mirror/bsnes
1
0
Fork 0
mirror of https://github.com/bsnes-emu/bsnes.git synced 2025-09-02 05:32:59 +02:00
Code Issues Projects Releases Wiki Activity

Update to v101r18 release.

Browse Source 
byuu says:

Changelog:

  - added 30 new PAL games to icarus (courtesy of Mikerochip)
  - new version of libco no longer requires mprotect nor W|X permissions
  - nall: default C compiler to -std=c11 instead of -std=c99
  - nall: use `-fno-strict-aliasing` during compilation
  - updated nall/certificates (hopefully for the last time)
  - updated nall/http to newer coding conventions
  - nall: improve handling of range() function

I didn't really work on higan at all, this is mostly just a release
because lots of other things have changed.

The most interesting is `-fno-strict-aliasing` ... basically, it joins
`-fwrapv` as being "stop the GCC developers from doing *really* evil
shit that could lead to security vulnerabilities or instabilities."

For the most part, it's a ~2% speed penalty for higan. Except for the
Sega Genesis, where it's a ~10% speedup. I have no idea how that's
possible, but clearly something's going very wrong with strict aliasing
on the Genesis core.

So ... it is what it is. If you need the performance for the non-Genesis
cores, you can turn it off in your builds. But I'm getting quite sick of
C++'s "surprises" and clever compiler developers, so I'm keeping it on
in all of my software going forward.
This commit is contained in:
Tim Allen
2016-09-14 21:55:53 +10:00
parent 750af6ebc3
commit 45a725e4b7
9 changed files with 421 additions and 412 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
amd64.c
View File

@@ -1,15 +1,12 @@
/* /*
libco.amd64 (2015-06-19) libco.amd64 (2016-09-14)
author: byuu author: byuu
license: public domain license: public domain
*/ */
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
#include "settings.h"
//Win64 only: provides a substantial speed-up, but will thrash XMM regs
//do not use this unless you are certain your application won't use SSE
//#define LIBCO_AMD64_NO_SSE
#include <assert.h> #include <assert.h>
#include <stdlib.h> #include <stdlib.h>
@@ -22,9 +19,14 @@ static thread_local long long co_active_buffer[64];
static thread_local cothread_t co_active_handle = 0; static thread_local cothread_t co_active_handle = 0;
static void (*co_swap)(cothread_t, cothread_t) = 0; static void (*co_swap)(cothread_t, cothread_t) = 0;
#ifdef LIBCO_MPROTECT
alignas(4096)
#else
section(text)
#endif
#ifdef _WIN32 #ifdef _WIN32
/* ABI: Win64 */ /* ABI: Win64 */
static unsigned char co_swap_function[] = { static const unsigned char co_swap_function[4096] = {
0x48, 0x89, 0x22, /* mov [rdx],rsp */ 0x48, 0x89, 0x22, /* mov [rdx],rsp */
0x48, 0x8b, 0x21, /* mov rsp,[rcx] */ 0x48, 0x8b, 0x21, /* mov rsp,[rcx] */
0x58, /* pop rax */ 0x58, /* pop rax */
@@ -36,7 +38,7 @@ static void (*co_swap)(cothread_t, cothread_t) = 0;
0x4c, 0x89, 0x6a, 0x30, /* mov [rdx+48],r13 */ 0x4c, 0x89, 0x6a, 0x30, /* mov [rdx+48],r13 */
0x4c, 0x89, 0x72, 0x38, /* mov [rdx+56],r14 */ 0x4c, 0x89, 0x72, 0x38, /* mov [rdx+56],r14 */
0x4c, 0x89, 0x7a, 0x40, /* mov [rdx+64],r15 */ 0x4c, 0x89, 0x7a, 0x40, /* mov [rdx+64],r15 */
#if !defined(LIBCO_AMD64_NO_SSE) #if !defined(LIBCO_NO_SSE)
0x0f, 0x29, 0x72, 0x50, /* movaps [rdx+ 80],xmm6 */ 0x0f, 0x29, 0x72, 0x50, /* movaps [rdx+ 80],xmm6 */
0x0f, 0x29, 0x7a, 0x60, /* movaps [rdx+ 96],xmm7 */ 0x0f, 0x29, 0x7a, 0x60, /* movaps [rdx+ 96],xmm7 */
0x44, 0x0f, 0x29, 0x42, 0x70, /* movaps [rdx+112],xmm8 */ 0x44, 0x0f, 0x29, 0x42, 0x70, /* movaps [rdx+112],xmm8 */
@@ -57,7 +59,7 @@ static void (*co_swap)(cothread_t, cothread_t) = 0;
0x4c, 0x8b, 0x69, 0x30, /* mov r13,[rcx+48] */ 0x4c, 0x8b, 0x69, 0x30, /* mov r13,[rcx+48] */
0x4c, 0x8b, 0x71, 0x38, /* mov r14,[rcx+56] */ 0x4c, 0x8b, 0x71, 0x38, /* mov r14,[rcx+56] */
0x4c, 0x8b, 0x79, 0x40, /* mov r15,[rcx+64] */ 0x4c, 0x8b, 0x79, 0x40, /* mov r15,[rcx+64] */
#if !defined(LIBCO_AMD64_NO_SSE) #if !defined(LIBCO_NO_SSE)
0x0f, 0x28, 0x71, 0x50, /* movaps xmm6, [rcx+ 80] */ 0x0f, 0x28, 0x71, 0x50, /* movaps xmm6, [rcx+ 80] */
0x0f, 0x28, 0x79, 0x60, /* movaps xmm7, [rcx+ 96] */ 0x0f, 0x28, 0x79, 0x60, /* movaps xmm7, [rcx+ 96] */
0x44, 0x0f, 0x28, 0x41, 0x70, /* movaps xmm8, [rcx+112] */ 0x44, 0x0f, 0x28, 0x41, 0x70, /* movaps xmm8, [rcx+112] */
@@ -75,13 +77,15 @@ static void (*co_swap)(cothread_t, cothread_t) = 0;
#include <windows.h> #include <windows.h>
void co_init() { static void co_init() {
#ifdef LIBCO_MPROTECT
DWORD old_privileges; DWORD old_privileges;
VirtualProtect(co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READWRITE, &old_privileges); VirtualProtect((void*)co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READ, &old_privileges);
#endif
} }
#else #else
/* ABI: SystemV */ /* ABI: SystemV */
static unsigned char co_swap_function[] = { static const unsigned char co_swap_function[4096] = {
0x48, 0x89, 0x26, /* mov [rsi],rsp */ 0x48, 0x89, 0x26, /* mov [rsi],rsp */
0x48, 0x8b, 0x27, /* mov rsp,[rdi] */ 0x48, 0x8b, 0x27, /* mov rsp,[rdi] */
0x58, /* pop rax */ 0x58, /* pop rax */
@@ -103,11 +107,13 @@ static void (*co_swap)(cothread_t, cothread_t) = 0;
#include <unistd.h> #include <unistd.h>
#include <sys/mman.h> #include <sys/mman.h>
void co_init() { static void co_init() {
#ifdef LIBCO_MPROTECT
unsigned long long addr = (unsigned long long)co_swap_function; unsigned long long addr = (unsigned long long)co_swap_function;
unsigned long long base = addr - (addr % sysconf(_SC_PAGESIZE)); unsigned long long base = addr - (addr % sysconf(_SC_PAGESIZE));
unsigned long long size = (addr - base) + sizeof co_swap_function; unsigned long long size = (addr - base) + sizeof co_swap_function;
mprotect((void*)base, size, PROT_READ | PROT_WRITE | PROT_EXEC); mprotect((void*)base, size, PROT_READ | PROT_EXEC);
#endif
} }
#endif #endif

