From 7871e50f7d48444f1a6333128308acb0a1c8ad2b Mon Sep 17 00:00:00 2001 From: MerryMage Date: Sun, 19 Apr 2020 17:35:46 +0100 Subject: [PATCH] First commit --- README.md | 27 +++ aarch64.c | 108 ++++++++++++ amd64.c | 165 ++++++++++++++++++ arm.c | 84 +++++++++ doc/style.css | 12 ++ doc/targets.html | 89 ++++++++++ doc/usage.html | 108 ++++++++++++ fiber.c | 55 ++++++ libco.c | 37 ++++ libco.h | 28 +++ ppc.c | 431 +++++++++++++++++++++++++++++++++++++++++++++++ ppc64v2.c | 279 ++++++++++++++++++++++++++++++ settings.h | 38 +++++ sjlj.c | 145 ++++++++++++++++ ucontext.c | 86 ++++++++++ x86.c | 119 +++++++++++++ 16 files changed, 1811 insertions(+) create mode 100644 README.md create mode 100644 aarch64.c create mode 100755 amd64.c create mode 100644 arm.c create mode 100755 doc/style.css create mode 100755 doc/targets.html create mode 100755 doc/usage.html create mode 100755 fiber.c create mode 100755 libco.c create mode 100755 libco.h create mode 100755 ppc.c create mode 100644 ppc64v2.c create mode 100644 settings.h create mode 100755 sjlj.c create mode 100755 ucontext.c create mode 100755 x86.c diff --git a/README.md b/README.md new file mode 100644 index 00000000..11e0c34e --- /dev/null +++ b/README.md @@ -0,0 +1,27 @@ +libco +----- + +libco is a cooperative multithreading library written in C89. + +Although cooperative multithreading is limited to a single CPU core, it scales substantially better than preemptive multithreading. + +For applications that need 100,000 or more context switches per second, the kernel overhead involved in preemptive multithreading can end up becoming the bottleneck in the application. libco can easily scale to 10,000,000 or more context switches per second. + +Ideal use cases include servers (HTTP, RDBMS) and emulators (CPU cores, etc.) + +It currently includes backends for: + +* x86 CPUs +* amd64 CPUs +* PowerPC CPUs +* PowerPC64 ELFv1 CPUs +* PowerPC64 ELFv2 CPUs +* ARM 32-bit CPUs +* ARM 64-bit (AArch64) CPUs +* POSIX platforms (setjmp) +* Windows platforms (fibers) + +License +======= + +libco is released under the ISC license. diff --git a/aarch64.c b/aarch64.c new file mode 100644 index 00000000..8505b675 --- /dev/null +++ b/aarch64.c @@ -0,0 +1,108 @@ +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include +#include +#ifdef LIBCO_MPROTECT + #include + #include +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +static thread_local unsigned long co_active_buffer[64]; +static thread_local cothread_t co_active_handle = 0; +static void (*co_swap)(cothread_t, cothread_t) = 0; + +#ifdef LIBCO_MPROTECT + alignas(4096) +#else + section(text) +#endif +static const uint32_t co_swap_function[1024] = { + 0x910003f0, /* mov x16,sp */ + 0xa9007830, /* stp x16,x30,[x1] */ + 0xa9407810, /* ldp x16,x30,[x0] */ + 0x9100021f, /* mov sp,x16 */ + 0xa9015033, /* stp x19,x20,[x1, 16] */ + 0xa9415013, /* ldp x19,x20,[x0, 16] */ + 0xa9025835, /* stp x21,x22,[x1, 32] */ + 0xa9425815, /* ldp x21,x22,[x0, 32] */ + 0xa9036037, /* stp x23,x24,[x1, 48] */ + 0xa9436017, /* ldp x23,x24,[x0, 48] */ + 0xa9046839, /* stp x25,x26,[x1, 64] */ + 0xa9446819, /* ldp x25,x26,[x0, 64] */ + 0xa905703b, /* stp x27,x28,[x1, 80] */ + 0xa945701b, /* ldp x27,x28,[x0, 80] */ + 0xf900303d, /* str x29, [x1, 96] */ + 0xf940301d, /* ldr x29, [x0, 96] */ + 0x6d072428, /* stp d8, d9, [x1,112] */ + 0x6d472408, /* ldp d8, d9, [x0,112] */ + 0x6d082c2a, /* stp d10,d11,[x1,128] */ + 0x6d482c0a, /* ldp d10,d11,[x0,128] */ + 0x6d09342c, /* stp d12,d13,[x1,144] */ + 0x6d49340c, /* ldp d12,d13,[x0,144] */ + 0x6d0a3c2e, /* stp d14,d15,[x1,160] */ + 0x6d4a3c0e, /* ldp d14,d15,[x0,160] */ + 0xd61f03c0, /* br x30 */ +}; + +static void co_init() { + #ifdef LIBCO_MPROTECT + unsigned long addr = (unsigned long)co_swap_function; + unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE)); + unsigned long size = (addr - base) + sizeof co_swap_function; + mprotect((void*)base, size, PROT_READ | PROT_EXEC); + #endif +} + +cothread_t co_active() { + if(!co_active_handle) co_active_handle = &co_active_buffer; + return co_active_handle; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*entrypoint)(void)) { + unsigned long* handle; + if(!co_swap) { + co_init(); + co_swap = (void (*)(cothread_t, cothread_t))co_swap_function; + } + if(!co_active_handle) co_active_handle = &co_active_buffer; + + if(handle = (unsigned long*)memory) { + unsigned int offset = (size & ~15); + unsigned long* p = (unsigned long*)((unsigned char*)handle + offset); + handle[0] = (unsigned long)p; /* x16 (stack pointer) */ + handle[1] = (unsigned long)entrypoint; /* x30 (link register) */ + handle[12] = (unsigned long)p; /* x29 (frame pointer) */ + } + + return handle; +} + +cothread_t co_create(unsigned int size, void (*entrypoint)(void)) { + void* memory = malloc(size); + if(!memory) return (cothread_t)0; + return co_derive(memory, size, entrypoint); +} + +void co_delete(cothread_t handle) { + free(handle); +} + +void co_switch(cothread_t handle) { + cothread_t co_previous_handle = co_active_handle; + co_swap(co_active_handle = handle, co_previous_handle); +} + +int co_serializable() { + return 1; +} + +#ifdef __cplusplus +} +#endif diff --git a/amd64.c b/amd64.c new file mode 100755 index 