abort();
}
return map_start;
}
#else
static void *new_code_mapping(void)
{
return mmap(0, SEGMENT_NRBYTES,
PROT_EXEC|PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANONYMOUS,
-1, 0);
}
#endif
static int check_callees(Eterm callees)
{
Eterm *tuple;
Uint arity;
Uint i;
if (is_not_tuple(callees))
return -1;
tuple = tuple_val(callees);
arity = arityval(tuple[0]);
for (i = 1; i <= arity; ++i) {
Eterm mfa = tuple[i];
if (is_not_tuple(mfa) ||
tuple_val(mfa)[0] != make_arityval(3) ||
is_not_atom(tuple_val(mfa)[1]) ||
is_not_atom(tuple_val(mfa)[2]) ||
is_not_small(tuple_val(mfa)[3]) ||
unsigned_val(tuple_val(mfa)[3]) > 255)
return -1;
}
return arity;
}
static unsigned int *try_alloc(Uint nrwords, int nrcallees, Eterm callees, unsigned int **trampvec)
{
unsigned int *base, *address, *tramp_pos, nrfreewords;
int trampnr;
tramp_pos = curseg.tramp_pos;
address = curseg.code_pos;
nrfreewords = tramp_pos - address;
if (nrwords > nrfreewords)
return NULL;
curseg.code_pos = address + nrwords;
nrfreewords -= nrwords;
base = curseg.base;
for (trampnr = 1; trampnr <= nrcallees; ++trampnr) {
Eterm mfa = tuple_val(callees)[trampnr];
Eterm m = tuple_val(mfa)[1];
Eterm f = tuple_val(mfa)[2];
unsigned int a = unsigned_val(tuple_val(mfa)[3]);
unsigned int *trampoline = hipe_mfa_get_trampoline(m, f, a);
if (!in_area(trampoline, base, SEGMENT_NRBYTES)) {
if (nrfreewords < 4)
return NULL;
nrfreewords -= 4;
tramp_pos = trampoline = tramp_pos - 4;
#if defined(__powerpc64__)
trampoline[0] = 0x3D600000; /* addis r11,0,0 */
trampoline[1] = 0x616B0000; /* ori r11,r11,0 */
#else
trampoline[0] = 0x39600000; /* addi r11,r0,0 */
trampoline[1] = 0x3D6B0000; /* addis r11,r11,0 */
#endif
trampoline[2] = 0x7D6903A6; /* mtctr r11 */
trampoline[3] = 0x4E800420; /* bctr */
hipe_flush_icache_range(trampoline, 4*sizeof(int));
hipe_mfa_set_trampoline(m, f, a, trampoline);
}
trampvec[trampnr-1] = trampoline;
}
curseg.tramp_pos = tramp_pos;
return address;
}
void *hipe_alloc_code(Uint nrbytes, Eterm callees, Eterm *trampolines, Process *p)
{
Uint nrwords;
int nrcallees;
Eterm trampvecbin;
unsigned int **trampvec;
unsigned int *address;
unsigned int *base;
struct segment oldseg;
if (nrbytes & 0x3)
return NULL;
nrwords = nrbytes >> 2;
nrcallees = check_callees(callees);
if (nrcallees < 0)
return NULL;
trampvecbin = new_binary(p, NULL, nrcallees*sizeof(unsigned int*));
trampvec = (unsigned int**)binary_bytes(trampvecbin);
address = try_alloc(nrwords, nrcallees, callees, trampvec);
if (!address) {
base = new_code_mapping();
if (base == MAP_FAILED)
return NULL;
oldseg = curseg;
curseg.base = base;
curseg.code_pos = base;
curseg.tramp_pos = (unsigned int*)((char*)base + SEGMENT_NRBYTES);
address = try_alloc(nrwords, nrcallees, callees, trampvec);
if (!address) {
munmap(base, SEGMENT_NRBYTES);
curseg = oldseg;
return NULL;
}
/* commit to new segment, ignore leftover space in old segment */
}
