Add custom context switch library.

The rationale to do an custom implementation is that the existing solutions are quite a bit slower and/or require more memory.

CMakeLists.txt

@@ -32,6 +32,7 @@ add_subdirectory(extern/benchmark_base)
 add_subdirectory(extern/benchmark_runner)
 
 # Include all internal subprojects (library, examples, testing).
+add_subdirectory(lib/context_switcher)
 add_subdirectory(lib/pls)
 
 # Include examples

app/context_switch/main.cpp

@@ -15,7 +15,7 @@ const size_t STACK_SIZE = 512 * 1;
 const char MAGIC_NUMBER = (unsigned char) 0xAB;
 
 // Memory for custom stack and continuation semantics
-char custom_stack[STACK_SIZE] = {0};
+char custom_stack_1[STACK_SIZE] = {0};
 jmp_buf buffer;
 
 // Example callback function and declaration of our assembly stack switching routine
@@ -50,7 +50,7 @@ long measure_function_call() {
 long measure_stack_switch() {
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
-    custom_stack_callback(&custom_stack[STACK_SIZE]);
+    custom_stack_callback(&custom_stack_1[STACK_SIZE]);
   }
   auto end_time = chrono::steady_clock::now();
   return chrono::duration_cast<chrono::nanoseconds>(end_time - start_time).count();
@@ -60,7 +60,7 @@ long measure_continuation() {
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
     if (setjmp(buffer) == 0) {
-      custom_stack_callback(&custom_stack[STACK_SIZE]);
+      custom_stack_callback(&custom_stack_1[STACK_SIZE]);
     }
   }
   auto end_time = chrono::steady_clock::now();
@@ -71,7 +71,7 @@ long measure_continuation_and_jump() {
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
     if (setjmp(buffer) == 0) {
-      custom_stack_callback(&custom_stack[STACK_SIZE]);
+      custom_stack_callback(&custom_stack_1[STACK_SIZE]);
       longjmp(buffer, 1);
     }
   }
@@ -87,7 +87,7 @@ void fcontext_callback_fast(fcontext_transfer_t transfer) {
 }
 
 long measure_fcontext_fast() {
-  fcontext_t context = make_fcontext(&custom_stack[STACK_SIZE], STACK_SIZE, &fcontext_callback_fast);
+  fcontext_t context = make_fcontext(&custom_stack_1[STACK_SIZE], STACK_SIZE, &fcontext_callback_fast);
 
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
@@ -105,7 +105,7 @@ void fcontext_callback_clean(fcontext_transfer_t transfer) {
 long measure_fcontext_clean() {
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
-    fcontext_t context = make_fcontext(&custom_stack[STACK_SIZE], STACK_SIZE, &fcontext_callback_clean);
+    fcontext_t context = make_fcontext(&custom_stack_1[STACK_SIZE], STACK_SIZE, &fcontext_callback_clean);
     jump_fcontext(context, nullptr);
   }
   auto end_time = chrono::steady_clock::now();
@@ -120,7 +120,7 @@ void fcontext_callcc(fcontext_transfer_t transfer) {
 long measure_fcontext_callcc() {
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
-    fcontext_t context = make_fcontext(&custom_stack[STACK_SIZE], STACK_SIZE, &fcontext_callcc);
+    fcontext_t context = make_fcontext(&custom_stack_1[STACK_SIZE], STACK_SIZE, &fcontext_callcc);
     jump_fcontext(jump_fcontext(context, nullptr).ctx, nullptr);
   }
   auto end_time = chrono::steady_clock::now();
@@ -132,7 +132,7 @@ long measure_custom() {
 
   auto start_time = chrono::steady_clock::now();
   for (unsigned int i = 0; i < NUM_RUNS; i++) {
-    fiber_call(custom_stack, STACK_SIZE, [](continuation_t continuation) {
+    fiber_call(custom_stack_1, STACK_SIZE, [](continuation_t continuation) {
       callback();
       return continuation;
     });
@@ -142,7 +142,7 @@ long measure_custom() {
 }
 
 int main() {
-  memset(custom_stack, MAGIC_NUMBER, STACK_SIZE);
+  memset(custom_stack_1, MAGIC_NUMBER, STACK_SIZE);
 
   auto time_cont_jump = measure_continuation_and_jump();
   auto time_cont = measure_continuation();
@@ -169,7 +169,7 @@ int main() {
   printf("Custom Fast Call : %10ld, %5.5f\n", time_custom, ((float) time_custom / NUM_RUNS));
 
   for (unsigned int i = 0; i < STACK_SIZE; i++) {
-    if (custom_stack[i] != MAGIC_NUMBER) {
+    if (custom_stack_1[i] != MAGIC_NUMBER) {
       printf("\n\nUsed stack size about %u bytes.\n", (STACK_SIZE - i));
       break;
     }

app/playground/CMakeLists.txt

 add_executable(playground main.cpp)
 