00000000..e96e5247 --- /dev/null +++ b/amd64.c @@ -0,0 +1,165 @@ +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +static thread_local long long co_active_buffer[64]; +static thread_local cothread_t co_active_handle = 0; +static void (*co_swap)(cothread_t, cothread_t) = 0; + +#ifdef LIBCO_MPROTECT + alignas(4096) +#else + section(text) +#endif +#ifdef _WIN32 + /* ABI: Win64 */ + static const unsigned char co_swap_function[4096] = { + 0x48, 0x89, 0x22, /* mov [rdx],rsp */ + 0x48, 0x8b, 0x21, /* mov rsp,[rcx] */ + 0x58, /* pop rax */ + 0x48, 0x89, 0x6a, 0x08, /* mov [rdx+ 8],rbp */ + 0x48, 0x89, 0x72, 0x10, /* mov [rdx+16],rsi */ + 0x48, 0x89, 0x7a, 0x18, /* mov [rdx+24],rdi */ + 0x48, 0x89, 0x5a, 0x20, /* mov [rdx+32],rbx */ + 0x4c, 0x89, 0x62, 0x28, /* mov [rdx+40],r12 */ + 0x4c, 0x89, 0x6a, 0x30, /* mov [rdx+48],r13 */ + 0x4c, 0x89, 0x72, 0x38, /* mov [rdx+56],r14 */ + 0x4c, 0x89, 0x7a, 0x40, /* mov [rdx+64],r15 */ + #if !defined(LIBCO_NO_SSE) + 0x0f, 0x29, 0x72, 0x50, /* movaps [rdx+ 80],xmm6 */ + 0x0f, 0x29, 0x7a, 0x60, /* movaps [rdx+ 96],xmm7 */ + 0x44, 0x0f, 0x29, 0x42, 0x70, /* movaps [rdx+112],xmm8 */ + 0x48, 0x83, 0xc2, 0x70, /* add rdx,112 */ + 0x44, 0x0f, 0x29, 0x4a, 0x10, /* movaps [rdx+ 16],xmm9 */ + 0x44, 0x0f, 0x29, 0x52, 0x20, /* movaps [rdx+ 32],xmm10 */ + 0x44, 0x0f, 0x29, 0x5a, 0x30, /* movaps [rdx+ 48],xmm11 */ + 0x44, 0x0f, 0x29, 0x62, 0x40, /* movaps [rdx+ 64],xmm12 */ + 0x44, 0x0f, 0x29, 0x6a, 0x50, /* movaps [rdx+ 80],xmm13 */ + 0x44, 0x0f, 0x29, 0x72, 0x60, /* movaps [rdx+ 96],xmm14 */ + 0x44, 0x0f, 0x29, 0x7a, 0x70, /* movaps [rdx+112],xmm15 */ + #endif + 0x48, 0x8b, 0x69, 0x08, /* mov rbp,[rcx+ 8] */ + 0x48, 0x8b, 0x71, 0x10, /* mov rsi,[rcx+16] */ + 0x48, 0x8b, 0x79, 0x18, /* mov rdi,[rcx+24] */ + 0x48, 0x8b, 0x59, 0x20, /* mov rbx,[rcx+32] */ + 0x4c, 0x8b, 0x61, 0x28, /* mov r12,[rcx+40] */ + 0x4c, 0x8b, 0x69, 0x30, /* mov r13,[rcx+48] */ + 0x4c, 0x8b, 0x71, 0x38, /* mov r14,[rcx+56] */ + 0x4c, 0x8b, 0x79, 0x40, /* mov r15,[rcx+64] */ + #if !defined(LIBCO_NO_SSE) + 0x0f, 0x28, 0x71, 0x50, /* movaps xmm6, [rcx+ 80] */ + 0x0f, 0x28, 0x79, 0x60, /* movaps xmm7, [rcx+ 96] */ + 0x44, 0x0f, 0x28, 0x41, 0x70, /* movaps xmm8, [rcx+112] */ + 0x48, 0x83, 0xc1, 0x70, /* add rcx,112 */ + 0x44, 0x0f, 0x28, 0x49, 0x10, /* movaps xmm9, [rcx+ 16] */ + 0x44, 0x0f, 0x28, 0x51, 0x20, /* movaps xmm10,[rcx+ 32] */ + 0x44, 0x0f, 0x28, 0x59, 0x30, /* movaps xmm11,[rcx+ 48] */ + 0x44, 0x0f, 0x28, 0x61, 0x40, /* movaps xmm12,[rcx+ 64] */ + 0x44, 0x0f, 0x28, 0x69, 0x50, /* movaps xmm13,[rcx+ 80] */ + 0x44, 0x0f, 0x28, 0x71, 0x60, /* movaps xmm14,[rcx+ 96] */ + 0x44, 0x0f, 0x28, 0x79, 0x70, /* movaps xmm15,[rcx+112] */ + #endif + 0xff, 0xe0, /* jmp rax */ + }; + + #include + + static void co_init() { + #ifdef LIBCO_MPROTECT + DWORD old_privileges; + VirtualProtect((void*)co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READ, &old_privileges); + #endif + } +#else + /* ABI: SystemV */ + static const unsigned char co_swap_function[4096] = { + 0x48, 0x89, 0x26, /* mov [rsi],rsp */ + 0x48, 0x8b, 0x27, /* mov rsp,[rdi] */ + 0x58, /* pop rax */ + 0x48, 0x89, 0x6e, 0x08, /* mov [rsi+ 8],rbp */ + 0x48, 0x89, 0x5e, 0x10, /* mov [rsi+16],rbx */ + 0x4c, 0x89, 0x66, 0x18, /* mov [rsi+24],r12 */ + 0x4c, 0x89, 0x6e, 0x20, /* mov [rsi+32],r13 */ + 0x4c, 0x89, 0x76, 0x28, /* mov [rsi+40],r14 */ + 0x4c, 0x89, 0x7e, 0x30, /* mov [rsi+48],r15 */ + 0x48, 0x8b, 0x6f, 0x08, /* mov rbp,[rdi+ 8] */ + 0x48, 0x8b, 0x5f, 0x10, /* mov rbx,[rdi+16] */ + 0x4c, 0x8b, 0x67, 0x18, /* mov r12,[rdi+24] */ + 0x4c, 0x8b, 0x6f, 0x20, /* mov r13,[rdi+32] */ + 0x4c, 0x8b, 0x77, 0x28, /* mov r14,[rdi+40] */ + 0x4c, 0x8b, 0x7f, 0x30, /* mov r15,[rdi+48] */ + 0xff, 0xe0, /* jmp rax */ + }; + + #ifdef LIBCO_MPROTECT + #include + #include + #endif + + static void co_init() { + #ifdef LIBCO_MPROTECT + unsigned long long addr = (unsigned long long)co_swap_function; + unsigned long long base = addr - (addr % sysconf(_SC_PAGESIZE)); + unsigned long long size = (addr - base) + sizeof co_swap_function; + mprotect((void*)base, size, PROT_READ | PROT_EXEC); + #endif + } +#endif + +static void crash() { + assert(0); /* called only if cothread_t entrypoint returns */ +} + +cothread_t co_active() { + if(!co_active_handle) co_active_handle = &co_active_buffer; + return co_active_handle; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*entrypoint)(void)) { + cothread_t handle; + if(!co_swap) { + co_init(); + co_swap = (void (*)(cothread_t, cothread_t))co_swap_function; + } + if(!co_active_handle) co_active_handle = &co_active_buffer; + + if(handle = (cothread_t)memory) { + unsigned int offset = (size & ~15) - 32; + long long *p = (long long*)((char*)handle + offset); /* seek to top of stack */ + *--p = (long long)crash; /* crash if entrypoint returns */ + *--p = (long long)entrypoint; /* start of function */ + *(long long*)handle = (long long)p; /* stack pointer */ + } + + return handle; +} + +cothread_t co_create(unsigned int size, void (*entrypoint)(void)) { + void* memory = malloc(size); + if(!memory) return (cothread_t)0; + return co_derive(memory, size, entrypoint); +} + +void co_delete(cothread_t handle) { + free(handle); +} + +void co_switch(cothread_t handle) { + register