16
arm.c
View File

@@ -1,11 +1,12 @@
/* /*
libco.arm (2015-06-18) libco.arm (2016-09-14)
author: byuu author: byuu
license: public domain license: public domain
*/ */
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
#include "settings.h"
#include <assert.h> #include <assert.h>
#include <stdlib.h> #include <stdlib.h>
@@ -20,17 +21,24 @@ static thread_local unsigned long co_active_buffer[64];
static thread_local cothread_t co_active_handle = 0; static thread_local cothread_t co_active_handle = 0;
static void (*co_swap)(cothread_t, cothread_t) = 0; static void (*co_swap)(cothread_t, cothread_t) = 0;
static unsigned long co_swap_function[] = { #ifdef LIBCO_MPROTECT
alignas(4096)
#else
section(text)
#endif
static const unsigned long co_swap_function[1024] = {
0xe8a16ff0, /* stmia r1!, {r4-r11,sp,lr} */ 0xe8a16ff0, /* stmia r1!, {r4-r11,sp,lr} */
0xe8b0aff0, /* ldmia r0!, {r4-r11,sp,pc} */ 0xe8b0aff0, /* ldmia r0!, {r4-r11,sp,pc} */
0xe12fff1e, /* bx lr */ 0xe12fff1e, /* bx lr */
}; };
void co_init() { static void co_init() {
#ifdef LIBCO_MPROTECT
unsigned long addr = (unsigned long)co_swap_function; unsigned long addr = (unsigned long)co_swap_function;
unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE)); unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE));
unsigned long size = (addr - base) + sizeof co_swap_function; unsigned long size = (addr - base) + sizeof co_swap_function;
mprotect((void*)base, size, PROT_READ | PROT_WRITE | PROT_EXEC); mprotect((void*)base, size, PROT_READ | PROT_EXEC);
#endif
} }
cothread_t co_active() { cothread_t co_active() {

2
fiber.c
View File

@@ -17,7 +17,7 @@ extern "C" {
static thread_local cothread_t co_active_ = 0; static thread_local cothread_t co_active_ = 0;
static void __stdcall co_thunk(void *coentry) { static void __stdcall co_thunk(void* coentry) {
((void (*)(void))coentry)(); ((void (*)(void))coentry)();
} }

11
libco.h
View File

@@ -1,6 +1,5 @@
/* /*
libco libco v18 (2016-09-14)
version: 0.17 (2015-06-18)
author: byuu author: byuu
license: public domain license: public domain
*/ */
@@ -8,14 +7,6 @@
#ifndef LIBCO_H #ifndef LIBCO_H
#define LIBCO_H #define LIBCO_H
#ifdef LIBCO_C
#ifdef LIBCO_MP
#define thread_local __thread
#else
#define thread_local
#endif
#endif
#ifdef __cplusplus #ifdef __cplusplus
extern "C" { extern "C" {
#endif #endif

