From c58c1a79e8bb718967d256929ac0ab164d0d3d84 Mon Sep 17 00:00:00 2001
From: Doug Coleman <doug.coleman@gmail.com>
Date: Mon, 23 Nov 2020 19:01:26 -0600
Subject: [PATCH] vm: Add a bunch of half-baked arm files. Also fix some .h
 issues.

---
 vm/Config.arm            |   3 +-
 vm/Config.arm64.64       |   1 +
 vm/Config.linux.arm.32   |   4 ++
 vm/Config.linux.arm.64   |   4 ++
 vm/Config.macosx         |   4 +-
 vm/Config.macosx.arm64   |   5 ++
 vm/Config.macosx.x86.32  |   3 +
 vm/Config.macosx.x86.64  |   6 ++
 vm/Config.macosx.x86.fat |   1 +
 vm/Config.unix           |   4 +-
 vm/Config.windows.x86.32 |   2 +-
 vm/Config.windows.x86.64 |   2 +-
 vm/cpu-arm.32.S          | 127 +++++++++++++++++++++++++++++++++++++++
 vm/cpu-arm.32.hpp        |   8 +++
 vm/cpu-arm.64.cpp        |  25 ++++++++
 vm/cpu-arm.64.hpp        |  13 ++++
 vm/cpu-arm.hpp           |  67 ++++++++++++++++++++-
 17 files changed, 269 insertions(+), 10 deletions(-)
 create mode 100644 vm/Config.arm64.64
 create mode 100644 vm/Config.linux.arm.32
 create mode 100644 vm/Config.linux.arm.64
 create mode 100644 vm/Config.macosx.arm64
 create mode 100644 vm/cpu-arm.32.S
 create mode 100644 vm/cpu-arm.32.hpp
 create mode 100644 vm/cpu-arm.64.cpp
 create mode 100644 vm/cpu-arm.64.hpp

diff --git a/vm/Config.arm b/vm/Config.arm
index 958a9b7c6d..b5a14149b3 100644
--- a/vm/Config.arm
+++ b/vm/Config.arm
@@ -1,2 +1 @@
-PLAF_MASTER_HEADERS += \
-	cpu-arm.hpp
+PLAF_MASTER_HEADERS += vm/cpu-arm.hpp
diff --git a/vm/Config.arm64.64 b/vm/Config.arm64.64
new file mode 100644
index 0000000000..b5a14149b3
--- /dev/null
+++ b/vm/Config.arm64.64
@@ -0,0 +1 @@
+PLAF_MASTER_HEADERS += vm/cpu-arm.hpp
diff --git a/vm/Config.linux.arm.32 b/vm/Config.linux.arm.32
new file mode 100644
index 0000000000..1677219341
--- /dev/null
+++ b/vm/Config.linux.arm.32
@@ -0,0 +1,4 @@
+include vm/Config.linux
+include vm/Config.arm
+PLAF_DLL_OBJS += vm/cpu-arm.32.o
+PLAF_MASTER_HEADERS += vm/cpu-arm.32.hpp
\ No newline at end of file
diff --git a/vm/Config.linux.arm.64 b/vm/Config.linux.arm.64
new file mode 100644
index 0000000000..d29c988fee
--- /dev/null
+++ b/vm/Config.linux.arm.64
@@ -0,0 +1,4 @@
+include vm/Config.linux
+include vm/Config.arm
+PLAF_DLL_OBJS += vm/cpu-arm.64.o
+PLAF_MASTER_HEADERS += vm/cpu-arm.64.hpp
\ No newline at end of file
diff --git a/vm/Config.macosx b/vm/Config.macosx
index 74f4fcacef..67cd585827 100644
--- a/vm/Config.macosx
+++ b/vm/Config.macosx
@@ -1,6 +1,4 @@
 include vm/Config.unix
-#SITE_CFLAGS +=
-SITE_CFLAGS += -mmacosx-version-min=10.9
 
 PLAF_DLL_OBJS += vm/os-macosx.o vm/mach_signal.o vm/mvm-unix.o
 PLAF_MASTER_HEADERS += vm/os-macosx.hpp vm/mach_signal.hpp
@@ -16,7 +14,7 @@ else
 	LIBS = -lm -framework Cocoa -framework AppKit
 endif
 
