Add new files, ujit_compile.c, ujit_compile.h

common.mk

@@ -151,6 +151,7 @@ COMMONOBJS    = array.$(OBJEXT) \
 		vm_sync.$(OBJEXT) \
 		vm_trace.$(OBJEXT) \
 		ujit_asm.$(OBJEXT) \
+		ujit_compile.$(OBJEXT) \
 		$(COROUTINE_OBJ) \
 		$(DTRACE_OBJ) \
 		$(BUILTIN_ENCOBJS) \

compile.c

@@ -43,6 +43,7 @@
 #include "builtin.h"
 #include "insns.inc"
 #include "insns_info.inc"
+#include "ujit_compile.h"
 
 #undef RUBY_UNTYPED_DATA_WARNING
 #define RUBY_UNTYPED_DATA_WARNING 0
@@ -856,8 +857,6 @@ rb_iseq_compile_node(rb_iseq_t *iseq, const NODE *node)
     return iseq_setup(iseq, ret);
 }
 
-extern uint8_t *native_pop_code; // TODO global hack
-
 static int
 rb_iseq_translate_threaded_code(rb_iseq_t *iseq)
 {
@@ -870,17 +869,17 @@ rb_iseq_translate_threaded_code(rb_iseq_t *iseq)
     {
     	int insn = (int)iseq->body->iseq_encoded[i];
     	int len = insn_len(insn);
-    	encoded[i] = (VALUE)table[insn];
 
-        if (insn == BIN(pop))
-            encoded[i] = (VALUE)native_pop_code;
-
-        const char* name = insn_name(insn);
-        printf("%s\n", name);
+        uint8_t* native_code_ptr = ujit_compile_insn(iseq, i);
 
+        if (native_code_ptr)
+            encoded[i] = (VALUE)native_code_ptr;
+        else
+	        encoded[i] = (VALUE)table[insn];
 
     	i += len;
     }
+
     FL_SET((VALUE)iseq, ISEQ_TRANSLATED);
 #endif
     return COMPILE_OK;

ujit_compile.c

@@ -0,0 +1,63 @@
+#include <assert.h>
+#include "insns.inc"
+#include "internal.h"
+#include "vm_core.h"
+#include "vm_callinfo.h"
+#include "builtin.h"
+#include "insns_info.inc"
+#include "ujit_compile.h"
+#include "ujit_asm.h"
+
+// NOTE: do we have to deal with multiple Ruby processes/threads compiling
+// functions with the new Ractor in Ruby 3.0? If so, we need to think about
+// a strategy for handling that. What does Ruby currently do for its own
+// iseq translation?
+static codeblock_t block;
+static codeblock_t* cb = NULL;
+
+extern uint8_t* native_pop_code; // FIXME global hack
+
+// Generate a chunk of machinecode for one individual bytecode instruction
+// Eventually, this will handle multiple instructions in a sequence
+uint8_t* ujit_compile_insn(rb_iseq_t *iseq, size_t insn_idx)
+{
+    // Allocate the code block if not previously allocated
+    if (!cb)
+    {
+        // 4MB ought to be enough for anybody
+        cb = &block;
+        cb_init(cb, 4000000);
+    }
+
+	int insn = (int)iseq->body->iseq_encoded[insn_idx];
+
+    //const char* name = insn_name(insn);
+    //printf("%s\n", name);
+
+    if (insn == BIN(pop))
+    {
+        //printf("COMPILING %ld\n", cb->write_pos);
+
+        return native_pop_code;
+
+        /*
+        // Get a pointer to the current write position in the code block
+        uint8_t* code_ptr = &cb->mem_block[cb->write_pos];
+
+        // Write the pre call bytes
+        cb_write_prologue(cb);
+
+        sub(cb, mem_opnd(64, RDI, 8), imm_opnd(8)); // decrement SP
+        add(cb, RSI, imm_opnd(8));                  // increment PC
+        mov(cb, mem_opnd(64, RDI, 0), RSI);         // write new PC to EC object, not necessary for pop bytecode?
+        mov(cb, RAX, RSI);                          // return new PC
+
+        // Write the post call bytes
+        cb_write_epilogue(cb);
+
+        return code_ptr;
+        */
+    }
+
+    return 0;
+}

ujit_compile.h

@@ -0,0 +1,14 @@
+#ifndef UJIT_COMPILE_H
+#define UJIT_COMPILE_H 1
+
+#include "stddef.h"
+#include "stdint.h"
+
+#ifndef rb_iseq_t
+typedef struct rb_iseq_struct rb_iseq_t;
+#define rb_iseq_t rb_iseq_t
+#endif
+
+uint8_t* ujit_compile_insn(rb_iseq_t* iseq, size_t insn_idx);
+
+#endif