675
ppc.c
View File

@@ -1,37 +1,33 @@
/* /*
libco.ppc (2010-10-17) libco.ppc (2016-09-14)
author: blargg author: blargg
license: public domain license: public domain
*/ */
/* PowerPC 32/64 using embedded or external asm, with optional
floating-point and AltiVec save/restore */
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
#include "settings.h"
#include <stdlib.h> #include <stdlib.h>
#include <stdint.h> #include <stdint.h>
#include <string.h> #include <string.h>
#define LIBCO_MPROTECT (__unix__ && !LIBCO_PPC_ASM)
#if LIBCO_MPROTECT #if LIBCO_MPROTECT
#include <unistd.h> #include <unistd.h>
#include <sys/mman.h> #include <sys/mman.h>
#endif #endif
/* State format (offsets in 32-bit words) /* state format (offsets in 32-bit words)
+0 Pointer to swap code +0 pointer to swap code
Rest of function descriptor for entry function rest of function descriptor for entry function
+8 PC +8 PC
+10 SP +10 SP
Special regs special registers
GPRs GPRs
FPRs FPRs
VRs VRs
stack stack
*/ */
enum { state_size = 1024 }; enum { state_size = 1024 };
@@ -40,369 +36,332 @@ enum { stack_align = 256 };
static thread_local cothread_t co_active_handle = 0; static thread_local cothread_t co_active_handle = 0;
/**** Determine environment ****/ /* determine environment */
#define LIBCO_PPC64 (_ARCH_PPC64 || __PPC64__ || __ppc64__ || __powerpc64__) #define LIBCO_PPC64 (_ARCH_PPC64 || __PPC64__ || __ppc64__ || __powerpc64__)
/* Whether function calls are indirect through a descriptor, /* whether function calls are indirect through a descriptor, or are directly to function */
or are directly to function */
#ifndef LIBCO_PPCDESC #ifndef LIBCO_PPCDESC
#if !_CALL_SYSV && (_CALL_AIX || _CALL_AIXDESC || LIBCO_PPC64) #if !_CALL_SYSV && (_CALL_AIX || _CALL_AIXDESC || LIBCO_PPC64)
#define LIBCO_PPCDESC 1 #define LIBCO_PPCDESC 1
#endif #endif
#endif #endif
#ifdef LIBCO_PPC_ASM #ifdef LIBCO_MPROTECT
alignas(4096)
#ifdef __cplusplus
extern "C"
#endif
/* Swap code is in ppc.S */
void co_swap_asm( cothread_t, cothread_t );
#define CO_SWAP_ASM( x, y ) co_swap_asm( x, y )
#else #else
section(text)
/* Swap code is here in array. Please leave dieassembly comments,
as they make it easy to see what it does, and reorder instructions
if one wants to see whether that improves performance. */
static const uint32_t libco_ppc_code [] = {
#if LIBCO_PPC64
0x7d000026, /* mfcr r8 */
0xf8240028, /* std r1,40(r4) */
0x7d2802a6, /* mflr r9 */
0xf9c40048, /* std r14,72(r4) */
0xf9e40050, /* std r15,80(r4) */
0xfa040058, /* std r16,88(r4) */
0xfa240060, /* std r17,96(r4) */
0xfa440068, /* std r18,104(r4) */
0xfa640070, /* std r19,112(r4) */
0xfa840078, /* std r20,120(r4) */
0xfaa40080, /* std r21,128(r4) */
0xfac40088, /* std r22,136(r4) */
0xfae40090, /* std r23,144(r4) */
0xfb040098, /* std r24,152(r4) */
0xfb2400a0, /* std r25,160(r4) */
0xfb4400a8, /* std r26,168(r4) */
0xfb6400b0, /* std r27,176(r4) */
0xfb8400b8, /* std r28,184(r4) */
0xfba400c0, /* std r29,192(r4) */
0xfbc400c8, /* std r30,200(r4) */
0xfbe400d0, /* std r31,208(r4) */
0xf9240020, /* std r9,32(r4) */
0xe8e30020, /* ld r7,32(r3) */
0xe8230028, /* ld r1,40(r3) */
0x48000009, /* bl 1 */
0x7fe00008, /* trap */
0x91040030,/*1:stw r8,48(r4) */
0x80c30030, /* lwz r6,48(r3) */
0x7ce903a6, /* mtctr r7 */
0xe9c30048, /* ld r14,72(r3) */
0xe9e30050, /* ld r15,80(r3) */
0xea030058, /* ld r16,88(r3) */
0xea230060, /* ld r17,96(r3) */
0xea430068, /* ld r18,104(r3) */
0xea630070, /* ld r19,112(r3) */
0xea830078, /* ld r20,120(r3) */
0xeaa30080, /* ld r21,128(r3) */
0xeac30088, /* ld r22,136(r3) */
0xeae30090, /* ld r23,144(r3) */
0xeb030098, /* ld r24,152(r3) */
0xeb2300a0, /* ld r25,160(r3) */
0xeb4300a8, /* ld r26,168(r3) */
0xeb6300b0, /* ld r27,176(r3) */
0xeb8300b8, /* ld r28,184(r3) */
0xeba300c0, /* ld r29,192(r3) */
0xebc300c8, /* ld r30,200(r3) */
0xebe300d0, /* ld r31,208(r3) */
0x7ccff120, /* mtcr r6 */
#else
0x7d000026, /* mfcr r8 */
0x90240028, /* stw r1,40(r4) */
0x7d2802a6, /* mflr r9 */
0x91a4003c, /* stw r13,60(r4) */
0x91c40040, /* stw r14,64(r4) */
