jyapayne/Nim
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Coroutine rework.

Browse source 
* ucontext backend (default on unix)
* setjmp backend
* fibers backend (default and required on windows)
* Fixed coroutine loop timing issues
* Fixed saving of xmm registers on x64 windows
* Fixed alignment issues
* Updated coroutine sample with cooperative fibonacci calculation.
* Disable glibc security features only when platform jump functions are used
* Removed dependency on fasm.
    * Using fiber api on windows.
    * Other platforms and compilers  will use built in assembler and .S files or API provided by platform libc.
* Replaced stack switching procs with `coroExecWithStack()` which never returns. This makes compiler always generate proper code.
This commit is contained in:
Rokas Kupstys 2017-02-20 17:10:50 +02:00
commit 373e667dbc
11 changed files with 421 additions and 446 deletions
Download patch file Download diff file Expand all files Collapse all files

96
lib/arch/x86/amd64.S Normal file
View file

@ -0,0 +1,96 @@
#
#
# Nim's Runtime Library
# (c) Copyright 2015 Rokas Kupstys
#
# See the file "copying.txt", included in this
# distribution, for details about the copyright.
#
# Partially based on code from musl libc Copyright © 2005-2014 Rich Felker, et al.
.globl narch_coroExecWithStack
.globl narch_setjmp
.globl narch_longjmp
.text
# SysV ABI - first argument is rdi.
# MS ABI - first argument is rcx.
#if defined(__MINGW32__) || defined(__MINGW64__)
#define REG_ARG1 rcx
#define REG_ARG2 rdx
#else
#define REG_ARG1 rdi
#define REG_ARG2 rsi
#endif
narch_coroExecWithStack:
mov %REG_ARG2, %rsp # swap stack with one passed to func
sub $0x30, %rsp # shadow space (for ms ABI) 0x20 + 0x10 for possible misalignment
and $-0x10, %rsp # 16-byte stack alignment
call *%REG_ARG1
narch_setjmp:
add $0x10, %REG_ARG1 # 16-byte alignment
and $-0x10, %REG_ARG1
mov %rbx, 0x00(%REG_ARG1) # jmp_buf, move registers onto it
mov %rbp, 0x08(%REG_ARG1)
mov %r12, 0x10(%REG_ARG1)
mov %r13, 0x18(%REG_ARG1)
mov %r14, 0x20(%REG_ARG1)
mov %r15, 0x28(%REG_ARG1)
lea 0x08(%rsp), %rdx # this is our rsp WITHOUT current ret addr
mov %rdx, 0x30(%REG_ARG1)
mov (%rsp), %rdx # save return addr ptr for new rip
mov %rdx, 0x38(%REG_ARG1)
mov %rsi, 0x40(%REG_ARG1)
mov %rdi, 0x48(%REG_ARG1)
#if defined(__MINGW32__) || defined(__MINGW64__)
movaps %xmm6, 0x50(%REG_ARG1)
movaps %xmm7, 0x60(%REG_ARG1)
movaps %xmm8, 0x70(%REG_ARG1)
movaps %xmm9, 0x80(%REG_ARG1)
movaps %xmm10, 0x90(%REG_ARG1)
movaps %xmm11, 0xA0(%REG_ARG1)
movaps %xmm12, 0xB0(%REG_ARG1)
movaps %xmm13, 0xC0(%REG_ARG1)
movaps %xmm14, 0xD0(%REG_ARG1)
movaps %xmm15, 0xE0(%REG_ARG1)
#endif
xor %rax, %rax # always return 0
ret
narch_longjmp:
add $0x10, %REG_ARG1 # 16-byte alignment
and $-0x10, %REG_ARG1 #
mov %REG_ARG2, %rax # val will be longjmp return
test %rax, %rax
jnz narch_longjmp_1
inc %rax # if val==0, val=1 per longjmp semantics
narch_longjmp_1:
mov 0x00(%REG_ARG1), %rbx # jmp_buf, restore regs from it
mov 0x08(%REG_ARG1), %rbp
mov 0x10(%REG_ARG1), %r12
mov 0x18(%REG_ARG1), %r13
mov 0x20(%REG_ARG1), %r14
mov 0x28(%REG_ARG1), %r15
mov 0x30(%REG_ARG1), %rsp # this ends up being the stack pointer
mov 0x38(%REG_ARG1), %rdx # this is the instruction pointer
mov 0x40(%REG_ARG1), %rsi
mov 0x48(%REG_ARG1), %rdi
#if defined(__MINGW32__) || defined(__MINGW64__)
movaps 0x50(%REG_ARG1), %xmm6
movaps 0x60(%REG_ARG1), %xmm7
movaps 0x70(%REG_ARG1), %xmm8
movaps 0x80(%REG_ARG1), %xmm9
movaps 0x90(%REG_ARG1), %xmm10
movaps 0xA0(%REG_ARG1), %xmm11
movaps 0xB0(%REG_ARG1), %xmm12
movaps 0xC0(%REG_ARG1), %xmm13
movaps 0xD0(%REG_ARG1), %xmm14
movaps 0xE0(%REG_ARG1), %xmm15
#endif
jmp *%rdx # goto saved address without altering rsp

64
lib/arch/x86/i386.S Normal file
View file

@ -0,0 +1,64 @@
#
#
# Nim's Runtime Library
# (c) Copyright 2015 Rokas Kupstys
#
# See the file "copying.txt", included in this
# distribution, for details about the copyright.
#
# Partially based on code from musl libc Copyright © 2005-2014 Rich Felker, et al.
.globl narch_coroExecWithStack
.globl narch_setjmp
.globl narch_longjmp
#if defined(__MINGW32__) || defined(__MINGW64__)
.globl @narch_coroExecWithStack@8
.globl @narch_setjmp@4
.globl @narch_longjmp@8
#endif
.text
#if defined(__MINGW32__) || defined(__MINGW64__)
@narch_coroExecWithStack@8:
#endif
narch_coroExecWithStack:
mov %edx, %esp # swap stack with one passed to func
sub $0x10, %esp # 16-byte alignment
and $-0x10, %esp #
sub $4, %esp # Simulate misalignment caused by return addr
jmp *%ecx
#if defined(__MINGW32__) || defined(__MINGW64__)
@narch_setjmp@4:
#endif
narch_setjmp:
mov %ebx, (%ecx)
mov %esi, 0x04(%ecx)
mov %edi, 0x08(%ecx)
mov %ebp, 0x0C(%ecx)
lea 0x04(%esp), %eax
mov %eax, 0x10(%ecx)
mov (%esp), %eax
mov %eax, 0x14(%ecx)
xor %eax, %eax
ret
#if defined(__MINGW32__) || defined(__MINGW64__)
@narch_longjmp@8:
#endif
narch_longjmp:
mov %edx, %eax
test %eax, %eax
jnz narch_longjmp_1
inc %eax
narch_longjmp_1:
mov (%ecx), %ebx
mov 0x04(%ecx), %esi
mov 0x08(%ecx), %edi
mov 0x0C(%ecx), %ebp
mov 0x10(%ecx), %esp
mov 0x14(%ecx), %edx
jmp *%edx