added basic arm64 support

7 files changed, 144 insertions, 10 deletions

diff --git a/Makefile b/Makefile
index b3665cc..47c7c90 100644
--- a/Makefile
+++ b/Makefile
@@ -1,10 +1,16 @@
 # Copyright (c) 2020 Johannes Stoelp
 
-CXX      := g++
+ARCH ?= x86_64
+
+ifeq ($(ARCH),arm64)
+PREFIX := aarch64-linux-gnu-
+endif
+
+CXX      := $(PREFIX)g++
 CXXFLAGS := -g -O0 -Wall -Wextra -I.
 
 lib/libmatcha.a:
-	make -C lib
+	make -C lib ARCH=$(ARCH)
 
 %.o: %.cc
 	$(CXX) -c -o $@ $^ $(CXXFLAGS)
@@ -13,12 +19,29 @@ fmt:
 	fd --type f '.+\.h$$|.+\.cc$$' --exec clang-format -i {}
 
 example/demo1: example/demo1.o lib/libmatcha.a
-	$(CXX) -o $@ $^
+	$(CXX) -o $@ $^ $(LDFLAGS)
+ifeq ($(ARCH),arm64)
+	qemu-aarch64                                      \
+	    -L /usr/aarch64-linux-gnu                     \
+	    -E LD_LIBRARY_PATH=/usr/aarch64-linux-gnu/lib \
+	    example/demo1
+else
+	example/demo1
+endif
 
 gdb: example/demo1
+ifeq ($(ARCH),arm64)
+	qemu-aarch64                                      \
+	    -L /usr/aarch64-linux-gnu                     \
+	    -E LD_LIBRARY_PATH=/usr/aarch64-linux-gnu/lib \
+	    -g 1234                                       \
+	    example/demo1 &
+	gdb-multiarch -ex 'target remote :1234' -ex 'b main' -ex 'c' example/demo1
+else
 	which cgdb && cgdb -x util.gdb -ex 'start' example/demo1 \
 	           ||  gdb -x util.gdb -ex 'start' example/demo1
+endif
 
 clean:
-	rm -f example/demo1
-	rm -f **/*.o **/lib*.a
+	make -C lib clean
+	rm -f example/demo1 example/demo1.o
diff --git a/lib/Makefile b/lib/Makefile
index 78b051e..c880346 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -1,12 +1,16 @@
 # Copyright (c) 2020 Johannes Stoelp
 
-AS       := gcc
+ARCH ?= x86_64
+
+ifeq ($(ARCH),arm64)
+PREFIX := aarch64-linux-gnu-
+endif
+
+AS       := $(PREFIX)gcc
 ASFLAGS  := -g -O0
-CXX      := g++
+CXX      := $(PREFIX)g++
 CXXFLAGS := -g -O0 -Wall -Wextra -I.
-AR       := ar
-
-ARCH := x86_64
+AR       := $(PREFIX)ar
 
 libmatcha.a: thread.o executor.o                               \
              thread.h executor.h                               \
diff --git a/lib/arch.h b/lib/arch.h
index fc86cad..9604be8 100644
--- a/lib/arch.h
+++ b/lib/arch.h
@@ -8,6 +8,8 @@ static_assert(false, "Matcha Threads only supported on Linux!");
 
