Add alternative context switch implementation (similar to boost).

CMakeLists.txt

@@ -44,6 +44,7 @@ add_subdirectory(app/benchmark_matrix)
 add_subdirectory(app/benchmark_prefix)
 add_subdirectory(app/benchmark_prefix)
 add_subdirectory(app/benchmark_pipeline)
 add_subdirectory(app/benchmark_pipeline)
 add_subdirectory(app/benchmark_fib)
 add_subdirectory(app/benchmark_fib)
+add_subdirectory(app/context_switch)
 
 
 # Add optional tests
 # Add optional tests
 option(PACKAGE_TESTS "Build the tests" ON)
 option(PACKAGE_TESTS "Build the tests" ON)

app/context_switch/CMakeLists.txt

@@ -7,7 +7,7 @@ else ()
 endif ()
 endif ()
 
 
 
 
-add_executable(playground main.cpp ${SWITCH_ASSEMBLY})
+add_executable(context_switch main.cpp ${SWITCH_ASSEMBLY} fiber_call_x86_64.s fiber_continue_x86_64.s)
 
 
 # Example for adding the library to your app (as a cmake project dependency)
 # Example for adding the library to your app (as a cmake project dependency)
-target_link_libraries(playground)
+target_link_libraries(context_switch)

app/context_switch/fiber_call_x86_64.s

+	.file	"fiber_call_x86_64.s"
+	.text
+	.global	fiber_call
+	.type	fiber_call, @function
+
+.align 16
+fiber_call:
+        # rdi = address of new stack top and to store current stack top (parameter to function)
+        # r12 temporary for restoring old stack (callee saved, so we can take a 'fast' return path)
+
+        ############### Register State Storage ################
+        # make space for all register state we will store
+        leaq -0x38(%rsp), %rsp
+
+        # store calee saved general registers
+        movq %r12, 0x00(%rsp)
+        movq %r13, 0x08(%rsp)
+        movq %r14, 0x10(%rsp)
+        movq %r15, 0x18(%rsp)
+        movq %rbx, 0x20(%rsp)
+        movq %rbp, 0x28(%rsp)
+        # store MMX control- and status-word
+        stmxcsr 0x30(%rsp)
+        # store x87 control-word
+        fnstcw 0x34(%rsp)
+
+        ############### Stack Pointer Manipulation #############
+        # keep the old stack pointer in a callee-saved register for 'fast' return path
+        movq %rsp, %r12
+        # switch to new stack pointer
+        movq (%rdi), %rsp
+        # store old stack pointer
+        movq %r12, (%rdi)
+
+        ############# Function Call ############################
+        # perform actual function call, this will now be on the new stack
+        call callback
+        # If someone else continued/jumped into this stored context we must not return here,
+        # thus we will never execute the rest of the function.
+        # Otherwise we can return out of the function as if it was a simple function call.
+
+        ############ Stack Pointer Manipulation ################
+        # restore stack pointer that we had before
+        movq %r12, %rsp
+
+        ############ Restore State Storage (as needed) #########
+        # ...we don't need to restore most registers (as they are calee saved).
+        # We only used %r12, so restore that and move back the stack pointer.
+        movq 0x00(%rsp), %r12
+
+        # release the space we used for storage
+        leaq 0x38(%rsp), %rsp
+
+        # just return back from the call
+        ret

app/context_switch/fiber_continue_x86_64.s

+	.file	"fiber_continue_x86_64.s"
+	.text
+	.global	fiber_continue
+	.type	fiber_continue, @function
+
+.align 16
+fiber_continue:
+        # rdi = address of new stack top and to store current stack top (parameter to function)
+
+        ############### Register State Storage ################
+        # make space for all register state we will store
+        leaq -0x38(%rsp), %rsp
+
+        # store calee saved general registers
+        movq %r12, 0x00(%rsp)
+        movq %r13, 0x08(%rsp)
+        movq %r14, 0x10(%rsp)
+        movq %r15, 0x18(%rsp)
+        movq %rbx, 0x20(%rsp)
+        movq %rbp, 0x28(%rsp)
+        # store MMX control- and status-word
+        stmxcsr 0x30(%rsp)
+        # store x87 control-word
+        fnstcw 0x34(%rsp)
+
+        ############### Stack Pointer Manipulation #############
+        # keep the old stack
+        movq %rsp, %r12
+        # switch to new stack pointer
+        movq (%rdi), %rsp
+        # store old stack pointer
+        movq %r12, (%rdi)
+
+        # We stored our old execution context and are now on the target stack
+        # on which we would like to contiune our execution.
+
+        ############ Restore State Storage (as needed) #########
+        # restore calee saved general registers
+        movq 0x00(%rsp), %r12
+        movq 0x08(%rsp), %r13
+        movq 0x10(%rsp), %r14
+        movq 0x18(%rsp), %r15
+        movq 0x20(%rsp), %rbx
+        movq 0x28(%rsp), %rbp
+        # restore MMX control- and status-word
+        ldmxcsr 0x30(%rsp)
+        # restore x87 control-word
+        fldcw 0x34(%rsp)
+
+        # Free space for restored state
+        leaq 0x38(%rsp), %rsp
+
+        # just return back from the call
+        ret

app/context_switch/main.cpp

@@ -7,7 +7,7 @@ using namespace std;
 