0x91e40044, /* stw r15,68(r4) */
0x92040048, /* stw r16,72(r4) */
0x9224004c, /* stw r17,76(r4) */
0x92440050, /* stw r18,80(r4) */
0x92640054, /* stw r19,84(r4) */
0x92840058, /* stw r20,88(r4) */
0x92a4005c, /* stw r21,92(r4) */
0x92c40060, /* stw r22,96(r4) */
0x92e40064, /* stw r23,100(r4) */
0x93040068, /* stw r24,104(r4) */
0x9324006c, /* stw r25,108(r4) */
0x93440070, /* stw r26,112(r4) */
0x93640074, /* stw r27,116(r4) */
0x93840078, /* stw r28,120(r4) */
0x93a4007c, /* stw r29,124(r4) */
0x93c40080, /* stw r30,128(r4) */
0x93e40084, /* stw r31,132(r4) */
0x91240020, /* stw r9,32(r4) */
0x80e30020, /* lwz r7,32(r3) */
0x80230028, /* lwz r1,40(r3) */
0x48000009, /* bl 1 */
0x7fe00008, /* trap */
0x91040030,/*1:stw r8,48(r4) */
0x80c30030, /* lwz r6,48(r3) */
0x7ce903a6, /* mtctr r7 */
0x81a3003c, /* lwz r13,60(r3) */
0x81c30040, /* lwz r14,64(r3) */
0x81e30044, /* lwz r15,68(r3) */
0x82030048, /* lwz r16,72(r3) */
0x8223004c, /* lwz r17,76(r3) */
0x82430050, /* lwz r18,80(r3) */
0x82630054, /* lwz r19,84(r3) */
0x82830058, /* lwz r20,88(r3) */
0x82a3005c, /* lwz r21,92(r3) */
0x82c30060, /* lwz r22,96(r3) */
0x82e30064, /* lwz r23,100(r3) */
0x83030068, /* lwz r24,104(r3) */
0x8323006c, /* lwz r25,108(r3) */
0x83430070, /* lwz r26,112(r3) */
0x83630074, /* lwz r27,116(r3) */
0x83830078, /* lwz r28,120(r3) */
0x83a3007c, /* lwz r29,124(r3) */
0x83c30080, /* lwz r30,128(r3) */
0x83e30084, /* lwz r31,132(r3) */
0x7ccff120, /* mtcr r6 */
#endif #endif
static const uint32_t libco_ppc_code[1024] = {
#if LIBCO_PPC64
0x7d000026, /* mfcr r8 */
0xf8240028, /* std r1,40(r4) */
0x7d2802a6, /* mflr r9 */
0xf9c40048, /* std r14,72(r4) */
0xf9e40050, /* std r15,80(r4) */
0xfa040058, /* std r16,88(r4) */
0xfa240060, /* std r17,96(r4) */
0xfa440068, /* std r18,104(r4) */
0xfa640070, /* std r19,112(r4) */
0xfa840078, /* std r20,120(r4) */
0xfaa40080, /* std r21,128(r4) */
0xfac40088, /* std r22,136(r4) */
0xfae40090, /* std r23,144(r4) */
0xfb040098, /* std r24,152(r4) */
0xfb2400a0, /* std r25,160(r4) */
0xfb4400a8, /* std r26,168(r4) */
0xfb6400b0, /* std r27,176(r4) */
0xfb8400b8, /* std r28,184(r4) */
0xfba400c0, /* std r29,192(r4) */
0xfbc400c8, /* std r30,200(r4) */
0xfbe400d0, /* std r31,208(r4) */
0xf9240020, /* std r9,32(r4) */
0xe8e30020, /* ld r7,32(r3) */
0xe8230028, /* ld r1,40(r3) */
0x48000009, /* bl 1 */
0x7fe00008, /* trap */
0x91040030, /*1:stw r8,48(r4) */
0x80c30030, /* lwz r6,48(r3) */
0x7ce903a6, /* mtctr r7 */
0xe9c30048, /* ld r14,72(r3) */
0xe9e30050, /* ld r15,80(r3) */
0xea030058, /* ld r16,88(r3) */
0xea230060, /* ld r17,96(r3) */
0xea430068, /* ld r18,104(r3) */
0xea630070, /* ld r19,112(r3) */
0xea830078, /* ld r20,120(r3) */
0xeaa30080, /* ld r21,128(r3) */
0xeac30088, /* ld r22,136(r3) */
0xeae30090, /* ld r23,144(r3) */
0xeb030098, /* ld r24,152(r3) */
0xeb2300a0, /* ld r25,160(r3) */
0xeb4300a8, /* ld r26,168(r3) */
0xeb6300b0, /* ld r27,176(r3) */
0xeb8300b8, /* ld r28,184(r3) */
0xeba300c0, /* ld r29,192(r3) */
0xebc300c8, /* ld r30,200(r3) */
0xebe300d0, /* ld r31,208(r3) */
0x7ccff120, /* mtcr r6 */
#else
0x7d000026, /* mfcr r8 */
0x90240028, /* stw r1,40(r4) */
0x7d2802a6, /* mflr r9 */
0x91a4003c, /* stw r13,60(r4) */
0x91c40040, /* stw r14,64(r4) */
0x91e40044, /* stw r15,68(r4) */
0x92040048, /* stw r16,72(r4) */
0x9224004c, /* stw r17,76(r4) */
0x92440050, /* stw r18,80(r4) */
0x92640054, /* stw r19,84(r4) */
0x92840058, /* stw r20,88(r4) */
0x92a4005c, /* stw r21,92(r4) */
0x92c40060, /* stw r22,96(r4) */
0x92e40064, /* stw r23,100(r4) */
0x93040068, /* stw r24,104(r4) */
0x9324006c, /* stw r25,108(r4) */
0x93440070, /* stw r26,112(r4) */
0x93640074, /* stw r27,116(r4) */
0x93840078, /* stw r28,120(r4) */
0x93a4007c, /* stw r29,124(r4) */
0x93c40080, /* stw r30,128(r4) */
0x93e40084, /* stw r31,132(r4) */
0x91240020, /* stw r9,32(r4) */
0x80e30020, /* lwz r7,32(r3) */
0x80230028, /* lwz r1,40(r3) */
0x48000009, /* bl 1 */
0x7fe00008, /* trap */
0x91040030, /*1:stw r8,48(r4) */
0x80c30030, /* lwz r6,48(r3) */