*trampolines = trampvecbin;
return address;
}
static unsigned int *alloc_stub(Uint nrwords)
{
unsigned int *address;
unsigned int *base;
struct segment oldseg;
address = try_alloc(nrwords, 0, NIL, NULL);
if (!address) {
base = new_code_mapping();
if (base == MAP_FAILED)
return NULL;
oldseg = curseg;
curseg.base = base;
curseg.code_pos = base;
curseg.tramp_pos = (unsigned int*)((char*)base + SEGMENT_NRBYTES);
address = try_alloc(nrwords, 0, NIL, NULL);
if (!address) {
munmap(base, SEGMENT_NRBYTES);
curseg = oldseg;
return NULL;
}
/* commit to new segment, ignore leftover space in old segment */
}
return address;
}
static void patch_imm16(Uint32 *address, unsigned int imm16)
{
unsigned int insn = *address;
*address = (insn & ~0xFFFF) | (imm16 & 0xFFFF);
hipe_flush_icache_word(address);
}
#if defined(__powerpc64__)
static void patch_li64(Uint32 *address, Uint64 value)
{
patch_imm16(address+0, value >> 48);/* addis r,0,value@highest */
patch_imm16(address+1, value >> 32);/* ori r,r,value@higher */
/* sldi r,r,32 */
patch_imm16(address+3, value >> 16);/* oris r,r,value@h */
patch_imm16(address+4, value); /* ori r,r,value@l */
}
static int patch_li31(Uint32 *address, Uint32 value)
{
if ((value >> 31) != 0)
return -1;
patch_imm16(address, value >> 16); /* addis r,0,value@h */
patch_imm16(address+1, value); /* ori r,r,value@l */
}
void hipe_patch_load_fe(Uint *address, Uint value)
{
patch_li64((Uint32*)address, value);
}
int hipe_patch_insn(void *address, Uint64 value, Eterm type)
{
switch (type) {
case am_closure:
case am_constant:
patch_li64((Uint32*)address, value);
return 0;
case am_atom:
case am_c_const:
return patch_li31((Uint32*)address, value);
default:
return -1;
}
}
void *hipe_make_native_stub(void *beamAddress, unsigned int beamArity)
{
unsigned int *code;
if ((unsigned long)&nbif_callemu & ~0x01FFFFFCUL)
abort();
code = alloc_stub(7);
/* addis r12,0,beamAddress@highest */
code[0] = 0x3d800000 | (((unsigned long)beamAddress >> 48) & 0xffff);
/* ori r12,r12,beamAddress@higher */
code[1] = 0x618c0000 | (((unsigned long)beamAddress >> 32) & 0xffff);
/* sldi r12,r12,32 (rldicr r12,r12,32,31) */
code[2] = 0x798c07c6;
/* oris r12,r12,beamAddress@h */
code[3] = 0x658c0000 | (((unsigned long)beamAddress >> 16) & 0xffff);
/* ori r12,r12,beamAddress@l */
code[4] = 0x618c0000 | ((unsigned long)beamAddress & 0xffff);
/* addi r0,0,beamArity */
code[5] = 0x38000000 | (beamArity & 0x7FFF);
/* ba nbif_callemu */
code[6] = 0x48000002 | (unsigned long)&nbif_callemu;
hipe_flush_icache_range(code, 7*sizeof(int));
return code;
}
#else /* !__powerpc64__ */
/*
* To load a 32-bit immediate value 'val' into Rd (Rd != R0):
*
* addi Rd, 0, val@l // val & 0xFFFF
* addis Rd, Rd, val@ha // ((val + 0x8000) >> 16) & 0xFFFF
*
* The first addi sign-extends the low 16 bits, so if
* val&(1<<15), the high portion of Rd will be -1 not 0.
* val@ha compensates by adding 1 if val&(1<<15).