 # Example for adding the library to your app (as a cmake project dependency)
-target_link_libraries(playground)
+target_link_libraries(playground context_switcher)

app/playground/main.cpp

+#include <utility>
 #include <cstdio>
-#include <csetjmp>
+
+#include "context_switcher/context_switcher.h"
+
+// Memory for custom stack and continuation semantics
+const size_t STACK_SIZE = 512 * 8;
+char custom_stack_1[STACK_SIZE];
+
+// Force disable optimization
+volatile int value = 0;
 
 int main() {
-  printf("Buffer Size: %u\n", sizeof(jmp_buf));
-  printf("Big Buffer Size: %u\n", sizeof(sigjmp_buf));
+  using namespace context_switcher;
 
+  printf("Main 1!\n");
+  auto cont_2 = enter_context(custom_stack_1, STACK_SIZE, [](continuation &&cont_main) {
+    printf("Stack 1!\n");
+    cont_main = switch_context(std::move(cont_main));
+    printf("Stack 2!\n");
+    return std::move(cont_main);
+  });
+  printf("Main 2!\n");
+  cont_2 = switch_context(std::move(cont_2));
+  printf("Main 3!\n");
   return 0;
 }

lib/context_switcher/CMakeLists.txt

+cmake_minimum_required(VERSION 3.10)
+project(context_switcher
+        VERSION 0.0.1
+        DESCRIPTION "allows to execute functions and lambdas on a new stack and switch between them"
+        LANGUAGES CXX ASM)
+
+set(CMAKE_CXX_STANDARD 11)
+
+# Platform Support - Edit this when porting the context switch facility.
+# We are rather conservative with our flags, maybe more systems follow implemented calling conventions.
+# See our documentation and boost context for help when adding a new system.
+if (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" AND CMAKE_SYSTEM_NAME STREQUAL "Linux")
+    SET(CONTEXT_SWITCH_ASSEMBLY "asm/enter_context_x86_64.s" "asm/switch_context_x86_64.s")
+elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7l" AND CMAKE_SYSTEM_NAME STREQUAL "Linux")
+    SET(CONTEXT_SWITCH_ASSEMBLY "asm/enter_context_arm32.s" "asm/switch_context_arm32.s")
+else ()
+    MESSAGE(FATAL_ERROR "Platform (${CMAKE_SYSTEM_PROCESSOR} on ${CMAKE_SYSTEM_NAME}) not supported! Please see Readme for instructions to port.")
+endif ()
+message("-- Context Switcher: ${CMAKE_SYSTEM_PROCESSOR} running ${CMAKE_SYSTEM_NAME}")
+
+add_library(context_switcher STATIC
+        ${CONTEXT_SWITCH_ASSEMBLY}
+        include/context_switcher/context_switcher.h src/context_switcher.cpp include/context_switcher/assembly_bindings.h include/context_switcher/continuation.h include/context_switcher/lambda_capture.h)
+
+# Add everything in `./include` to be in the include path of this project
+target_include_directories(context_switcher
+        PUBLIC
+        $<INSTALL_INTERFACE:include>
+        $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+        PRIVATE
+        ${CMAKE_CURRENT_SOURCE_DIR}/src
+        )
+

