9 files changed, 449 insertions, 0 deletions
diff --git a/src/backend/initializestack_32.cpp b/src/backend/initializestack_32.cpp
new file mode 100644
index 0000000..d387460
--- /dev/null
+++ b/src/backend/initializestack_32.cpp
@@ -0,0 +1,21 @@
+#include <stdlib.h>
+
+void initializeStack(void *data, int size, void (*entry)(), void **stackPointer)
+{
+    void* stackBottom = (char*)data + size;
+    // align to 16 byte
+    stackBottom = (void*)((size_t)stackBottom & ~0xF);
+    
+    void **p = (void**)stackBottom;
+
+    *(--p) = 0; // align
+    *(--p) = (void*)entry; // rip
+    *(--p) = stackBottom; // ebp
+    *(--p) = 0; // ebx
+    *(--p) = 0; // esi
+    *(--p) = 0; // edi
+    *(--p) = (void*)0x00001f80; // SIMD floating point control default
+    *(--p) = (void*)0x0000033f; // floating point control default
+    
+    *stackPointer = p;
+}
diff --git a/src/backend/initializestack_64_linux_mac.cpp b/src/backend/initializestack_64_linux_mac.cpp
new file mode 100644
index 0000000..5465dfb
--- /dev/null
+++ b/src/backend/initializestack_64_linux_mac.cpp
@@ -0,0 +1,22 @@
+#include <stdlib.h>
+
+void initializeStack(void *data, int size, void (*entry)(), void **stackPointer)
+{
+    void* stackBottom = (char*)data + size;
+    // align to 16 byte
+    stackBottom = (void*)((size_t)stackBottom & ~0xF);
+    
+    void **p = (void**)stackBottom;
+
+    *(--p) = 0; // align
+    *(--p) = (void*)entry; // rip
+    *(--p) = stackBottom; // rbp
+    *(--p) = 0; // rbx
+    *(--p) = 0; // r12
+    *(--p) = 0; // r13
+    *(--p) = 0; // r14
+    *(--p) = 0; // r15
+    *(--p) = (void*)0x00001f800000033f; // SIMD and regular floating point control defaults
+    
+    *stackPointer = p;
+}
diff --git a/src/backend/initializestack_64_win.cpp b/src/backend/initializestack_64_win.cpp
new file mode 100644
index 0000000..3b29caf
--- /dev/null
+++ b/src/backend/initializestack_64_win.cpp
@@ -0,0 +1,44 @@
+#include <stdlib.h>
+
+void initializeStack(void *data, int size, void (*entry)(), void **stackPointer)
+{
+    void* stackBottom = (char*)data + size;
+    // align to 16 byte
+    stackBottom = (void*)((size_t)stackBottom & ~0xF);
+    
+    void **p = (void**)stackBottom;
+
+    //*(--p) = 0; // align to 16 bytes
+    *(--p) = (void*)fn; // rip
+    *(--p) = entry; // rbp
+    *(--p) = 0; // rbx
+    *(--p) = 0; // r12
+    *(--p) = 0; // r13
+    *(--p) = 0; // r14
+    *(--p) = 0; // r15
+    *(--p) = 0; // rsi
+    *(--p) = 0; // rdi
+    *(--p) = 0; // xmm6
+    *(--p) = 0; // xmm6
+    *(--p) = 0; // xmm7
+    *(--p) = 0; // xmm7
+    *(--p) = 0; // xmm8
+    *(--p) = 0; // xmm8
+    *(--p) = 0; // xmm9
+    *(--p) = 0; // xmm9
+    *(--p) = 0; // xmm10
+    *(--p) = 0; // xmm10
+    *(--p) = 0; // xmm11
+    *(--p) = 0; // xmm11
+    *(--p) = 0; // xmm12
+    *(--p) = 0; // xmm12
+    *(--p) = 0; // xmm13
+    *(--p) = 0; // xmm13
+    *(--p) = 0; // xmm14
+    *(--p) = 0; // xmm14
+    *(--p) = 0; // xmm15
+    *(--p) = 0; // xmm15
+    *(--p) = (void*)0x00001f800000033f; // SIMD and regular floating point control defaults
+    
+    *stackPointer = p;
+}
diff --git a/src/backend/switchstack_gcc_32_linux_mac.s b/src/backend/switchstack_gcc_32_linux_mac.s
new file mode 100644
index 0000000..e4eac63
--- /dev/null
+++ b/src/backend/switchstack_gcc_32_linux_mac.s
@@ -0,0 +1,41 @@
+.text
+.globl _switchStackInternal
+
+_switchStackInternal:
+    // save callee-saved registers
+    push %ebp
+    movl %esp, %ebp
+    push %ebx
+    push %esi
+    push %edi
+
+    // store SIMD floating point control word
+    sub $4, %esp
+    stmxcsr (%esp)
+
+    // store floating point control bytes
+    sub $4, %esp
+    fstcw (%esp)
+
+    // save the old stack pointer
+    movl 0xc(%ebp), %edx
+    movl %esp, (%edx)
+    // set the new stack pointer
+    movl 0x8(%ebp), %esp
+
+    // restore floating point control bytes