-LINKER = $(CXX) $(CFLAGS) -dynamiclib -single_module \
+LINKER = $(CXX) $(TARGET) $(CFLAGS) $(CXXFLAGS) $(SHARED_FLAG) -single_module \
 	-current_version $(VERSION) \
 	-compatibility_version $(VERSION) \
 	-fvisibility=hidden \
diff --git a/vm/Config.macosx.arm64 b/vm/Config.macosx.arm64
new file mode 100644
index 0000000000..95affb370d
--- /dev/null
+++ b/vm/Config.macosx.arm64
@@ -0,0 +1,5 @@
+include vm/Config.macosx
+include vm/Config.arm64.64
+
+PLAF_DLL_OBJS += vm/cpu-arm.64.cpp
+PLAF_MASTER_HEADERS += vm/os-macosx-arm64.64.hpp
diff --git a/vm/Config.macosx.x86.32 b/vm/Config.macosx.x86.32
index 47a9473601..5ed3d2bf85 100644
--- a/vm/Config.macosx.x86.32
+++ b/vm/Config.macosx.x86.32
@@ -1,6 +1,9 @@
 include vm/Config.macosx
 include vm/Config.x86.32
 
+PLAF_MASTER_HEADERS += vm/os-macosx-x86.32.hpp
+
 # The last SDK to support x86 is 10.13
 CFLAGS += --sysroot=$(XCODE_PATH)/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/$(MACOSX_32_SDK)
 CXXFLAGS += --sysroot=$(XCODE_PATH)/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/$(MACOSX_32_SDK)
+SITE_CFLAGS += -mmacosx-version-min=10.9
diff --git a/vm/Config.macosx.x86.64 b/vm/Config.macosx.x86.64
index f987a55694..6726f7f855 100644
--- a/vm/Config.macosx.x86.64
+++ b/vm/Config.macosx.x86.64
@@ -1,2 +1,8 @@
 include vm/Config.macosx
 include vm/Config.x86.64
+
+TARGET = -target x86_64-apple-darwin
+
+PLAF_MASTER_HEADERS += vm/os-macosx-x86.64.hpp
+
+SITE_CFLAGS += -mmacosx-version-min=10.9
diff --git a/vm/Config.macosx.x86.fat b/vm/Config.macosx.x86.fat
index e8a0e10d9f..d899efd4d4 100644
--- a/vm/Config.macosx.x86.fat
+++ b/vm/Config.macosx.x86.fat
@@ -1,3 +1,4 @@
 include vm/Config.macosx
 include vm/Config.x86.fat
 SITE_CFLAGS += -arch i386 -arch x86_64
+SITE_CFLAGS += -mmacosx-version-min=10.9
diff --git a/vm/Config.unix b/vm/Config.unix
index 79fa467826..e667dc61d8 100644
--- a/vm/Config.unix
+++ b/vm/Config.unix
@@ -8,9 +8,9 @@ DLL_EXTENSION = .a
 SHARED_DLL_EXTENSION = .so
 SHARED_FLAG = -shared
 
-PLAF_DLL_OBJS = vm/os-unix.o
+PLAF_DLL_OBJS += vm/os-unix.o
 PLAF_EXE_OBJS += vm/main-unix.o
-PLAF_MASTER_HEADERS += vm/os-unix.hpp
+PLAF_MASTER_HEADERS += vm/os-unix.hpp vm/atomic-gcc.hpp vm/atomic.hpp
 