lib/context_switcher/asm/enter_context_arm32.s

+	.arm
+	.text
+	.global	__cs_enter_context
+	.type	__cs_enter_context, %function
+
+__cs_enter_context:
+        /* Parameter List (in order)
+        * r0 = new stack pointer
+        * r1 = first parameter to callback
+        * r2 = callback function pointer
+        * r3 = new stack limit (not used on most platforms)
+        *
+        * Return
+        * r0 = continuation that returned control back to the caller (null if fallthrough)
+        *
+        * Variables
+        * r4 = temporary for the old stack pointer */
+
+        /* ========== Save State ========== */
+        /* store programm counter for later return */
+        push {lr}
+        /* store callee saved registers */
+        push {r4-r12,lr}
+        /* ========== Save State ========== */
+
+        /* Perform change to new stack */
+        /* Keep old stack as second parameter to our callback function. */
+        mov r4, sp
+        /* Make sure that stack start is properly aligned. */
+        and r0, r0, #-16
+        /* Switch to new stack pointer. */
+        mov sp, r0
+
+        /* Perform actual function call, this will now be on the new stack */
+        /* r0 = first parametor to callback (continuation) */
+        /* r1 = second parameter to callback (arbetary pointer) */
+        mov r0, r4
+        blx r2
+
+        /* Restore state of returned continuation. */
+        /* To do so we first reset the stack pointer (which we get returned in r0). */
+        /* After that we execute our standard restore procedere to pop the state from the stack. */
+        mov sp, r0
+
+        /* ========== Restore State ========== */
+        /* restore callee saved registers */
+        pop {r4-r12,lr}
+        /* ========== Restore State ========== */
+
+        /* Just return back from the call. */
+        /* This is the end of a fiber, so we have no continuation. */
+        eor r0, r0, r0
+        pop {pc}

lib/context_switcher/asm/enter_context_x86_64.s

+	.file	"enter_context_x86_64.s"
+	.text
+	.global	__cs_enter_context
+	.type	__cs_enter_context, @function
+
+.align 16
+__cs_enter_context:
+        # Parameter List (in order)
+        # rdi = new stack pointer
+        # rsi = first parameter to callback
+        # rdx = callback function pointer
+        # rcx = new stack limit (not used on most platforms)
+
+        # Return
+        # rax = continuation that returned control back to the caller (null if fallthrough)
+
+        # Variables
+        # r12 = temporary for the old stack pointer
+
+        ############### Save State ###############
+        # 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)
+        ############### Save State ###############
+
+        # Perform change to new stack.
+        # Keep old stack as second parameter to our callback function.
+        movq %rsp, %r12
+        # Make sure that stack start is properly aligned.
+        andq $-16, %rdi
+        # Switch to new stack pointer.
+        movq %rdi, %rsp
+
+        # Perform actual function call, this will now be on the new stack
+        # rdi = first parametor to callback (continuation)
+        # rsi = second parameter to callback (arbetary pointer)
+        movq %r12, %rdi
+        call *%rdx
+
+        # Restore state of returned continuation.
+        # To do so we first reset the stack pointer (which we get returned in rax).
+        # After that we execute our standard restore procedere to pop the state from the stack.
+        movq %rax, %rsp
+
+        ############ Restore State  ############
+        # 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
+        ############ Restore State  ############
+
+        # TODO: Maybe look into a 'cleanup' hook for freeing the stack space here.
+
+        # Just return back from the call.
+        # This is the end of a fiber, so we have no continuation.
+        xor %rax, %rax
+        ret

lib/context_switcher/asm/switch_context_arm32.s

+	.arm
+	.text
+	.global	__cs_switch_context
+	.type	__cs_switch_context, %function
+
+__cs_switch_context:
+        /* Parameter List (in order)
+        * r0 = pointer to continuation (should hold value of target stack will be filled with this continuation)
+        *
+        * Return
+        * r0 = continuation that returned control back to the caller (null if fallthrough)
+        *
+        * Variables
+        * r1 = temporary for the old stack pointer */
+
+        /* ========== Save State ========== */
+        /* store programm counter for later return */
+        push {lr}
+        /* store callee saved registers */
+        push {r4-r12,lr}
+        /* ========== Save State ========== */
+
+        /* Perform change to new stack */
+        /* Keep old stack as result from this function. */
+        mov r1, sp
+        /* Switch to new stack pointer. */
+        mov sp, r0
+
+
+        /* ========== Restore State ========== */
+        /* restore callee saved registers */
+        pop {r4-r12,lr}
+        /* ========== Restore State ========== */
+
+        /* Just return back from the call. */
+        /* This is the end of a fiber, so we have no continuation. */
+        mov r0, r1
+        pop {pc}