cothread_t co_previous_handle = co_active_handle; + co_swap(co_active_handle = handle, co_previous_handle); +} + +int co_serializable() { + return 1; +} + +#ifdef __cplusplus +} +#endif diff --git a/arm.c b/arm.c new file mode 100644 index 00000000..1c9dff5c --- /dev/null +++ b/arm.c @@ -0,0 +1,84 @@ +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include +#ifdef LIBCO_MPROTECT + #include + #include +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +static thread_local unsigned long co_active_buffer[64]; +static thread_local cothread_t co_active_handle = 0; +static void (*co_swap)(cothread_t, cothread_t) = 0; + +#ifdef LIBCO_MPROTECT + alignas(4096) +#else + section(text) +#endif +static const unsigned long co_swap_function[1024] = { + 0xe8a16ff0, /* stmia r1!, {r4-r11,sp,lr} */ + 0xe8b0aff0, /* ldmia r0!, {r4-r11,sp,pc} */ + 0xe12fff1e, /* bx lr */ +}; + +static void co_init() { + #ifdef LIBCO_MPROTECT + unsigned long addr = (unsigned long)co_swap_function; + unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE)); + unsigned long size = (addr - base) + sizeof co_swap_function; + mprotect((void*)base, size, PROT_READ | PROT_EXEC); + #endif +} + +cothread_t co_active() { + if(!co_active_handle) co_active_handle = &co_active_buffer; + return co_active_handle; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*entrypoint)(void)) { + unsigned long* handle; + if(!co_swap) { + co_init(); + co_swap = (void (*)(cothread_t, cothread_t))co_swap_function; + } + if(!co_active_handle) co_active_handle = &co_active_buffer; + + if(handle = (unsigned long*)memory) { + unsigned int offset = (size & ~15); + unsigned long* p = (unsigned long*)((unsigned char*)handle + offset); + handle[8] = (unsigned long)p; + handle[9] = (unsigned long)entrypoint; + } + + return handle; +} + +cothread_t co_create(unsigned int size, void (*entrypoint)(void)) { + void* memory = malloc(size); + if(!memory) return (cothread_t)0; + return co_derive(memory, size, entrypoint); +} + +void co_delete(cothread_t handle) { + free(handle); +} + +void co_switch(cothread_t handle) { + cothread_t co_previous_handle = co_active_handle; + co_swap(co_active_handle = handle, co_previous_handle); +} + +int co_serializable() { + return 1; +} + +#ifdef __cplusplus +} +#endif diff --git a/doc/style.css b/doc/style.css new file mode 100755 index 00000000..ab070256 --- /dev/null +++ b/doc/style.css @@ -0,0 +1,12 @@ +body { + background: #333; + color: #fff; +} + +code { + background: #444; +} + +a { + color: #aaf; +} diff --git a/doc/targets.html b/doc/targets.html new file mode 100755 index 00000000..d6211a15 --- /dev/null +++ b/doc/targets.html @@ -0,0 +1,89 @@ + + + + + + + +Supported targets:

+ +Note that supported targets are only those that have been tested and confirmed +working. It is quite possible that libco will work on more processors, compilers +and operating systems than those listed below. +
+ +libco.x86
+Overhead: ~5x
+Supported processor(s): 32-bit x86
+Supported compiler(s): any
+Supported operating system(s):
    +
  • Windows
    • +
  • Mac OS X
    • +
  • Linux
    • +
  • BSD
    • +
+
+ +libco.amd64
+Overhead: ~10x (Windows), ~6x (all other platforms)
+Supported processor(s): 64-bit amd64
+Supported compiler(s): any
+Supported operating system(s):
    +
  • Windows
    • +
  • Mac OS X
    • +
  • Linux
    • +
  • BSD
    • +
+
+ +libco.ppc
+Overhead: ~20x
+Supported processor(s): 32-bit PowerPC, 64-bit PowerPC
+Supported compiler(s): GNU GCC
+Supported operating system(s):
    +
+
  • Mac OS X
    • +
  • Linux
    • +
  • BSD
    • +
  • Playstation 3
    • + +
    + +Note: this module contains compiler flags to enable/disable FPU and Altivec +support. + +
    + +libco.fiber
    +Overhead: ~15x
    +Supported processor(s): Processor independent
    +Supported compiler(s): any
    +Supported operating system(s):
      +
    • Windows
      • +
    +
    + +libco.sjlj
    +Overhead: ~30x
    +Supported processor(s): Processor independent
    +Supported compiler(s): any
    +Supported operating system(s):
      +
    • Mac OS X
      • +
    • Linux
      • +
    • BSD
      • +
    • Solaris
      • +
    +
    + +libco.ucontext
    +Overhead: ~300x
    +Supported processor(s): Processor independent
    +Supported compiler(s): any
    +Supported operating system(s):
      +
    • Linux
      • +
    • BSD
      • +
    +
    + + + diff --git a/doc/usage.html b/doc/usage.html new file mode 100755 index 00000000..994072f9 --- /dev/null +++ b/doc/usage.html @@ -0,0 +1,108 @@ + + + + + + + +License:

    +libco is released under the ISC license. +
    + +Foreword:

    +libco is a cross-platform, permissively licensed implementation of +cooperative-multithreading; a feature that is sorely lacking from the ISO C/C++ +standard.
    +The library is designed for maximum speed and portability, and not for safety or +features. If safety or extra functionality is desired, a wrapper API can easily +be written to encapsulate all library functions.
    +Behavior of executing operations that are listed as not permitted below result +in undefined behavior. They may work anyway, they may cause undesired / unknown +behavior, or they may crash the program entirely.