 
 // Settings for stack and benchmark
 // Settings for stack and benchmark
 const unsigned int NUM_RUNS = 100000;
 const unsigned int NUM_RUNS = 100000;
-const unsigned int STACK_SIZE = 512 * 1;
+const unsigned int STACK_SIZE = 512 * 8;
 const unsigned char MAGIC_NUMBER = (unsigned char) 0xAB;
 const unsigned char MAGIC_NUMBER = (unsigned char) 0xAB;
 
 
 // Memory for custom stack and continuation semantics
 // Memory for custom stack and continuation semantics
@@ -18,6 +18,9 @@ jmp_buf buffer;
 extern "C" {
 extern "C" {
 void custom_stack_callback(void *);
 void custom_stack_callback(void *);
 
 
+void fiber_call(void *);
+void fiber_continue(void *);
+
 void __attribute__ ((noinline)) callback() {
 void __attribute__ ((noinline)) callback() {
   static volatile int tmp;
   static volatile int tmp;
   tmp = 0; // Force at least a single memory write
   tmp = 0; // Force at least a single memory write
@@ -65,6 +68,16 @@ long __attribute__ ((noinline)) measure_continuation_and_jump() {
   return chrono::duration_cast<chrono::nanoseconds>(end_time - start_time).count();
   return chrono::duration_cast<chrono::nanoseconds>(end_time - start_time).count();
 }
 }
 
 
+long __attribute__ ((noinline)) measure_fiber_call() {
+  auto start_time = chrono::steady_clock::now();
+  for (unsigned int i = 0; i < NUM_RUNS; i++) {
+    void *stack_pointer = custom_stack;
+    fiber_call(&stack_pointer);
+  }
+  auto end_time = chrono::steady_clock::now();
+  return chrono::duration_cast<chrono::nanoseconds>(end_time - start_time).count();
+}
+
 int main() {
 int main() {
   memset(custom_stack, MAGIC_NUMBER, STACK_SIZE);
   memset(custom_stack, MAGIC_NUMBER, STACK_SIZE);
 
 
@@ -72,6 +85,7 @@ int main() {
   auto time_cont = measure_continuation();
   auto time_cont = measure_continuation();
   auto time_stack = measure_stack_switch();
   auto time_stack = measure_stack_switch();
   auto time_func = measure_function_call();
   auto time_func = measure_function_call();
+  auto time_fiber = measure_fiber_call();
 
 
   for (unsigned int i = 0; i < STACK_SIZE; i++) {
   for (unsigned int i = 0; i < STACK_SIZE; i++) {
     if (custom_stack[i] != MAGIC_NUMBER) {
     if (custom_stack[i] != MAGIC_NUMBER) {
@@ -84,6 +98,7 @@ int main() {
   printf("Stack Switching  : %10ld, %5.5f\n", time_stack, ((float) time_stack / NUM_RUNS));
   printf("Stack Switching  : %10ld, %5.5f\n", time_stack, ((float) time_stack / NUM_RUNS));
   printf("Full Continuation: %10ld, %5.5f\n", time_cont, ((float) time_cont / NUM_RUNS));
   printf("Full Continuation: %10ld, %5.5f\n", time_cont, ((float) time_cont / NUM_RUNS));
   printf("Jump Continuation: %10ld, %5.5f\n", time_cont_jump, ((float) time_cont_jump / NUM_RUNS));
   printf("Jump Continuation: %10ld, %5.5f\n", time_cont_jump, ((float) time_cont_jump / NUM_RUNS));
+  printf("Fiber Call       : %10ld, %5.5f\n", time_fiber, ((float) time_fiber / NUM_RUNS));
 
 
   return 0;
   return 0;
 }
 }

app/playground/main.cpp

+#include <cstdio>
+#include <csetjmp>
 
 
 int main() {
 int main() {
+  printf("Buffer Size: %u\n", sizeof(jmp_buf));
+  printf("Big Buffer Size: %u\n", sizeof(sigjmp_buf));
 
 
   return 0;
   return 0;
 }
 }