lib/context_switcher/asm/switch_context_x86_64.s

+	.file	"switch_context_x86_64.s"
+	.text
+	.global	__cs_switch_context
+	.type	__cs_switch_context, @function
+
+.align 16
+__cs_switch_context:
+        # Parameter List (in order)
+        # rdi = pointer to continuation (should hold value of target stack will be filled with this continuation)
+
+        # Return
+        # rax = continuation that returned control back to the caller (null if fallthrough)
+
+        ############### Save State ###############
+        # 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)
+        ############### Save State ###############
+
+        # Perform change to new stack.
+        # Keep old stack as result from this function
+        movq %rsp, %rax
+        # switch to new stack pointer
+        movq %rdi, %rsp
+
+        ############ Restore State  ############
+        # 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
+        ############ Restore State  ############
+
+        # Return the context we came from as a continuation.
+        # rax has already the correct value
+        ret

lib/context_switcher/include/context_switcher/assembly_bindings.h

+
+#ifndef CONTEXT_SWITCHER_ASSEMBLY_BINDINGS_H_
+#define CONTEXT_SWITCHER_ASSEMBLY_BINDINGS_H_
+
+/**
+ * Low level bindings to the two assembly functions used to switch context.
+ * Can be used standalone to build new control structures, but should be handled with care.
+ *
+ * Basic usage:
+ * // Switch control to a new stack.
+ * continuation_t cont = __cs_enter_context(my_stack, arg_pointer, callback_func, my_stack_limit);
+ * // Will only be run when 'jumped back' by either returning the callback or an explicit jump.
+ * // cont will then hold the context of the jump origin (or null if it was a return form a finished callback).
+ * cont = __switch_context(cont);
+ * // Re-Enter the context that jumped to us.
+ */
+
+namespace context_switcher {
+namespace assembly_bindings {
+
+using continuation_t = void *;
+using stack_pointer_t = char *;
+using callback_t = continuation_t (*)(continuation_t, void *);
+
+extern "C" {
+continuation_t __cs_enter_context(stack_pointer_t stack_base,
+                                  void *callback_arg,
+                                  callback_t callback,
+                                  stack_pointer_t stack_limit);
+continuation_t __cs_switch_context(continuation_t continuation);
+}
+
+}
+}
+
+#endif //CONTEXT_SWITCHER_ASSEMBLY_BINDINGS_H_

lib/context_switcher/include/context_switcher/context_switcher.h

+#ifndef CONTEXT_SWITCHER_CONTEXT_SWITCHER_H_
+#define CONTEXT_SWITCHER_CONTEXT_SWITCHER_H_
+
+#include <utility>
+#include <cstdio>
+#include <cstdint>
+#include <new>
+
+#include "assembly_bindings.h"
+#include "continuation.h"
+#include "lambda_capture.h"
+
+namespace context_switcher {
+
+template<typename F>
+continuation enter_context(assembly_bindings::stack_pointer_t stack_memory, size_t stack_size, F &&lambda) {
+  assembly_bindings::stack_pointer_t lambda_memory = stack_memory + stack_size - sizeof(lambda_capture<F>);
+  auto *captured_lambda = place_lambda_capture(std::forward<F>(lambda), lambda_memory);
+
+  assembly_bindings::stack_pointer_t stack_base = lambda_memory;
+  assembly_bindings::stack_pointer_t stack_limit = stack_memory;
+
+  assembly_bindings::callback_t callback = lambda_capture_callback<lambda_capture<F>>;
+  return continuation{assembly_bindings::__cs_enter_context(stack_base, captured_lambda, callback, stack_limit)};
+}
+
+continuation switch_context(continuation &&cont) {
+  assembly_bindings::continuation_t cont_pointer = cont.consume();
+
+  return continuation{assembly_bindings::__cs_switch_context(cont_pointer)};
+}
+
+}
+
+#endif //CONTEXT_SWITCHER_CONTEXT_SWITCHER_H_