    +The goal of this library was to simplify the base API as much as possible, +implementing only that which cannot be implemented using pure C. Additional +functionality after this would only complicate ports of this library to new +platforms. +
    + +Porting:

    +This document is included as a reference for porting libco. Please submit any +ports you create to me, so that libco can become more useful. Please note that +since libco is permissively licensed, you must submit your code as a work of the +public domain in order for it to be included in the official distribution. +Full credit will be given in the source code of the official release. Please +do not bother submitting code to me under any other license -- including GPL, +LGPL, BSD or CC -- I am not interested in creating a library with multiple +different licenses depending on which targets are used. +
    + +Synopsis:

    + +typedef void* cothread_t;
    +
    +cothread_t co_active();
    +cothread_t co_create(unsigned int heapsize, void (*coentry)(void));
    +void       co_delete(cothread_t cothread);
    +void       co_switch(cothread_t cothread);
    +     +
    + +Usage: +
    + +typedef void* cothread_t;

    +Handle to cothread.
    +Handle must be of type void*.
    +A value of null (0) indicates an uninitialized or invalid +handle, whereas a non-zero value indicates a valid handle. +
    + +cothread_t co_active();

    +Return handle to current cothread. Always returns a valid handle, even when +called from the main program thread. +
    + +cothread_t co_derive(void* memory, unsigned int heapsize, void (*coentry)(void));

    +Initializes new cothread.
    +This function is identical to co_create, only it attempts to use the provided +memory instead of allocating new memory on the heap. Please note that certain +implementations (currently only Windows Fibers) cannot be created using existing +memory, and as such, this function will fail. +
    + +cothread_t co_create(unsigned int heapsize, void (*coentry)(void));

    +Create new cothread.
    +Heapsize is the amount of memory allocated for the cothread stack, specified +in bytes. This is unfortunately impossible to make fully portable. It is +recommended to specify sizes using `n * sizeof(void*)'. It is better to err +on the side of caution and allocate more memory than will be needed to ensure +compatibility with other platforms, within reason. A typical heapsize for a +32-bit architecture is ~1MB.
    +When the new cothread is first called, program execution jumps to coentry. +This function does not take any arguments, due to portability issues with +passing function arguments. However, arguments can be simulated by the use +of global variables, which can be set before the first call to each cothread.
    +coentry() must not return, and should end with an appropriate co_switch() +statement. Behavior is undefined if entry point returns normally.
    +Library is responsible for allocating cothread stack memory, to free +the user from needing to allocate special memory capable of being used +as program stack memory on platforms where this is required.
    +User is always responsible for deleting cothreads with co_delete().
    +Return value of null (0) indicates cothread creation failed. +
    + +void co_delete(cothread_t cothread);

    +Delete specified cothread.
    +Null (0) or invalid cothread handle is not allowed.
    +Passing handle of active cothread to this function is not allowed.
    +Passing handle of primary cothread is not allowed. +
    + +void co_switch(cothread_t cothread);

    +Switch to specified cothread.
    +Null (0) or invalid cothread handle is not allowed.
    +Passing handle of active cothread to this function is not allowed. +
    + + + diff --git a/fiber.c b/fiber.c new file mode 100755 index 00000000..dd539c37 --- /dev/null +++ b/fiber.c @@ -0,0 +1,55 @@ +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#define WINVER 0x0400 +#define _WIN32_WINNT 0x0400 +#include + +#ifdef __cplusplus +extern "C" { +#endif + +static thread_local cothread_t co_active_ = 0; + +static void __stdcall co_thunk(void* coentry) { + ((void (*)(void))coentry)(); +} + +cothread_t co_active() { + if(!co_active_) { + ConvertThreadToFiber(0); + co_active_ = GetCurrentFiber(); + } + return co_active_; +} + +cothread_t co_derive(void* memory, unsigned int heapsize, void (*coentry)(void)) { + //Windows fibers do not allow users to supply their own memory + return (cothread_t)0; +} + +cothread_t co_create(unsigned int heapsize, void (*coentry)(void)) { + if(!co_active_) { + ConvertThreadToFiber(0); + co_active_ = GetCurrentFiber(); + } + return (cothread_t)CreateFiber(heapsize, co_thunk, (void*)coentry); +} + +void co_delete(cothread_t cothread) { + DeleteFiber(cothread); +} + +void co_switch(cothread_t cothread) { + co_active_ = cothread; + SwitchToFiber(cothread); +} + +int co_serializable() { + return 0; +} + +#ifdef __cplusplus +} +#endif diff --git a/libco.c b/libco.c new file mode 100755 index 00000000..21fe4cab --- /dev/null +++ b/libco.c @@ -0,0 +1,37 @@ +#if defined(__clang__) + #pragma clang diagnostic ignored "-Wparentheses" + + /* placing code in section(text) does not mark it executable with Clang. */ + #undef LIBCO_MPROTECT + #define LIBCO_MPROTECT +#endif + +#if defined(__clang__) || defined(__GNUC__) + #if defined(__i386__) + #include "x86.c" + #elif defined(__amd64__) + #include "amd64.c" + #elif defined(__arm__) + #include "arm.c" + #elif defined(__aarch64__) + #include "aarch64.c" + #elif defined(__powerpc64__) && defined(_CALL_ELF) && _CALL_ELF == 2 + #include "ppc64v2.c" + #elif defined(_ARCH_PPC) && !defined(__LITTLE_ENDIAN__) + #include "ppc.c" + #elif defined(_WIN32) + #include "fiber.c" + #else + #include "sjlj.c" + #endif +#elif defined(_MSC_VER) + #if defined(_M_IX86) + #include "x86.c" + #elif defined(_M_AMD64) + #include "amd64.c" + #else + #include "fiber.c" + #endif +#else + #error "libco: unsupported processor, compiler or operating system" +#endif diff --git a/libco.h b/libco.h new file mode 100755 index 00000000..88d00a72 --- /dev/null +++ b/libco.h @@ -0,0 +1,28 @@ +/* + libco v20 (2019-10-16) + author: byuu + license: ISC +*/ + +#ifndef LIBCO_H +#define LIBCO_H + +#ifdef __cplusplus +extern "C" { +#endif + +typedef void* cothread_t; + +cothread_t co_active(); +cothread_t co_derive(void*, unsigned int, void (*)(void)); +cothread_t co_create(unsigned int, void (*)(void)); +void co_delete(cothread_t); +void co_switch(cothread_t); +int co_serializable(); + +#ifdef __cplusplus +} +#endif + +/* ifndef LIBCO_H */ +#endif diff --git a/ppc.c b/ppc.c new file mode 100755 index 00000000..ee6a9a87 --- /dev/null +++ b/ppc.c @@ -0,0 +1,431 @@ +/* ppc64le (ELFv2) is not currently supported */ + +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include +#include + +#ifdef LIBCO_MPROTECT + #include + #include +#endif + +/* state format (offsets in 32-bit words) + + +0 pointer to swap code + rest of function descriptor for entry function + +8 PC ++10 SP + special registers + GPRs + FPRs + VRs + stack +*/ + +enum { state_size = 1024 }; +enum { above_stack = 2048 }; +enum { stack_align = 256 }; + +static thread_local cothread_t co_active_handle = 0; + +/* determine environment */ + +#define LIBCO_PPC64 (_ARCH_PPC64 || __PPC64__ || __ppc64__ || __powerpc64__) + +/* whether function calls are indirect through a descriptor, or are directly to function */ +#ifndef LIBCO_PPCDESC + #if !_CALL_SYSV && (_CALL_AIX || _CALL_AIXDESC || (LIBCO_PPC64 && (!defined(_CALL_ELF) || _CALL_ELF == 1))) + #define LIBCO_PPCDESC 1 + #endif +#endif + +#ifdef LIBCO_MPROTECT + alignas(4096) +#else + section(text) +#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 + 0xd9c400e0, /* stfd f14,224(r4) */ + 0xd9e400e8, /* stfd f15,232(r4) */ + 0xda0400f0, /* stfd f16,240(r4) */ + 0xda2400f8, /* stfd f17,248(r4) */ + 0xda440100, /* stfd f18,256(r4) */ + 0xda640108, /* stfd f19,264(r4) */ + 0xda840110, /* stfd f20,272(r4) */ + 0xdaa40118, /* stfd f21,280(r4) */ + 0xdac40120, /* stfd f22,288(r4) */ + 0xdae40128, /* stfd f23,296(r4) */ + 0xdb040130, /* stfd f24,304(r4) */ + 0xdb240138, /* stfd f25,312(r4) */ + 0xdb440140, /* stfd f26,320(r4) */ + 0xdb640148, /* stfd f27,328(r4) */ + 0xdb840150, /* stfd f28,336(r4) */ + 0xdba40158, /* stfd f29,344(r4) */ + 0xdbc40160, /* stfd f30,352(r4) */ + 0xdbe40168, /* stfd f31,360(r4) */ + 0xc9c300e0, /* lfd f14,224(r3) */ + 0xc9e300e8, /* lfd f15,232(r3) */ + 0xca0300f0, /* lfd f16,240(r3) */ + 0xca2300f8, /* lfd f17,248(r3) */ + 0xca430100, /* lfd f18,256(r3) */ + 0xca630108, /* lfd f19,264(r3) */ + 0xca830110, /* lfd f20,272(r3) */ + 0xcaa30118, /* lfd f21,280(r3) */ + 0xcac30120, /* lfd f22,288(r3) */ + 0xcae30128, /* lfd f23,296(r3) */ + 0xcb030130, /* lfd f24,304(r3) */ + 0xcb230138, /* lfd f25,312(r3) */ + 0xcb430140, /* lfd f26,320(r3) */ + 0xcb630148, /* lfd f27,328(r3) */ + 0xcb830150, /* lfd f28,336(r3) */ + 0xcba30158, /* lfd f29,344(r3) */ + 0xcbc30160, /* lfd f30,352(r3) */ + 0xcbe30168, /* lfd f31,360(r3) */ + #endif + + #ifdef __ALTIVEC__ + 0x7ca042a6, /* mfvrsave r5 */ + 0x39040180, /* addi r8,r4,384 */ + 0x39240190, /* addi r9,r4,400 */ + 0x70a00fff, /* andi. r0,r5,4095 */ + 0x90a40034, /* stw r5,52(r4) */ + 0x4182005c, /* beq- 2 */ + 0x7e8041ce, /* stvx v20,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7ea049ce, /* stvx v21,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7ec041ce, /* stvx v22,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7ee049ce, /* stvx v23,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f0041ce, /* stvx v24,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7f2049ce, /* stvx v25,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f4041ce, /* stvx v26,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7f6049ce, /* stvx v27,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f8041ce, /* stvx v28,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7fa049ce, /* stvx v29,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7fc041ce, /* stvx v30,r0,r8 */ + 0x7fe049ce, /* stvx v31,r0,r9 */ + 0x80a30034, /*2:lwz r5,52(r3) */ + 0x39030180, /* addi r8,r3,384 */ + 0x39230190, /* addi r9,r3,400 */ + 0x70a00fff, /* andi. r0,r5,4095 */ + 0x7ca043a6, /* mtvrsave r5 */ + 0x4d820420, /* beqctr */ + 0x7e8040ce, /* lvx v20,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7ea048ce, /* lvx v21,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7ec040ce, /* lvx v22,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7ee048ce, /* lvx v23,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f0040ce, /* lvx v24,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7f2048ce, /* lvx v25,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f4040ce, /* lvx v26,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7f6048ce, /* lvx v27,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7f8040ce, /* lvx v28,r0,r8 */ + 0x39080020, /* addi r8,r8,32 */ + 0x7fa048ce, /* lvx v29,r0,r9 */ + 0x39290020, /* addi r9,r9,32 */ + 0x7fc040ce, /* lvx v30,r0,r8 */ + 0x7fe048ce, /* lvx v31,r0,r9 */ + #endif + + 0x4e800420, /* bctr */ +}; + +#if LIBCO_PPCDESC + /* function call goes through indirect descriptor */ + #define CO_SWAP_ASM(x, y) ((void (*)(cothread_t, cothread_t))(uintptr_t)x)(x, y) +#else + /* function call goes directly to code */ + #define CO_SWAP_ASM(x, y) ((void (*)(cothread_t, cothread_t))(uintptr_t)libco_ppc_code)(x, y) +#endif + +static uint32_t* co_derive_(void* memory, unsigned size, uintptr_t entry) { + (void)entry; + + uint32_t* t = (uint32_t*)memory; + + #if LIBCO_PPCDESC + 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 */ + } + #endif + + return t; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*entry_)(void)) { + uintptr_t entry = (uintptr_t)entry_; + uint32_t* t = 0; + + /* be sure main thread was successfully allocated */ + if(co_active()) { + t = co_derive_(memory, size, entry); + } + + if(t) { + uintptr_t sp; + int shift; + + /* save current registers into new thread, so that any special ones will have proper values when thread is begun */ + CO_SWAP_ASM(t, t); + + #if LIBCO_PPCDESC + entry = (uintptr_t)*(void**)entry; /* get real address */ + #endif + + /* put stack near end of block, and align */ + sp = (uintptr_t)t + size - above_stack; + sp -= sp % stack_align; + + /* 