*/
static unsigned int at_ha(unsigned int val)
{
return ((val + 0x8000) >> 16) & 0xFFFF;
}
static void patch_li(Uint32 *address, Uint32 value)
{
patch_imm16(address, value);
patch_imm16(address+1, at_ha(value));
}
void hipe_patch_load_fe(Uint32 *address, Uint value)
{
patch_li(address, value);
}
int hipe_patch_insn(void *address, Uint32 value, Eterm type)
{
switch (type) {
case am_closure:
case am_constant:
case am_atom:
case am_c_const:
break;
default:
return -1;
}
patch_li((Uint32*)address, value);
return 0;
}
/* called from hipe_bif0.c:hipe_bifs_make_native_stub_2()
and hipe_bif0.c:hipe_make_stub() */
void *hipe_make_native_stub(void *beamAddress, unsigned int beamArity)
{
unsigned int *code;
/*
* Native code calls BEAM via a stub looking as follows:
*
* addi r12,0,beamAddress@l
* addi r0,0,beamArity
* addis r12,r12,beamAddress@ha
* ba nbif_callemu
*
* I'm using r0 and r12 since the standard SVR4 ABI allows
* them to be modified during function linkage. Trampolines
* (for b/bl to distant targets) may modify r11.
*
* The runtime system code is linked completely below the
* 32MB address boundary. Hence the branch to nbif_callemu
* is done with a 'ba' instruction.
*/
/* verify that 'ba' can reach nbif_callemu */
if ((unsigned long)&nbif_callemu & ~0x01FFFFFCUL)
abort();
code = alloc_stub(4);
/* addi r12,0,beamAddress@l */
code[0] = 0x39800000 | ((unsigned long)beamAddress & 0xFFFF);
/* addi r0,0,beamArity */
code[1] = 0x38000000 | (beamArity & 0x7FFF);
/* addis r12,r12,beamAddress@ha */
code[2] = 0x3D8C0000 | at_ha((unsigned long)beamAddress);
/* ba nbif_callemu */
code[3] = 0x48000002 | (unsigned long)&nbif_callemu;
hipe_flush_icache_range(code, 4*sizeof(int));
return code;
}
#endif /* !__powerpc64__ */
static void patch_b(Uint32 *address, Sint32 offset, Uint32 AA)
{
Uint32 oldI = *address;
Uint32 newI = (oldI & 0xFC000001) | ((offset & 0x00FFFFFF) << 2) | (AA & 2);
*address = newI;
hipe_flush_icache_word(address);
}
int hipe_patch_call(void *callAddress, void *destAddress, void *trampoline)
{
if ((Uint32)destAddress == ((Uint32)destAddress & 0x01FFFFFC)) {
/* The destination is in the [0,32MB[ range.
We can reach it with a ba/bla instruction.
This is the typical case for BIFs and primops.
It's also common for trap-to-BEAM stubs (on ppc32). */
patch_b((Uint32*)callAddress, (Uint32)destAddress >> 2, 2);
} else {
Sint32 destOffset = ((Sint32)destAddress - (Sint32)callAddress) >> 2;
if (destOffset >= -0x800000 && destOffset <= 0x7FFFFF) {
/* The destination is within a [-32MB,+32MB[ range from us.
We can reach it with a b/bl instruction.
This is typical for nearby Erlang code. */
patch_b((Uint32*)callAddress, destOffset, 0);
} else {
/* The destination is too distant for b/bl/ba/bla.
Must do a b/bl to the trampoline. */
Sint32 trampOffset = ((Sint32)trampoline - (Sint32)callAddress) >> 2;
if (trampOffset >= -0x800000 && trampOffset <= 0x7FFFFF) {
/* Update the trampoline's address computation.
(May be redundant, but we can't tell.) */
#if defined(__powerpc64__)
/* This relies on the fact that we allocate code below 2GB. */
patch_li31((Uint32*)trampoline, (Uint32)destAddress);
#else
patch_li((Uint32*)trampoline, (Uint32)destAddress);
#endif
/* Update this call site. */
patch_b((Uint32*)callAddress, trampOffset, 0);
} else
return -1;
}
}
return 0;
}
void hipe_arch_print_pcb(struct hipe_process_state *p)
{
#define U(n,x) \
printf(" % 4d | %s | 0x%0*lx | %*s |\r\n", (int)offsetof(struct hipe_process_state,x), n, 2*(int)sizeof(long), (unsigned long)p->x, 2+2*(int)sizeof(long), "")
U("nra ", nra);
U("narity ", narity);
#undef U
}