+    fnclex
+    fldcw (%esp)
+    add $4, %esp
+
+    // restore SIMD floating point control word
+    ldmxcsr (%esp)
+    add $4, %esp
+
+    // pop callee-saved registers
+    pop %edi
+    pop %esi
+    pop %ebx
+    pop %ebp
+
+    ret
diff --git a/src/backend/switchstack_gcc_32_win.s b/src/backend/switchstack_gcc_32_win.s
new file mode 100644
index 0000000..2e8c140
--- /dev/null
+++ b/src/backend/switchstack_gcc_32_win.s
@@ -0,0 +1,42 @@
+.global _switchStackInternal 
+.section .text 
+.def _switchStackInternal ; .scl 2 ; .type 32 ; .endef 
+
+_switchStackInternal:
+    // save callee-saved registers
+    push %ebp
+    movl %esp, %ebp
+    push %ebx
+    push %esi
+    push %edi
+
+    // store SIMD floating point control word
+    sub $4, %esp
+    stmxcsr (%esp)
+
+    // store floating point control bytes
+    sub $4, %esp
+    fstcw (%esp)
+
+    // save the old stack pointer
+    movl 0xc(%ebp), %edx
+    movl %esp, (%edx)
+    // set the new stack pointer
+    movl 0x8(%ebp), %esp
+
+    // restore floating point control bytes
+    fnclex
+    fldcw (%esp)
+    add $4, %esp
+
+    // restore SIMD floating point control word
+    ldmxcsr (%esp)
+    add $4, %esp
+
+    // pop callee-saved registers
+    pop %edi
+    pop %esi
+    pop %ebx
+    pop %ebp
+    
+    ret
diff --git a/src/backend/switchstack_gcc_64_linux_mac.s b/src/backend/switchstack_gcc_64_linux_mac.s
new file mode 100644
index 0000000..ced6a5d
--- /dev/null
+++ b/src/backend/switchstack_gcc_64_linux_mac.s
@@ -0,0 +1,44 @@
+.text
+.globl _switchStackInternal
+
+_switchStackInternal:
+    // save callee-saved registers
+    push %rbp
+    movq %rsp, %rbp
+    push %rbx
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    // store SIMD floating point control word
+    sub $4, %rsp
+    stmxcsr (%rsp)
+
+    // store floating point control bytes
+    sub $4, %rsp
+    fstcw (%rsp)
+
+    // save the old stack pointer (second arg in rsi)
+    movq %rsp, (%rsi)
+    // set the new stack pointer (first arg in rdi)
+    movq %rdi, %rsp
+
+    // restore floating point control bytes
+    fnclex
+    fldcw (%rsp)
+    add $4, %rsp
+
+    // restore SIMD floating point control word
+    ldmxcsr (%rsp)
+    add $4, %rsp
+
+    // pop callee-saved registers
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %rbx
+    pop %rbp
+
+    retq
diff --git a/src/backend/switchstack_gcc_64_win.s b/src/backend/switchstack_gcc_64_win.s
new file mode 100644
index 0000000..f75c53c
--- /dev/null
+++ b/src/backend/switchstack_gcc_64_win.s
@@ -0,0 +1,93 @@
+.global _switchStackInternal
+.section .text 
+.def _switchStackInternal ; .scl 2 ; .type 32 ; .endef 
+
+_switchStackInternal:
+    // save callee-saved registers
+    push %rbp
+    movq %rsp, %rbp
+    push %rbx
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    push %rsi
+    push %rdi
+
+    sub $0x10, %rsp
+    movupd %xmm6, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm7, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm8, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm9, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm10, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm11, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm12, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm13, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm14, (%rsp)
+    sub $0x10, %rsp
+    movupd %xmm15, (%rsp)
+
+    // store SIMD floating point control word
+    sub $4, %rsp
+    stmxcsr (%rsp)
+
+    // store floating point control bytes
+    sub $4, %rsp
+    fstcw (%rsp)
+
+    // save the old stack pointer (second arg in rsi)
+    movq %rsp, (%rsi)
+    // set the new stack pointer (first arg in rdi)
+    movq %rdi, %rsp
+
+    // restore floating point control bytes
+    fnclex
+    fldcw (%rsp)
+    add $4, %rsp
+
+    // restore SIMD floating point control word
+    ldmxcsr (%rsp)