 FFI_TEST_CFLAGS = -fPIC
 
diff --git a/vm/Config.windows.x86.32 b/vm/Config.windows.x86.32
index 56b7f7fa48..d4c0a88e26 100644
--- a/vm/Config.windows.x86.32
+++ b/vm/Config.windows.x86.32
@@ -1,5 +1,5 @@
 PLAF_DLL_OBJS += vm/os-windows-x86.32.o
-PLAF_MASTER_HEADERS += vm/os-windows.32.hpp
+PLAF_MASTER_HEADERS += vm/os-windows.32.hpp vm/atomic-cl-32.hpp vm/atomic.hpp
 DLL_PATH=http://factorcode.org/dlls
 WINDRES=windres -F pe-i386
 include vm/Config.windows
diff --git a/vm/Config.windows.x86.64 b/vm/Config.windows.x86.64
index 26a1903f94..ade7071c9e 100644
--- a/vm/Config.windows.x86.64
+++ b/vm/Config.windows.x86.64
@@ -1,5 +1,5 @@
 PLAF_DLL_OBJS += vm/os-windows-x86.64.o
-PLAF_MASTER_HEADERS += vm/os-windows.64.hpp
+PLAF_MASTER_HEADERS += vm/os-windows.64.hpp vm/atomic-cl-64.hpp vm/atomic.hpp
 DLL_PATH=http://factorcode.org/dlls/64
 CC=$(WIN64_PATH)-gcc.exe
 WINDRES=$(WIN64_PATH)-windres.exe
diff --git a/vm/cpu-arm.32.S b/vm/cpu-arm.32.S
new file mode 100644
index 0000000000..390cc2bf48
--- /dev/null
+++ b/vm/cpu-arm.32.S
@@ -0,0 +1,127 @@
+#include "asm.h"
+
+/* Note that the XT is passed to the quotation in r12 */
+#define CALL_QUOT \
+        ldr r12,[r0, #9]     /* load quotation-xt slot */ ; \
+       mov lr,pc ; \
+        mov pc,r12
+
+#define JUMP_QUOT \
+        ldr r12,[r0, #9]     /* load quotation-xt slot */ ; \
+       mov pc,r12
+
+#define SAVED_REGS_SIZE 32
+
+#define FRAME (RESERVED_SIZE + SAVED_REGS_SIZE + 8)
+
+#define LR_SAVE [sp, #-4]
+#define RESERVED_SIZE 8
+
+#define SAVE_LR str lr,LR_SAVE
+
+#define LOAD_LR ldr lr,LR_SAVE
+
+#define SAVE_AT(offset) (RESERVED_SIZE + 4 * offset)
+
+#define SAVE(register,offset) str register,[sp, #SAVE_AT(offset)]
+
+#define RESTORE(register,offset) ldr register,[sp, #SAVE_AT(offset)]
+
+#define PROLOGUE \
+       SAVE_LR ; \
+       sub sp,sp,#FRAME
+
+#define EPILOGUE \
+       add sp,sp,#FRAME ; \
+       LOAD_LR
+
+DEF(void,c_to_factor,(CELL quot)):
+        PROLOGUE
+
+       SAVE(r4,0)           /* save GPRs */
+                             /* don't save ds pointer */
+                             /* don't save rs pointer */
+        SAVE(r7,3)
+        SAVE(r8,4)
+        SAVE(r9,5)
+        SAVE(r10,6)
+        SAVE(r11,7)
+       SAVE(r0,8)           /* save quotation since we're about to mangle it */
+
+        sub r0,sp,#4         /* pass call stack pointer as an argument */
+       bl MANGLE(save_callstack_bottom)
+
+       RESTORE(r0,8)        /* restore quotation */
+        CALL_QUOT
+
+        RESTORE(r11,7)       /* restore GPRs */
+        RESTORE(r10,6)
+        RESTORE(r9,5)
+        RESTORE(r8,4)
+        RESTORE(r7,3)
+                             /* don't restore rs pointer */
+                             /* don't restore ds pointer */
+        RESTORE(r4,0)
+
+        EPILOGUE
+        mov pc,lr
+
+/* The JIT compiles an 'mov r1',sp in front of every primitive call, since a
+word which was defined as a primitive will not change its definition for the
+lifetime of the image -- adding new primitives requires a bootstrap. However,
+an undefined word can certainly become defined,
+
+DEFER: foo
+...
+: foo ... ;
+
+And calls to non-primitives do not have this one-instruction prologue, so we
+set the XT of undefined words to this symbol. */
+DEF(void,undefined,(CELL word)):
+       sub r1,sp,#4
+       b MANGLE(undefined_error)
+
+/* Here we have two entry points. The first one is taken when profiling is
+enabled */
+DEF(void,docol_profiling,(CELL word)):
+        ldr r1,[r0, #25]     /* load profile-count slot */
+        add r1,r1,#8         /* increment count */
+        str r1,[r0, #25]     /* store profile-count slot */
+DEF(void,docol,(CELL word)):
+        ldr r0,[r0, #13]     /* load word-def slot */
+        JUMP_QUOT
+
+/* We must pass the XT to the quotation in r12. */
+DEF(void,primitive_call,(void)):
+        ldr r0,[r5], #-4     /* load quotation from data stack */
+        JUMP_QUOT
+
+/* We must preserve r1 here in case we're calling a primitive */
+DEF(void,primitive_execute,(void)):
+        ldr r0,[r5], #-4     /* load word from data stack */
+        ldr pc,[r0, #29]     /* jump to word-xt */
+
+@ DEF(void,set_callstack,(F_STACK_FRAME *to, F_STACK_FRAME *from, CELL length)):
+@         sub sp,r0,r2         /* compute new stack pointer */
+@         mov r0,sp            /* start of destination of memcpy() */
+@        sub sp,sp,#12        /* alignment */
+@         bl MANGLE(memcpy)    /* go */
+@        add sp,sp,#16        /* point SP at innermost frame */
+@         ldr pc,LR_SAVE       /* return */
+
+DEF(void,throw_impl,(CELL quot, F_STACK_FRAME *rewind_to)):
+       add sp,r1,#4         /* compute new stack pointer */
+       ldr lr,LR_SAVE       /* we have rewound the stack; load return address */
+       JUMP_QUOT            /* call the quotation */
+
+DEF(void,lazy_jit_compile,(CELL quot)):
+       mov r1,sp            /* save stack pointer */
+       PROLOGUE
+       bl MANGLE(lazy_jit_compile_impl)
+       EPILOGUE
+        JUMP_QUOT            /* call the quotation */
+
+#ifdef WINCE
+       .section .drectve
+       .ascii " -export:c_to_factor"
+#endif
diff --git a/vm/cpu-arm.32.hpp b/vm/cpu-arm.32.hpp
new file mode 100644
index 0000000000..4748efaf93
--- /dev/null
+++ b/vm/cpu-arm.32.hpp
@@ -0,0 +1,8 @@
+namespace factor {
+
+#define FACTOR_CPU_STRING "arm.32"
+
+register cell ds asm("r5");
+register cell rs asm("r6");
+
+}
diff --git a/vm/cpu-arm.64.cpp b/vm/cpu-arm.64.cpp
new file mode 100644
index 0000000000..6951b3de0e
--- /dev/null
+++ b/vm/cpu-arm.64.cpp
@@ -0,0 +1,25 @@
+#include "master.hpp"
+
+namespace factor {
+
+void factor_vm::dispatch_signal_handler(cell* sp, cell* pc, cell newpc) {
+
+  // bool in_code_seg = code->seg->in_segment_p(*pc);
+  // cell cs_limit = ctx->callstack_seg->start + stack_reserved;
+  // signal_resumable = in_code_seg && *sp >= cs_limit;
+
+  // if (signal_resumable) {
+  //   dispatch_resumable_signal(sp, pc, handler);
+  // } else {
+  //   dispatch_non_resumable_signal(sp, pc, handler, cs_limit);
+  // }
+
+  // Poking with the stack pointer, which the above code does, means
+  // that pointers to stack-allocated objects will become
+  // corrupted. Therefore the root vectors needs to be cleared because
+  // their pointers to stack variables are now garbage.
+  data_roots.clear();
+  code_roots.clear();
+}
+
+}
\ No newline at end of file
diff --git a/vm/cpu-arm.64.hpp b/vm/cpu-arm.64.hpp
new file mode 100644
index 0000000000..b59abac052
--- /dev/null
+++ b/vm/cpu-arm.64.hpp
@@ -0,0 +1,13 @@
+namespace factor {
+
+#define FACTOR_CPU_STRING "arm.64"
+
+// register cell ds asm("r5");
+// register cell rs asm("r6");
+inline static void flush_icache(cell start, cell len) {
+  // int result;
+}
+
+
+
+}
diff --git a/vm/cpu-arm.hpp b/vm/cpu-arm.hpp
index 4cd5316099..a0d4de56ec 100644
--- a/vm/cpu-arm.hpp
+++ b/vm/cpu-arm.hpp
@@ -1,5 +1,70 @@
 namespace factor {
 