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 + uses. so, we manually write pointers at the proper size. we always + save and restore at the same address, and since PPC is big-endian, + we must put the low byte first on PPC32. */ + + /* if uintptr_t is 32 bits, >>32 is undefined behavior, + so we do two shifts and don't have to care how many bits uintptr_t is. */ + #if LIBCO_PPC64 + shift = 16; + #else + shift = 0; + #endif + + /* set up so entry will be called on next swap */ + t[ 8] = (uint32_t)(entry >> shift >> shift); + t[ 9] = (uint32_t)entry; + + t[10] = (uint32_t)(sp >> shift >> shift); + t[11] = (uint32_t)sp; + } + + return t; +} + +static uint32_t* co_create_(unsigned size, uintptr_t entry) { + (void)entry; + + uint32_t* t = (uint32_t*)malloc(size); + + #if LIBCO_PPCDESC + 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 */ + } + #endif + + return t; +} + +cothread_t co_create(unsigned int size, void (*entry_)(void)) { + uintptr_t entry = (uintptr_t)entry_; + uint32_t* t = 0; + + /* be sure main thread was successfully allocated */ + if(co_active()) { + size += state_size + above_stack + stack_align; + t = co_create_(size, entry); + } + + if(t) { + uintptr_t sp; + int shift; + + /* save current registers into new thread, so that any special ones will have proper values when thread is begun */ + CO_SWAP_ASM(t, t); + + #if LIBCO_PPCDESC + entry = (uintptr_t)*(void**)entry; /* get real address */ + #endif + + /* put stack near end of block, and align */ + sp = (uintptr_t)t + size - above_stack; + sp -= sp % stack_align; + + /* 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 + uses. so, we manually write pointers at the proper size. we always + save and restore at the same address, and since PPC is big-endian, + we must put the low byte first on PPC32. */ + + /* if uintptr_t is 32 bits, >>32 is undefined behavior, + so we do two shifts and don't have to care how many bits uintptr_t is. */ + #if LIBCO_PPC64 + shift = 16; + #else + shift = 0; + #endif + + /* set up so entry will be called on next swap */ + t[ 8] = (uint32_t)(entry >> shift >> shift); + t[ 9] = (uint32_t)entry; + + t[10] = (uint32_t)(sp >> shift >> shift); + t[11] = (uint32_t)sp; + } + + return t; +} + +void co_delete(cothread_t t) { + free(t); +} + +static void co_init_(void) { + #if LIBCO_MPROTECT + long page_size = sysconf(_SC_PAGESIZE); + if(page_size > 0) { + uintptr_t align = page_size; + uintptr_t begin = (uintptr_t)libco_ppc_code; + uintptr_t end = begin + sizeof libco_ppc_code; + + /* align beginning and end */ + end += align - 1; + end -= end % align; + begin -= begin % align; + + mprotect((void*)begin, end - begin, PROT_READ | PROT_EXEC); + } + #endif + + co_active_handle = co_create_(state_size, (uintptr_t)&co_switch); +} + +cothread_t co_active() { + if(!co_active_handle) co_init_(); + + return co_active_handle; +} + +void co_switch(cothread_t t) { + cothread_t old = co_active_handle; + co_active_handle = t; + + CO_SWAP_ASM(t, old); +} + +int co_serializable() { + return 0; +} diff --git a/ppc64v2.c b/ppc64v2.c new file mode 100644 index 00000000..3c296959 --- /dev/null +++ b/ppc64v2.c @@ -0,0 +1,279 @@ +/* author: Shawn Anastasio */ + +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct ppc64_context { + //GPRs + uint64_t gprs[32]; + uint64_t lr; + uint64_t ccr; + + //FPRs + uint64_t fprs[32]; + + #ifdef __ALTIVEC__ + //Altivec (VMX) + uint64_t vmx[12 * 2]; + uint32_t vrsave; + #endif +}; + +static thread_local struct ppc64_context* co_active_handle = 0; + +#define MAX(x, y) ((x) > (y) ? (x) : (y)) +#define ALIGN(p, x) ((void*)((uintptr_t)(p) & ~((x) - 1))) + +#define MIN_STACK 0x10000lu +#define MIN_STACK_FRAME 0x20lu +#define STACK_ALIGN 0x10lu + +void swap_context(struct ppc64_context* read, struct ppc64_context* write); +__asm__( + ".text\n" + ".align 4\n" + ".type swap_context @function\n" + "swap_context:\n" + ".cfi_startproc\n" + + //save GPRs + "std 1, 8(4)\n" + "std 2, 16(4)\n" + "std 12, 96(4)\n" + "std 13, 104(4)\n" + "std 14, 112(4)\n" + "std 15, 120(4)\n" + "std 16, 128(4)\n" + "std 17, 136(4)\n" + "std 18, 144(4)\n" + "std 19, 152(4)\n" + "std 20, 160(4)\n" + "std 21, 168(4)\n" + "std 22, 176(4)\n" + "std 23, 184(4)\n" + "std 24, 192(4)\n" + "std 25, 200(4)\n" + "std 26, 208(4)\n" + "std 27, 216(4)\n" + "std 28, 224(4)\n" + "std 29, 232(4)\n" + "std 30, 240(4)\n" + "std 31, 248(4)\n" + + //save LR + "mflr 5\n" + "std 5, 256(4)\n" + + //save CCR + "mfcr 5\n" + "std 5, 264(4)\n" + + //save FPRs + "stfd 14, 384(4)\n" + "stfd 15, 392(4)\n" + "stfd 16, 400(4)\n" + "stfd 17, 408(4)\n" + "stfd 18, 416(4)\n" + "stfd 19, 424(4)\n" + "stfd 20, 432(4)\n" + "stfd 21, 440(4)\n" + "stfd 22, 448(4)\n" + "stfd 23, 456(4)\n" + "stfd 24, 464(4)\n" + "stfd 25, 472(4)\n" + "stfd 26, 480(4)\n" + "stfd 27, 488(4)\n" + "stfd 28, 496(4)\n" + "stfd 29, 504(4)\n" + "stfd 30, 512(4)\n" + "stfd 31, 520(4)\n" + + #ifdef __ALTIVEC__ + //save VMX + "li 5, 528\n" + "stvxl 20, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 21, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 22, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 23, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 24, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 25, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 26, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 27, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 28, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 29, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 30, 4, 5\n" + "addi 5, 5, 16\n" + "stvxl 31, 4, 5\n" + "addi 5, 5, 16\n" + + //save VRSAVE + "mfvrsave 5\n" + "stw 5, 736(4)\n" + #endif + + //restore GPRs + "ld 1, 8(3)\n" + "ld 2, 16(3)\n" + "ld 12, 96(3)\n" + "ld 13, 104(3)\n" + "ld 14, 112(3)\n" + "ld 15, 120(3)\n" + "ld 16, 128(3)\n" + "ld 17, 136(3)\n" + "ld 18, 144(3)\n" + "ld 19, 152(3)\n" + "ld 20, 160(3)\n" + "ld 21, 168(3)\n" + "ld 22, 176(3)\n" + "ld 23, 184(3)\n" + "ld 24, 192(3)\n" + "ld 25, 200(3)\n" + "ld 26, 208(3)\n" + "ld 27, 216(3)\n" + "ld 28, 224(3)\n" + "ld 29, 232(3)\n" + "ld 30, 240(3)\n" + "ld 31, 248(3)\n" + + //restore LR + "ld 5, 256(3)\n" + "mtlr 