0x7ce903a6, /* mtctr r7 */
0x81a3003c, /* lwz r13,60(r3) */
0x81c30040, /* lwz r14,64(r3) */
0x81e30044, /* lwz r15,68(r3) */
0x82030048, /* lwz r16,72(r3) */
0x8223004c, /* lwz r17,76(r3) */
0x82430050, /* lwz r18,80(r3) */
0x82630054, /* lwz r19,84(r3) */
0x82830058, /* lwz r20,88(r3) */
0x82a3005c, /* lwz r21,92(r3) */
0x82c30060, /* lwz r22,96(r3) */
0x82e30064, /* lwz r23,100(r3) */
0x83030068, /* lwz r24,104(r3) */
0x8323006c, /* lwz r25,108(r3) */
0x83430070, /* lwz r26,112(r3) */
0x83630074, /* lwz r27,116(r3) */
0x83830078, /* lwz r28,120(r3) */
0x83a3007c, /* lwz r29,124(r3) */
0x83c30080, /* lwz r30,128(r3) */
0x83e30084, /* lwz r31,132(r3) */
0x7ccff120, /* mtcr r6 */
#endif
#ifndef LIBCO_PPC_NOFP #ifndef LIBCO_PPC_NOFP
0xd9c400e0, /* stfd f14,224(r4) */ 0xd9c400e0, /* stfd f14,224(r4) */
0xd9e400e8, /* stfd f15,232(r4) */ 0xd9e400e8, /* stfd f15,232(r4) */
0xda0400f0, /* stfd f16,240(r4) */ 0xda0400f0, /* stfd f16,240(r4) */
0xda2400f8, /* stfd f17,248(r4) */ 0xda2400f8, /* stfd f17,248(r4) */
0xda440100, /* stfd f18,256(r4) */ 0xda440100, /* stfd f18,256(r4) */
0xda640108, /* stfd f19,264(r4) */ 0xda640108, /* stfd f19,264(r4) */
0xda840110, /* stfd f20,272(r4) */ 0xda840110, /* stfd f20,272(r4) */
0xdaa40118, /* stfd f21,280(r4) */ 0xdaa40118, /* stfd f21,280(r4) */
0xdac40120, /* stfd f22,288(r4) */ 0xdac40120, /* stfd f22,288(r4) */
0xdae40128, /* stfd f23,296(r4) */ 0xdae40128, /* stfd f23,296(r4) */
0xdb040130, /* stfd f24,304(r4) */ 0xdb040130, /* stfd f24,304(r4) */
0xdb240138, /* stfd f25,312(r4) */ 0xdb240138, /* stfd f25,312(r4) */
0xdb440140, /* stfd f26,320(r4) */ 0xdb440140, /* stfd f26,320(r4) */
0xdb640148, /* stfd f27,328(r4) */ 0xdb640148, /* stfd f27,328(r4) */
0xdb840150, /* stfd f28,336(r4) */ 0xdb840150, /* stfd f28,336(r4) */
0xdba40158, /* stfd f29,344(r4) */ 0xdba40158, /* stfd f29,344(r4) */
0xdbc40160, /* stfd f30,352(r4) */ 0xdbc40160, /* stfd f30,352(r4) */
0xdbe40168, /* stfd f31,360(r4) */ 0xdbe40168, /* stfd f31,360(r4) */
0xc9c300e0, /* lfd f14,224(r3) */ 0xc9c300e0, /* lfd f14,224(r3) */
0xc9e300e8, /* lfd f15,232(r3) */ 0xc9e300e8, /* lfd f15,232(r3) */
0xca0300f0, /* lfd f16,240(r3) */ 0xca0300f0, /* lfd f16,240(r3) */
0xca2300f8, /* lfd f17,248(r3) */ 0xca2300f8, /* lfd f17,248(r3) */
0xca430100, /* lfd f18,256(r3) */ 0xca430100, /* lfd f18,256(r3) */
0xca630108, /* lfd f19,264(r3) */ 0xca630108, /* lfd f19,264(r3) */
0xca830110, /* lfd f20,272(r3) */ 0xca830110, /* lfd f20,272(r3) */
0xcaa30118, /* lfd f21,280(r3) */ 0xcaa30118, /* lfd f21,280(r3) */
0xcac30120, /* lfd f22,288(r3) */ 0xcac30120, /* lfd f22,288(r3) */
0xcae30128, /* lfd f23,296(r3) */ 0xcae30128, /* lfd f23,296(r3) */
0xcb030130, /* lfd f24,304(r3) */ 0xcb030130, /* lfd f24,304(r3) */
0xcb230138, /* lfd f25,312(r3) */ 0xcb230138, /* lfd f25,312(r3) */
0xcb430140, /* lfd f26,320(r3) */ 0xcb430140, /* lfd f26,320(r3) */
0xcb630148, /* lfd f27,328(r3) */ 0xcb630148, /* lfd f27,328(r3) */
0xcb830150, /* lfd f28,336(r3) */ 0xcb830150, /* lfd f28,336(r3) */
0xcba30158, /* lfd f29,344(r3) */ 0xcba30158, /* lfd f29,344(r3) */
0xcbc30160, /* lfd f30,352(r3) */ 0xcbc30160, /* lfd f30,352(r3) */
0xcbe30168, /* lfd f31,360(r3) */ 0xcbe30168, /* lfd f31,360(r3) */
#endif #endif
#ifdef __ALTIVEC__ #ifdef __ALTIVEC__
0x7ca042a6, /* mfvrsave r5 */ 0x7ca042a6, /* mfvrsave r5 */
0x39040180, /* addi r8,r4,384 */ 0x39040180, /* addi r8,r4,384 */
0x39240190, /* addi r9,r4,400 */ 0x39240190, /* addi r9,r4,400 */
0x70a00fff, /* andi. r0,r5,4095 */ 0x70a00fff, /* andi. r0,r5,4095 */
0x90a40034, /* stw r5,52(r4) */ 0x90a40034, /* stw r5,52(r4) */
0x4182005c, /* beq- 2 */ 0x4182005c, /* beq- 2 */