 #if defined(__x86_64__) || defined(__amd64__)
 #    include "arch/x86_64/api.h"
+#elif defined(__aarch64__)
+#    include "arch/arm64/api.h"
 #else
 static_assert(false, "Matcha Threads only supported on x86_64!");
 #endif
diff --git a/lib/arch/arm64/api.h b/lib/arch/arm64/api.h
new file mode 100644
index 0000000..b2babc9
--- /dev/null
+++ b/lib/arch/arm64/api.h
@@ -0,0 +1,6 @@
+/* Copyright (c) 2020 Johannes Stoelp */
+
+#pragma once
+
+extern "C" void yield(const void* new_stack, void* const* old_stack);
+void* init_stack(void* stack_ptr, void (*entry)(void*), void* ctx);
diff --git a/lib/arch/arm64/init_stack.cc b/lib/arch/arm64/init_stack.cc
new file mode 100644
index 0000000..b28d957
--- /dev/null
+++ b/lib/arch/arm64/init_stack.cc
@@ -0,0 +1,37 @@
+#include <cassert>
+#include <cstdint>  // uintN_t
+
+extern "C" void thread_create();
+
+void* init_stack(void* stack_ptr, void (*entry)(void*), void* ctx) {
+    static_assert(sizeof(uint64_t) == sizeof(std::uintptr_t), "Pointer must be 64bit!");
+
+    // Setup initial stack frame which will be popped when yielding
+    // first time into the thread.
+    // Basic idea is to yield into Thread::entry() function which will
+    // then call the user function.
+
+    uint64_t* stack = static_cast<uint64_t*>(stack_ptr);
+    // Arguments for `thread_create`.
+    *(--stack) = reinterpret_cast<uint64_t>(ctx);
+    *(--stack) = reinterpret_cast<uint64_t>(entry);
+
+    // Yield epilogue.
+    *(--stack) = reinterpret_cast<uint64_t>(thread_create);  // x30 (LR)
+    *(--stack) = static_cast<uint64_t>(0);                   // x29 (FP)
+
+    // Callee saved registers.
+    *(--stack) = static_cast<uint64_t>(0);  // x28
+    *(--stack) = static_cast<uint64_t>(0);  // x27
+    *(--stack) = static_cast<uint64_t>(0);  // x26
+    *(--stack) = static_cast<uint64_t>(0);  // x25
+    *(--stack) = static_cast<uint64_t>(0);  // x24
+    *(--stack) = static_cast<uint64_t>(0);  // x23
+    *(--stack) = static_cast<uint64_t>(0);  // x22
+    *(--stack) = static_cast<uint64_t>(0);  // x21
+    *(--stack) = static_cast<uint64_t>(0);  // x20
+    *(--stack) = static_cast<uint64_t>(0);  // x19
+
+    assert((reinterpret_cast<std::uintptr_t>(stack) & 0xf) == 0);  // 16byte aligned
+    return static_cast<void*>(stack);
+}
diff --git a/lib/arch/arm64/thread_create.s b/lib/arch/arm64/thread_create.s
new file mode 100644
index 0000000..dc6fae2
--- /dev/null
+++ b/lib/arch/arm64/thread_create.s
@@ -0,0 +1,20 @@
+# Copyright (c) 2020 Johannes Stoelp
+
+    .arch armv8-a
+    .section .text, "ax", @progbits
+
+     # extern "C" void thread_create();
+    .global thread_create
+    .type   thread_create, @function
+thread_create:
+    .cfi_startproc
+    ldr x0, [sp, 0x8]
+    ldr x1, [sp]
+
+    blr x1
+
+    # FIXME: no return from thread after user fn finished.
+1:
+    b 1b
+    .cfi_endproc
+    .size thread_create, .-thread_create
diff --git a/lib/arch/arm64/yield.s b/lib/arch/arm64/yield.s
new file mode 100644
index 0000000..63a4051
--- /dev/null
+++ b/lib/arch/arm64/yield.s
@@ -0,0 +1,42 @@
+# Copyright (c) 2020 Johannes Stoelp
+
+    .arch armv8-a
+    .section .text, "ax", @progbits
+
+    # extern "C" void yield(const void* new_stack, void* const* old_stack);
+    #                                   ^^^^^^^^^               ^^^^^^^^^
+    #                                   rdi                     rsi
+    .global yield
+    .type   yield, @function
+yield:
+    .cfi_startproc
+    // prologue
+    stp x29, x30, [sp, -16]!
+    mov x29, sp
+
+    // push callee saved registers
+    stp x27, x28, [sp, -16]!
+    stp x25, x26, [sp, -16]!
+    stp x23, x24, [sp, -16]!
+    stp x21, x22, [sp, -16]!
+    stp x19, x20, [sp, -16]!
+
+    // arg0: x0 holds new stack
+    // arg1: x1 holds addr to location current stack must be saved
+    mov x2, sp
+    str x2, [x1]  // save current stack ptr
+    mov sp, x0    // switch to new stack ptr
+
+    // pop callee saved registers
+    ldp x19, x20, [sp], 16
+    ldp x21, x22, [sp], 16
+    ldp x23, x24, [sp], 16
+    ldp x25, x26, [sp], 16
+    ldp x27, x28, [sp], 16
+
+    // epilogue
+    ldp x29, x30, [sp], 16
+
+    ret
+    .cfi_endproc
+    .size yield, .-yield