5\n" + + //restore CCR + "ld 5, 264(3)\n" + "mtcr 5\n" + + //restore FPRs + "lfd 14, 384(3)\n" + "lfd 15, 392(3)\n" + "lfd 16, 400(3)\n" + "lfd 17, 408(3)\n" + "lfd 18, 416(3)\n" + "lfd 19, 424(3)\n" + "lfd 20, 432(3)\n" + "lfd 21, 440(3)\n" + "lfd 22, 448(3)\n" + "lfd 23, 456(3)\n" + "lfd 24, 464(3)\n" + "lfd 25, 472(3)\n" + "lfd 26, 480(3)\n" + "lfd 27, 488(3)\n" + "lfd 28, 496(3)\n" + "lfd 29, 504(3)\n" + "lfd 30, 512(3)\n" + "lfd 31, 520(3)\n" + + #ifdef __ALTIVEC__ + //restore VMX + "li 5, 528\n" + "lvxl 20, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 21, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 22, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 23, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 24, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 25, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 26, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 27, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 28, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 29, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 30, 3, 5\n" + "addi 5, 5, 16\n" + "lvxl 31, 3, 5\n" + "addi 5, 5, 16\n" + + //restore VRSAVE + "lwz 5, 720(3)\n" + "mtvrsave 5\n" + #endif + + //branch to LR + "blr\n" + + ".cfi_endproc\n" + ".size swap_context, .-swap_context\n" +); + +cothread_t co_active() { + if(!co_active_handle) { + co_active_handle = (struct ppc64_context*)malloc(MIN_STACK + sizeof(struct ppc64_context)); + } + return (cothread_t)co_active_handle; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*coentry)(void)) { + uint8_t* sp; + struct ppc64_context* context = (struct ppc64_context*)memory; + + //save current context into new context to initialize it + swap_context(context, context); + + //align stack + sp = (uint8_t*)memory + size - STACK_ALIGN; + sp = (uint8_t*)ALIGN(sp, STACK_ALIGN); + + //write 0 for initial backchain + *(uint64_t*)sp = 0; + + //create new frame with backchain + sp -= MIN_STACK_FRAME; + *(uint64_t*)sp = (uint64_t)(sp + MIN_STACK_FRAME); + + //update context with new stack (r1) and entrypoint (r12, lr) + context->gprs[ 1] = (uint64_t)sp; + context->gprs[12] = (uint64_t)coentry; + context->lr = (uint64_t)coentry; + + return (cothread_t)memory; +} + +cothread_t co_create(unsigned int size, void (*coentry)(void)) { + void* memory = malloc(size); + if(!memory) return (cothread_t)0; + return co_derive(memory, size, coentry); +} + +void co_delete(cothread_t handle) { + free(handle); +} + +void co_switch(cothread_t to) { + struct ppc64_context* from = co_active_handle; + co_active_handle = (struct ppc64_context*)to; + swap_context((struct ppc64_context*)to, from); +} + +int co_serializable() { + return 1; +} + +#ifdef __cplusplus +} +#endif diff --git a/settings.h b/settings.h new file mode 100644 index 00000000..d8037bc4 --- /dev/null +++ b/settings.h @@ -0,0 +1,38 @@ +#if defined(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 */ + +#if defined(LIBCO_C) + #if defined(LIBCO_MP) + #define thread_local __thread + #else + #define thread_local + #endif +#endif + +#if __STDC_VERSION__ >= 201112L + #if !defined(_MSC_VER) + #include + #endif +#else + #define alignas(bytes) +#endif + +#if defined(_MSC_VER) + #define section(name) __declspec(allocate("." #name)) +#elif defined(__APPLE__) + #define section(name) __attribute__((section("__TEXT,__" #name))) +#else + #define section(name) __attribute__((section("." #name "#"))) +#endif + +/* if defined(LIBCO_C) */ +#endif diff --git a/sjlj.c b/sjlj.c new file mode 100755 index 00000000..5af14729 --- /dev/null +++ b/sjlj.c @@ -0,0 +1,145 @@ +/* + note this was designed for UNIX systems. Based on ideas expressed in a paper by Ralf Engelschall. + for SJLJ on other systems, one would want to rewrite springboard() and co_create() and hack the jmb_buf stack pointer. +*/ + +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#define _BSD_SOURCE +#define _XOPEN_SOURCE 500 +#include +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct { + sigjmp_buf context; + void (*coentry)(void); + void* stack; +} cothread_struct; + +static thread_local cothread_struct co_primary; +static thread_local cothread_struct* creating; +static thread_local cothread_struct* co_running = 0; + +static void springboard(int ignored) { + if(sigsetjmp(creating->context, 0)) { + co_running->coentry(); + } +} + +cothread_t co_active() { + if(!co_running) co_running = &co_primary; + return (cothread_t)co_running; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*coentry)(void)) { + if(!co_running) co_running = &co_primary; + + cothread_struct* thread = (cothread_struct*)memory; + memory = (unsigned char*)memory + sizeof(cothread_struct); + size -= sizeof(cothread_struct); + if(thread) { + struct sigaction handler; + struct sigaction old_handler; + + stack_t stack; + stack_t old_stack; + + thread->coentry = thread->stack = 0; + + stack.ss_flags = 0; + stack.ss_size = size; + thread->stack = stack.ss_sp = memory; + if(stack.ss_sp && !sigaltstack(&stack, &old_stack)) { + handler.sa_handler = springboard; + handler.sa_flags = SA_ONSTACK; + sigemptyset(&handler.sa_mask); + creating = thread; + + if(!sigaction(SIGUSR1, &handler, &old_handler)) { + if(!raise(SIGUSR1)) { + thread->coentry = coentry; + } + sigaltstack(&old_stack, 0); + sigaction(SIGUSR1, &old_handler, 0); + } + } + + if(thread->coentry != coentry) { + co_delete(thread); + thread = 0; + } + } + + return (cothread_t)thread; +} + +cothread_t co_create(unsigned int size, void (*coentry)(void)) { + if(!co_running) co_running = &co_primary; + + cothread_struct* thread = (cothread_struct*)malloc(sizeof(cothread_struct)); + if(thread) { + struct sigaction handler; + struct sigaction old_handler; + + stack_t stack; + stack_t old_stack; + + thread->coentry = thread->stack = 0; + + stack.ss_flags = 0; + stack.ss_size = size; + thread->stack = stack.ss_sp = malloc(size); + if(stack.ss_sp && !sigaltstack(&stack, &old_stack)) { + handler.sa_handler = springboard; + handler.sa_flags = SA_ONSTACK; + sigemptyset(&handler.sa_mask); + creating = thread; + + if(!sigaction(SIGUSR1, &handler, &old_handler)) { + if(!raise(SIGUSR1)) { + thread->coentry = coentry; + } + sigaltstack(&old_stack, 0); + sigaction(SIGUSR1, &old_handler, 0); + } + } + + if(thread->coentry != coentry) { + co_delete(thread); + thread = 0; + } + } + + return (cothread_t)thread; +} + +void co_delete(cothread_t cothread) { + if(cothread) { + if(((cothread_struct*)cothread)->stack) { + free(((cothread_struct*)cothread)->stack); + } + free(cothread); + } +} + +void co_switch(cothread_t cothread) { + if(!sigsetjmp(co_running->context, 0)) { + co_running = (cothread_struct*)cothread; + siglongjmp(co_running->context, 1); + } +} + +int co_serializable() { + return 0; +} + +#ifdef __cplusplus +} +#endif diff --git a/ucontext.c b/ucontext.c new file mode 100755 index 00000000..5ff76af3 --- /dev/null +++ b/ucontext.c @@ -0,0 +1,86 @@ +/* + WARNING: the overhead of POSIX ucontext is very high, + assembly versions of libco or libco_sjlj should be much faster + + this library only exists for two reasons: + 1: as an initial test for the viability of a ucontext implementation + 2: to demonstrate the power and speed of libco over existing implementations, + such as pth (which defaults to wrapping ucontext on unix targets) + + use this library only as a *last resort* +*/ + +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#define _BSD_SOURCE +#define _XOPEN_SOURCE 500 +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +static thread_local ucontext_t co_primary; +static thread_local ucontext_t* co_running = 0; + +cothread_t co_active() { + if(!co_running) co_running = &co_primary; + return (cothread_t)co_running; +} + +cothread_t co_derive(void* memory, unsigned int heapsize, void (*coentry)(void)) { + if(!co_running) co_running = &co_primary; + ucontext_t* thread = (ucontext_t*)memory; + memory = (unsigned char*)memory + sizeof(ucontext_t); + heapsize -= sizeof(ucontext_t); + if(thread) { + if((!getcontext(thread) && !(thread->uc_stack.ss_sp = 0)) && (thread->uc_stack.ss_sp = memory)) { + thread->uc_link = co_running; + thread->uc_stack.ss_size = heapsize; + makecontext(thread, coentry, 0); + } else { + thread = 0; + } + } + return (cothread_t)thread; +} + +cothread_t co_create(unsigned int heapsize, void (*coentry)(void)) { + if(!co_running) co_running = &co_primary; + ucontext_t* thread = (ucontext_t*)malloc(sizeof(ucontext_t)); + if(thread) { + if((!getcontext(thread) && !(thread->uc_stack.ss_sp = 0)) && (thread->uc_stack.ss_sp = malloc(heapsize))) { + thread->uc_link = co_running; + thread->uc_stack.ss_size = heapsize; + makecontext(thread, coentry, 0); + } else { + co_delete((cothread_t)thread); + thread = 0; + } + } + return (cothread_t)thread; +} + +void co_delete(cothread_t cothread) { + if(cothread) { + if(((ucontext_t*)cothread)->uc_stack.ss_sp) { free(((ucontext_t*)cothread)->uc_stack.ss_sp); } + free(cothread); + } +} + +void co_switch(cothread_t cothread) { + ucontext_t* old_thread = co_running; + co_running = (ucontext_t*)cothread; + swapcontext(old_thread, co_running); +} + +int co_serializable() { + return 0; +} + +#ifdef __cplusplus +} +#endif diff --git a/x86.c b/x86.c new file mode 100755 index 00000000..c539a299 --- /dev/null +++ b/x86.c @@ -0,0 +1,119 @@ +#define LIBCO_C +#include "libco.h" +#include "settings.h" + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +#if defined(__clang__) || defined(__GNUC__) + #define fastcall __attribute__((fastcall)) +#elif defined(_MSC_VER) + #define fastcall __fastcall +#else + #error "libco: please define fastcall macro" +#endif + +static thread_local long co_active_buffer[64]; +static thread_local cothread_t co_active_handle = 0; +static void (fastcall *co_swap)(cothread_t, cothread_t) = 0; + +#ifdef LIBCO_MPROTECT + alignas(4096) +#else + section(text) +#endif +/* ABI: fastcall */ +static const unsigned char co_swap_function[4096] = { + 0x89, 0x22, /* mov [edx],esp */ + 0x8b, 0x21, /* mov esp,[ecx] */ + 0x58, /* pop eax */ + 0x89, 0x6a, 0x04, /* mov [edx+ 4],ebp */ + 0x89, 0x72, 0x08, /* mov [edx+ 8],esi */ + 0x89, 0x7a, 0x0c, /* mov [edx+12],edi */ + 0x89, 0x5a, 0x10, /* mov [edx+16],ebx */ + 0x8b, 0x69, 0x04, /* mov ebp,[ecx+ 4] */ + 0x8b, 0x71, 0x08, /* mov esi,[ecx+ 8] */ + 0x8b, 0x79, 0x0c, /* mov edi,[ecx+12] */ + 0x8b, 0x59, 0x10, /* mov ebx,[ecx+16] */ + 0xff, 0xe0, /* jmp eax */ +}; + +#ifdef _WIN32 + #include + + static void co_init() { + #ifdef LIBCO_MPROTECT + DWORD old_privileges; + VirtualProtect((void*)co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READ, &old_privileges); + #endif + } +#else + #ifdef LIBCO_MPROTECT + #include + #include + #endif + + static void co_init() { + #ifdef LIBCO_MPROTECT + unsigned long addr = (unsigned long)co_swap_function; + unsigned long base = addr - (addr % sysconf(_SC_PAGESIZE)); + unsigned long size = (addr - base) + sizeof co_swap_function; + mprotect((void*)base, size, PROT_READ | PROT_EXEC); + #endif + } +#endif + +static void crash() { + assert(0); /* called only if cothread_t entrypoint returns */ +} + +cothread_t co_active() { + if(!co_active_handle) co_active_handle = &co_active_buffer; + return co_active_handle; +} + +cothread_t co_derive(void* memory, unsigned int size, void (*entrypoint)(void)) { + cothread_t handle; + if(!co_swap) { + co_init(); + co_swap = (void (fastcall*)(cothread_t, cothread_t))co_swap_function; + } + if(!co_active_handle) co_active_handle = &co_active_buffer; + + if(handle = (cothread_t)memory) { + unsigned int offset = (size & ~15) - 32; + long *p = (long*)((char*)handle + offset); /* seek to top of stack */ + *--p = (long)crash; /* crash if entrypoint returns */ + *--p = (long)entrypoint; /* start of function */ + *(long*)handle = (long)p; /* stack pointer */ + } + + return handle; +} + +cothread_t co_create(unsigned int size, void (*entrypoint)(void)) { + void* memory = malloc(size); + if(!memory) return (cothread_t)0; + return co_derive(memory, size, entrypoint); +} + +void co_delete(cothread_t handle) { + free(handle); +} + +void co_switch(cothread_t handle) { + register cothread_t co_previous_handle = co_active_handle; + co_swap(co_active_handle = handle, co_previous_handle); +} + +int co_serializable() { + return 1; +} + +#ifdef __cplusplus +} +#endif