0x7e8041ce, /* stvx v20,r0,r8 */ 0x7e8041ce, /* stvx v20,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7ea049ce, /* stvx v21,r0,r9 */ 0x7ea049ce, /* stvx v21,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7ec041ce, /* stvx v22,r0,r8 */ 0x7ec041ce, /* stvx v22,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7ee049ce, /* stvx v23,r0,r9 */ 0x7ee049ce, /* stvx v23,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f0041ce, /* stvx v24,r0,r8 */ 0x7f0041ce, /* stvx v24,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7f2049ce, /* stvx v25,r0,r9 */ 0x7f2049ce, /* stvx v25,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f4041ce, /* stvx v26,r0,r8 */ 0x7f4041ce, /* stvx v26,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7f6049ce, /* stvx v27,r0,r9 */ 0x7f6049ce, /* stvx v27,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f8041ce, /* stvx v28,r0,r8 */ 0x7f8041ce, /* stvx v28,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7fa049ce, /* stvx v29,r0,r9 */ 0x7fa049ce, /* stvx v29,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7fc041ce, /* stvx v30,r0,r8 */ 0x7fc041ce, /* stvx v30,r0,r8 */
0x7fe049ce, /* stvx v31,r0,r9 */ 0x7fe049ce, /* stvx v31,r0,r9 */
0x80a30034,/*2:lwz r5,52(r3) */ 0x80a30034, /*2:lwz r5,52(r3) */
0x39030180, /* addi r8,r3,384 */ 0x39030180, /* addi r8,r3,384 */
0x39230190, /* addi r9,r3,400 */ 0x39230190, /* addi r9,r3,400 */
0x70a00fff, /* andi. r0,r5,4095 */ 0x70a00fff, /* andi. r0,r5,4095 */
0x7ca043a6, /* mtvrsave r5 */ 0x7ca043a6, /* mtvrsave r5 */
0x4d820420, /* beqctr */ 0x4d820420, /* beqctr */
0x7e8040ce, /* lvx v20,r0,r8 */ 0x7e8040ce, /* lvx v20,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7ea048ce, /* lvx v21,r0,r9 */ 0x7ea048ce, /* lvx v21,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7ec040ce, /* lvx v22,r0,r8 */ 0x7ec040ce, /* lvx v22,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7ee048ce, /* lvx v23,r0,r9 */ 0x7ee048ce, /* lvx v23,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f0040ce, /* lvx v24,r0,r8 */ 0x7f0040ce, /* lvx v24,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7f2048ce, /* lvx v25,r0,r9 */ 0x7f2048ce, /* lvx v25,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f4040ce, /* lvx v26,r0,r8 */ 0x7f4040ce, /* lvx v26,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7f6048ce, /* lvx v27,r0,r9 */ 0x7f6048ce, /* lvx v27,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7f8040ce, /* lvx v28,r0,r8 */ 0x7f8040ce, /* lvx v28,r0,r8 */
0x39080020, /* addi r8,r8,32 */ 0x39080020, /* addi r8,r8,32 */
0x7fa048ce, /* lvx v29,r0,r9 */ 0x7fa048ce, /* lvx v29,r0,r9 */
0x39290020, /* addi r9,r9,32 */ 0x39290020, /* addi r9,r9,32 */
0x7fc040ce, /* lvx v30,r0,r8 */ 0x7fc040ce, /* lvx v30,r0,r8 */
0x7fe048ce, /* lvx v31,r0,r9 */ 0x7fe048ce, /* lvx v31,r0,r9 */
#endif #endif
0x4e800420, /* bctr */ 0x4e800420, /* bctr */
}; };
#if LIBCO_PPCDESC #if LIBCO_PPCDESC
/* Function call goes through indirect descriptor */ /* function call goes through indirect descriptor */
#define CO_SWAP_ASM( x, y ) \ #define CO_SWAP_ASM(x, y) ((void (*)(cothread_t, cothread_t))(uintptr_t)x)(x, y)
((void (*)( cothread_t, cothread_t )) (uintptr_t) x)( x, y ) #else
#else /* function call goes directly to code */
/* Function call goes directly to code */ #define CO_SWAP_ASM(x, y) ((void (*)(cothread_t, cothread_t))(uintptr_t)libco_ppc_code)(x, y)
#define CO_SWAP_ASM( x, y ) \
((void (*)( cothread_t, cothread_t )) (uintptr_t) libco_ppc_code)( x, y )
#endif
#endif #endif
static uint32_t* co_create_( unsigned size, uintptr_t entry ) static uint32_t* co_create_(unsigned size, uintptr_t entry) {
{ (void)entry;
uint32_t* t = (uint32_t*) malloc( size );
uint32_t* t = (uint32_t*)malloc(size);
(void) entry;
#if LIBCO_PPCDESC
#if LIBCO_PPCDESC if(t) {
if ( t ) memcpy(t, (void*)entry, sizeof(void*) * 3); /* copy entry's descriptor */
{ *(const void**)t = libco_ppc_code; /* set function pointer to swap routine */
/* Copy entry's descriptor */ }
memcpy( t, (void*) entry, sizeof (void*) * 3 ); #endif
/* Set function pointer to swap routine */ return t;
#ifdef LIBCO_PPC_ASM
*(const void**) t = *(void**) &co_swap_asm;
#else
*(const void**) t = libco_ppc_code;
#endif
}
#endif
return t;
} }
cothread_t co_create( unsigned int size, void (*entry_)( void ) ) cothread_t co_create(unsigned int size, void (*entry_)(void)) {
{ uintptr_t entry = (uintptr_t)entry_;