+    add $4, %rsp
+
+    // pop callee-saved registers
+    movupd (%rsp), %xmm15
+    add $0x10, %rsp
+    movupd (%rsp), %xmm14
+    add $0x10, %rsp
+    movupd (%rsp), %xmm13
+    add $0x10, %rsp
+    movupd (%rsp), %xmm12
+    add $0x10, %rsp
+    movupd (%rsp), %xmm11
+    add $0x10, %rsp
+    movupd (%rsp), %xmm10
+    add $0x10, %rsp
+    movupd (%rsp), %xmm9
+    add $0x10, %rsp
+    movupd (%rsp), %xmm8
+    add $0x10, %rsp
+    movupd (%rsp), %xmm7
+    add $0x10, %rsp
+    movupd (%rsp), %xmm6
+    add $0x10, %rsp
+
+    pop %rdi
+    pop %rsi
+
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %rbx
+    pop %rbp
+
+    retq
diff --git a/src/backend/switchstack_msvc_32.cpp b/src/backend/switchstack_msvc_32.cpp
new file mode 100644
index 0000000..eae68ee
--- /dev/null
+++ b/src/backend/switchstack_msvc_32.cpp
@@ -0,0 +1,44 @@
+extern "C" {
+void __declspec(naked) _switchStackInternal(void* to, void** from)
+{
+    __asm {
+    // save callee-saved registers
+    PUSH EBP
+    MOV EBP, ESP
+    PUSH EBX
+    PUSH ESI
+    PUSH EDI
+
+    // store SIMD floating point control word
+    SUB ESP, 4
+    STMXCSR [ESP]
+
+    // store floating point control bytes
+    SUB ESP, 4
+    FSTCW [ESP]
+
+    // save the old stack pointer
+    MOV EDX, dword ptr 12[EBP]
+    MOV [EDX], ESP
+    // set the new stack pointer
+    MOV ESP, dword ptr 8[EBP]
+
+    // restore floating point control bytes
+    FNCLEX
+    FLDCW [ESP]
+    ADD ESP, 4
+
+    // restore SIMD floating point control word
+    LDMXCSR [ESP]
+    ADD ESP, 4
+
+    // pop callee-saved registers
+    POP EDI
+    POP ESI
+    POP EBX
+    POP EBP        
+    
+    RET
+    }
+}
+}
diff --git a/src/backend/switchstack_msvc_64.asm b/src/backend/switchstack_msvc_64.asm
new file mode 100644
index 0000000..3c6f321
--- /dev/null
+++ b/src/backend/switchstack_msvc_64.asm
@@ -0,0 +1,98 @@
+.model flat, c
+.code
+
+_switchStackInternal PROC to:QWORD, from:QWORD
+    ; save callee-saved registers
+    PUSH RBP
+    MOV RBP, RSP
+    PUSH RBX
+    PUSH ESI
+    PUSH EDI
+
+    PUSH R12
+    PUSH R13
+    PUSH R14
+    PUSH R15
+
+    PUSH RSI
+    PUSH RDI
+
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM6
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM7
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM8
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM9
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM10
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM11
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM12
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM13
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM14
+    SUB RSP, 0x10
+    MOVUPD [RSP], XMM15
+
+    ; store SIMD floating point control word
+    SUB RSP, 4
+    STMXCSR [RSP]
+
+    ; store floating point control bytes
+    SUB RSP, 4
+    FSTCW [RSP]
+
+    ; save the old stack pointer
+    MOV [from], RSP
+    ; set the new stack pointer
+    MOV RSP, to
+
+    ; restore floating point control bytes
+    FNCLEX
+    FLDCW [RSP]
+    ADD RSP, 4
+
+    ; restore SIMD floating point control word
+    LDMXCSR [RSP]
+    ADD RSP, 4
+
+    ; pop callee-saved registers
+    MOVUPD XMM15, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM14, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM13, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM12, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM11, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM10, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM9, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM8, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM7, [RSP]
+    ADD RSP, 0x10
+    MOVUPD XMM6, [RSP]
+    ADD RSP, 0x10
+
+    POP RDI
+    POP RSI
+
+    POP R15
+    POP R14
+    POP R13
+    POP R12
+    POP RBX
+    POP RBP
+
+    RET
+_switchStackInternal ENDP
+
+end