-#define FACTOR_CPU_STRING "arm"
+#define CALLSTACK_BOTTOM(ctx) (ctx->callstack_seg->end - sizeof(cell) * 5) // omg
+
+// void c_to_factor(cell quot);
+// void lazy_jit_compile(cell quot);
+
+static const fixnum xt_tail_pic_offset = 4 + 1; // or 4 or whatever else...
+
+
+static const unsigned char call_opcode = 0xe8; // omg, these are x86 values
+static const unsigned char jmp_opcode = 0xe9; // omg
+
+static const unsigned SIGNAL_HANDLER_STACK_FRAME_SIZE = 64; // omg
+
+
+// omg
+inline static unsigned char call_site_opcode(cell return_address) {
+  return *(unsigned char*)(return_address - 5);
+}
+
+// omg
+inline static void check_call_site(cell return_address) {
+  unsigned char opcode = call_site_opcode(return_address);
+  FACTOR_ASSERT(opcode == call_opcode || opcode == jmp_opcode);
+  (void)opcode; // suppress warning when compiling without assertions
+}
+
+// omg
+inline static void* get_call_target(cell return_address) {
+  check_call_site(return_address);
+  return (void*)(*(int*)(return_address - 4) + return_address);
+}
+
+// omg
+inline static void set_call_target(cell return_address, cell target) {
+  check_call_site(return_address);
+  *(int*)(return_address - 4) = (uint32_t)(target - return_address);
+}
+
+inline static bool tail_call_site_p(cell return_address) {
+  switch (call_site_opcode(return_address)) {
+    case jmp_opcode:
+      return true;
+    case call_opcode:
+      return false;
+    default:
+      abort();
+      return false;
+  }
+}
+
+// inline static unsigned int fpu_status(unsigned int status) {
+//   unsigned int r = 0;
+
+//   if (status & 0x01)
+//     r |= FP_TRAP_INVALID_OPERATION;
+//   if (status & 0x04)
+//     r |= FP_TRAP_ZERO_DIVIDE;
+//   if (status & 0x08)
+//     r |= FP_TRAP_OVERFLOW;
+//   if (status & 0x10)
+//     r |= FP_TRAP_UNDERFLOW;
+//   if (status & 0x20)
+//     r |= FP_TRAP_INEXACT;
+
+//   return r;
+// }
 
 }
-- 
2.34.1