uintptr_t entry = (uintptr_t) entry_; uint32_t* t = 0;
uint32_t* t = NULL;
/* be sure main thread was successfully allocated */
/* Be sure main thread was successfully allocated */ if(co_active()) {
if ( co_active() ) size += state_size + above_stack + stack_align;
{ t = co_create_(size, entry);
size += state_size + above_stack + stack_align; }
t = co_create_( size, entry );
} if(t) {
uintptr_t sp;
if ( t ) int shift;
{
uintptr_t sp; /* save current registers into new thread, so that any special ones will have proper values when thread is begun */
int shift; CO_SWAP_ASM(t, t);
/* Save current registers into new thread, so that any special ones will #if LIBCO_PPCDESC
have proper values when thread is begun */ entry = (uintptr_t)*(void**)entry; /* get real address */
CO_SWAP_ASM( t, t ); #endif
#if LIBCO_PPCDESC /* put stack near end of block, and align */
/* Get real address */ sp = (uintptr_t)t + size - above_stack;
entry = (uintptr_t) *(void**) entry; sp -= sp % stack_align;
#endif
/* on PPC32, we save and restore GPRs as 32 bits. for PPC64, we
/* Put stack near end of block, and align */ save and restore them as 64 bits, regardless of the size the ABI
sp = (uintptr_t) t + size - above_stack; uses. so, we manually write pointers at the proper size. we always
sp -= sp % stack_align; save and restore at the same address, and since PPC is big-endian,
we must put the low byte first on PPC32. */
/* On PPC32, we save and restore GPRs as 32 bits. For PPC64, we
save and restore them as 64 bits, regardless of the size the ABI /* if uintptr_t is 32 bits, >>32 is undefined behavior,
uses. So, we manually write pointers at the proper size. We always so we do two shifts and don't have to care how many bits uintptr_t is. */
save and restore at the same address, and since PPC is big-endian, #if LIBCO_PPC64
we must put the low byte first on PPC32. */ shift = 16;
#else
/* If uintptr_t is 32 bits, >>32 is undefined behavior, so we do two shifts shift = 0;
and don't have to care how many bits uintptr_t is. */ #endif
#if LIBCO_PPC64
shift = 16; /* set up so entry will be called on next swap */
#else t[ 8] = (uint32_t)(entry >> shift >> shift);
shift = 0; t[ 9] = (uint32_t)entry;
#endif
t[10] = (uint32_t)(sp >> shift >> shift);
/* Set up so entry will be called on next swap */ t[11] = (uint32_t)sp;
t [8] = (uint32_t) (entry >> shift >> shift); }
t [9] = (uint32_t) entry;
return t;
t [10] = (uint32_t) (sp >> shift >> shift);
t [11] = (uint32_t) sp;
}
return t;
} }
void co_delete( cothread_t t ) void co_delete(cothread_t t) {
{ free(t);
free( t );
} }
static void co_init_( void ) static void co_init_(void) {
{ #if LIBCO_MPROTECT
#if LIBCO_MPROTECT long page_size = sysconf(_SC_PAGESIZE);
/* TODO: pre- and post-pad PPC code so that this doesn't make other if(page_size > 0) {
data executable and writable */ uintptr_t align = page_size;
long page_size = sysconf( _SC_PAGESIZE ); uintptr_t begin = (uintptr_t)libco_ppc_code;
if ( page_size > 0 ) uintptr_t end = begin + sizeof libco_ppc_code;
{
uintptr_t align = page_size; /* align beginning and end */
uintptr_t begin = (uintptr_t) libco_ppc_code; end += align - 1;
uintptr_t end = begin + sizeof libco_ppc_code; end -= end % align;
begin -= begin % align;
/* Align beginning and end */
end += align - 1; mprotect((void*)begin, end - begin, PROT_READ | PROT_EXEC);
end -= end % align; }
begin -= begin % align; #endif
mprotect( (void*) begin, end - begin, PROT_READ | PROT_WRITE | PROT_EXEC ); co_active_handle = co_create_(state_size, (uintptr_t)&co_switch);
}
#endif
co_active_handle = co_create_( state_size, (uintptr_t) &co_switch );
} }
cothread_t co_active() cothread_t co_active() {
{ if(!co_active_handle) co_init_();
if ( !co_active_handle )
co_init_(); return co_active_handle;
return co_active_handle;
} }
void co_switch( cothread_t t ) void co_switch(cothread_t t) {
{ cothread_t old = co_active_handle;
cothread_t old = co_active_handle; co_active_handle = t;
co_active_handle = t;
CO_SWAP_ASM(t, old);
CO_SWAP_ASM( t, old );
} }