lib/context_switcher/include/context_switcher/continuation.h

+
+#ifndef CONTEXT_SWITCHER_CONTINUATION_H_
+#define CONTEXT_SWITCHER_CONTINUATION_H_
+
+#include "assembly_bindings.h"
+
+namespace context_switcher {
+
+/**
+ * One-Shot continuation. Represents the paused state of an execution thread.
+ * Can be used exactly once to jump back to that state.
+ *
+ * Move only to ensure semantics of one time use.
+ */
+struct continuation {
+ public:
+  continuation() : cont_pointer_{nullptr} {};
+  explicit continuation(assembly_bindings::continuation_t cont_pointer) : cont_pointer_{cont_pointer} {};
+
+  // Move-Only Semantics
+  continuation(const continuation &other) = delete;
+  continuation(continuation &&other) noexcept {
+    cont_pointer_ = other.cont_pointer_;
+    other.cont_pointer_ = nullptr;
+  }
+
+  continuation &operator=(const continuation &other) = delete;
+  continuation &operator=(continuation &&other) noexcept {
+    cont_pointer_ = other.cont_pointer_;
+    other.cont_pointer_ = nullptr;
+
+    return *this;
+  }
+
+  // Semantics as 'one time use'
+  bool valid() const {
+    return cont_pointer_ != nullptr;
+  }
+
+  assembly_bindings::continuation_t consume() {
+    auto tmp = cont_pointer_;
+    cont_pointer_ = nullptr;
+    return tmp;
+  }
+
+ private:
+  assembly_bindings::continuation_t cont_pointer_;
+};
+
+}
+
+#endif //CONTEXT_SWITCHER_CONTINUATION_H_

lib/context_switcher/include/context_switcher/lambda_capture.h

+
+#ifndef CONTEXT_SWITCHER_LAMBDA_CAPTURE_H_
+#define CONTEXT_SWITCHER_LAMBDA_CAPTURE_H_
+
+#include <utility>
+#include <new>
+#include <type_traits>
+
+#include "assembly_bindings.h"
+#include "continuation.h"
+
+/**
+ * Helpers to more easily use a lambda expression as the code run on a context switch.
+ * Captures lambdas by placing them in a memory region and offering a conversion form
+ * the assembly_bindings callback API to the user's lambda.
+ */
+
+namespace context_switcher {
+
+template<typename F>
+struct lambda_capture {
+  explicit lambda_capture(F &&lambda) : lambda_{std::forward<F>(lambda)} {}
+
+  assembly_bindings::continuation_t operator()(assembly_bindings::continuation_t continuation_pointer) {
+    continuation cont = lambda_(continuation{continuation_pointer});
+    return cont.consume();
+  }
+
+ private:
+  F lambda_;
+};
+
+template<typename T>
+assembly_bindings::continuation_t lambda_capture_callback(assembly_bindings::continuation_t continuation_pointer,
+                                                          void *lambda_capture_param) {
+  // Perform Call
+  T *lambda_capture = reinterpret_cast<T *>(lambda_capture_param);
+  auto result_continuation = (*lambda_capture)(continuation_pointer);
+
+  // Free resources and switch to result_continuation (this execution thread is finished with the return)
+  lambda_capture->~T();
+  return result_continuation;
+}
+
+template<typename F>
+static lambda_capture<F> *place_lambda_capture(F &&lambda, char *memory) {
+  return new(memory) lambda_capture<F>(std::forward<F>(lambda));
+}
+
+}
+
+#endif //CONTEXT_SWITCHER_LAMBDA_CAPTURE_H_

lib/context_switcher/src/context_switcher.cpp

+

lib/pls/CMakeLists.txt

@@ -58,7 +58,7 @@ add_library(pls STATIC
         include/pls/internal/scheduling/task.h src/internal/scheduling/task.cpp
         include/pls/internal/scheduling/cont_manager.h
         include/pls/internal/scheduling/cont.h
-        include/pls/internal/data_structures/bounded_ws_deque.h include/pls/internal/data_structures/optional.h include/pls/internal/scheduling/memory_block.h include/pls/internal/scheduling/thread_state_static.h src/internal/base/error_handling.cpp include/pls/internal/data_structures/bounded_trading_deque.h)
+        include/pls/internal/data_structures/bounded_ws_deque.h include/pls/internal/data_structures/optional.h include/pls/internal/scheduling/memory_block.h include/pls/internal/scheduling/thread_state_static.h src/internal/base/error_handling.cpp include/pls/internal/data_structures/bounded_trading_deque.h ../context_switcher/src/context_switcher.cpp)
 
 # Add everything in `./include` to be in the include path of this project
 target_include_directories(pls