36
settings.h Normal file
View File

@@ -0,0 +1,36 @@
#ifdef LIBCO_C
/*[amd64, arm, ppc, x86]:
by default, co_swap_function is marked as a text (code) section
if not supported, uncomment the below line to use mprotect instead */
/* #define LIBCO_MPROTECT */
/*[amd64]:
Win64 only: provides a substantial speed-up, but will thrash XMM regs
do not use this unless you are certain your application won't use SSE */
/* #define LIBCO_NO_SSE */
#ifdef LIBCO_C
#ifdef LIBCO_MP
#define thread_local __thread
#else
#define thread_local
#endif
#endif
#if __STDC_VERSION__ >= 201112L
#ifndef _MSC_VER
#include <stdalign.h>
#endif
#else
#define alignas(bytes)
#endif
#ifndef _MSC_VER
#define section(name) __attribute__((section("." #name "#")))
#else
#define section(name) __declspec(allocate("." #name))
#endif
/* ifdef LIBCO_C */
#endif

13
sjlj.c
View File

@@ -5,11 +5,9 @@
*/ */
/* /*
* Note this was designed for UNIX systems. Based on ideas expressed in a paper note this was designed for UNIX systems. Based on ideas expressed in a paper by Ralf Engelschall.
* by Ralf Engelschall. for SJLJ on other systems, one would want to rewrite springboard() and co_create() and hack the jmb_buf stack pointer.
* For SJLJ on other systems, one would want to rewrite springboard() and */
* co_create() and hack the jmb_buf stack pointer.
*/
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
@@ -25,11 +23,12 @@ extern "C" {
typedef struct { typedef struct {
sigjmp_buf context; sigjmp_buf context;
void (*coentry)(void); void (*coentry)(void);
void *stack; void* stack;
} cothread_struct; } cothread_struct;
static thread_local cothread_struct co_primary; static thread_local cothread_struct co_primary;
static thread_local cothread_struct *creating, *co_running = 0; static thread_local cothread_struct* creating;
static thread_local cothread_struct* co_running = 0;
static void springboard(int ignored) { static void springboard(int ignored) {
if(sigsetjmp(creating->context, 0)) { if(sigsetjmp(creating->context, 0)) {

26
ucontext.c
View File

@@ -5,16 +5,16 @@
*/ */
/* /*
* WARNING: the overhead of POSIX ucontext is very high, WARNING: the overhead of POSIX ucontext is very high,
* assembly versions of libco or libco_sjlj should be much faster assembly versions of libco or libco_sjlj should be much faster
*
* This library only exists for two reasons: this library only exists for two reasons:
* 1 - as an initial test for the viability of a ucontext implementation 1: as an initial test for the viability of a ucontext implementation
* 2 - to demonstrate the power and speed of libco over existing implementations, 2: to demonstrate the power and speed of libco over existing implementations,
* such as pth (which defaults to wrapping ucontext on unix targets) such as pth (which defaults to wrapping ucontext on unix targets)
*
* Use this library only as a *last resort* use this library only as a *last resort*
*/ */
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
@@ -28,7 +28,7 @@ extern "C" {
#endif #endif
static thread_local ucontext_t co_primary; static thread_local ucontext_t co_primary;
static thread_local ucontext_t *co_running = 0; static thread_local ucontext_t* co_running = 0;
cothread_t co_active() { cothread_t co_active() {
if(!co_running) co_running = &co_primary; if(!co_running) co_running = &co_primary;
@@ -37,7 +37,7 @@ cothread_t co_active() {
cothread_t co_create(unsigned int heapsize, void (*coentry)(void)) { cothread_t co_create(unsigned int heapsize, void (*coentry)(void)) {
if(!co_running) co_running = &co_primary; if(!co_running) co_running = &co_primary;
ucontext_t *thread = (ucontext_t*)malloc(sizeof(ucontext_t)); ucontext_t* thread = (ucontext_t*)malloc(sizeof(ucontext_t));
if(thread) { if(thread) {
if((!getcontext(thread) && !(thread->uc_stack.ss_sp = 0)) && (thread->uc_stack.ss_sp = malloc(heapsize))) { if((!getcontext(thread) && !(thread->uc_stack.ss_sp = 0)) && (thread->uc_stack.ss_sp = malloc(heapsize))) {
thread->uc_link = co_running; thread->uc_link = co_running;
@@ -59,7 +59,7 @@ void co_delete(cothread_t cothread) {
} }
void co_switch(cothread_t cothread) { void co_switch(cothread_t cothread) {
ucontext_t *old_thread = co_running; ucontext_t* old_thread = co_running;
co_running = (ucontext_t*)cothread; co_running = (ucontext_t*)cothread;
swapcontext(old_thread, co_running); swapcontext(old_thread, co_running);
} }

22
x86.c
View File

@@ -1,11 +1,12 @@
/* /*
libco.x86 (2009-10-12) libco.x86 (2016-09-14)
author: byuu author: byuu
license: public domain license: public domain
*/ */
#define LIBCO_C #define LIBCO_C
#include "libco.h" #include "libco.h"
#include "settings.h"
#include <assert.h> #include <assert.h>
#include <stdlib.h> #include <stdlib.h>
@@ -26,8 +27,13 @@ static thread_local long co_active_buffer[64];
static thread_local cothread_t co_active_handle = 0; static thread_local cothread_t co_active_handle = 0;
static void (fastcall *co_swap)(cothread_t, cothread_t) = 0; static void (fastcall *co_swap)(cothread_t, cothread_t) = 0;
#ifdef LIBCO_MPROTECT
alignas(4096)
#else
section(text)
#endif
/* ABI: fastcall */ /* ABI: fastcall */
static unsigned char co_swap_function[] = { static const unsigned char co_swap_function[4096] = {
0x89, 0x22, /* mov [edx],esp */ 0x89, 0x22, /* mov [edx],esp */
0x8b, 0x21, /* mov esp,[ecx] */ 0x8b, 0x21, /* mov esp,[ecx] */
0x58, /* pop eax */ 0x58, /* pop eax */
@@ -45,19 +51,23 @@ static unsigned char co_swap_function[] = {
#ifdef _WIN32 #ifdef _WIN32
#include <windows.h> #include <windows.h>
void co_init() { static void co_init() {
#ifdef LIBCO_MPROTECT
DWORD old_privileges; DWORD old_privileges;
VirtualProtect(co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READWRITE, &old_privileges); VirtualProtect((void*)co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READ, &old_privileges);
#endif
} }
#else #else
#include <unistd.h> #include <unistd.h>
#include <sys/mman.h> #include <sys/mman.h>
void co_init() { static void co_init() {
#ifdef LIBCO_MPROTECT
unsigned long addr = (unsigned long)co_swap_function; unsigned long addr = (unsigned long)co_swap_function;
unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE)); unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE));
unsigned long size = (addr - base) + sizeof co_swap_function; unsigned long size = (addr - base) + sizeof co_swap_function;
mprotect((void*)base, size, PROT_READ | PROT_WRITE | PROT_EXEC); mprotect((void*)base, size, PROT_READ | PROT_EXEC);
#endif
} }
#endif #endif