Removed native_pop_code, ported call with label

iseq.c

@@ -43,9 +43,6 @@
 #include "insns.inc"
 #include "insns_info.inc"
 #include <sys/mman.h>
-#include "ujit_examples.h"
-
-uint8_t *native_pop_code; // TODO: hack. see addr2insn
 
 VALUE rb_cISeq;
 static VALUE iseqw_new(const rb_iseq_t *iseq);
@@ -3217,52 +3214,6 @@ rb_vm_encoded_insn_data_table_init(void)
         st_add_direct(rb_encoded_insn_data, key1, (st_data_t)&insn_data[insn]);
         st_add_direct(rb_encoded_insn_data, key2, (st_data_t)&insn_data[insn]);
     }
-
-
-    /*
-    native_pop_code = mmap(0, 4096, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANON|MAP_PRIVATE, 0, 0);
-    if (native_pop_code == MAP_FAILED) rb_bug("mmap failed");
-    uint8_t *head = native_pop_code;
-    memcpy(head, ujit_pre_call_bytes, sizeof(ujit_pre_call_bytes));
-    head += sizeof(ujit_pre_call_bytes);
-    const uint8_t handmade_pop[] = {           // TODO assmeble this from a separate file
-       0x48, 0x83, 0x6f, 0x08, 0x08,         //	subq	$8, 8(%rdi)
-       0x48, 0x83, 0xc6, 0x08,               //	addq	$8, %rsi
-       0x48, 0x89, 0x37,                     //	movq	%rsi, (%rdi)
-       0x48, 0x89, 0xf0                      //	movq	%rsi, %rax
-    };
-    memcpy(head, handmade_pop, sizeof(handmade_pop));
-    head += sizeof(handmade_pop);
-    memcpy(head, ujit_post_call_bytes, sizeof(ujit_post_call_bytes));
-    // TODO this is small enough to fit in the page we allocated but that can change
-    */
-
-
-
-    // I decided to start by replicating Alan's code above using the new assembler
-    codeblock_t block;
-    codeblock_t* cb = &block;
-    cb_init(cb, 4096);
-
-    // 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);
-
-    native_pop_code = cb_get_ptr(cb, 0);
-
-
-
-
-
-
-
 }
 
 int
@@ -3276,12 +3227,6 @@ rb_vm_insn_addr2insn(const void *addr)
         return (int)e->insn;
     }
 
-    // TODO this is a hack. The proper way to do this is to refactor this so that it takes
-    // the iseq body.
-    if (addr && addr == native_pop_code) {
-        return BIN(pop);
-    }
-
     rb_bug("rb_vm_insn_addr2insn: invalid insn address: %p", addr);
 }
 
@@ -3312,12 +3257,6 @@ encoded_iseq_trace_instrument(VALUE *iseq_encoded_insn, rb_event_flag_t turnon,
         return e->insn_len;
     }
 
-    // TODO this is a hack. The proper way to do this is to refactor this so that it takes
-    // the iseq body.
-    if (key && (uint8_t *)key == native_pop_code) {
-        return insn_len(BIN(pop));
-    }
-
     rb_bug("trace_instrument: invalid insn address: %p", (void *)*iseq_encoded_insn);
 }
 

ujit_asm.c

@@ -723,22 +723,20 @@ void add(codeblock_t* cb, x86opnd_t opnd0, x86opnd_t opnd1)
     );
 }
 
-/*
 /// call - Call to label with 32-bit offset
-void call(CodeBlock cb, Label label)
+void call_label(codeblock_t* cb, size_t label_idx)
 {
-    cb.writeASM("call", label);
+    //cb.writeASM("call", label);
 
     // Write the opcode
-    cb.writeByte(0xE8);
+    cb_write_byte(cb, 0xE8);
 
     // Add a reference to the label
-    cb.addLabelRef(label);
+    cb_label_ref(cb, label_idx);
 
     // Relative 32-bit offset to be patched
-    cb.writeInt(0, 32);
+    cb_write_int(cb, 0, 32);
 }
-*/
 
 /// call - Indirect call with an R/M operand
 void call(codeblock_t* cb, x86opnd_t opnd)

ujit_asm.h

@@ -194,6 +194,7 @@ void cb_write_epilogue(codeblock_t* cb);
 
 // Encode individual instructions into a code block
 void add(codeblock_t* cb, x86opnd_t opnd0, x86opnd_t opnd1);
+void call_label(codeblock_t* cb, size_t label_idx);
 void call(codeblock_t* cb, x86opnd_t opnd);
 void ja(codeblock_t* cb, size_t label_idx);
 void jae(codeblock_t* cb, size_t label_idx);

ujit_asm_tests.c

@@ -83,12 +83,13 @@ void run_tests()
     cb_set_pos(cb, 0); add(cb, ECX, imm_opnd(255)); check_bytes(cb, "81C1FF000000");
 
     // call
-    /*
+    {
-    test(
+        cb_set_pos(cb, 0);
-        delegate void (CodeBlock cb) { auto l = cb.label("foo"); cb.instr(CALL, l); },
+        size_t fn_label = cb_new_label(cb, "foo");
-        "E8FBFFFFFF"
+        call_label(cb, fn_label);
-    );
+        cb_link_labels(cb);
-    */
+        check_bytes(cb, "E8FBFFFFFF");
+    }
     cb_set_pos(cb, 0); call(cb, RAX); check_bytes(cb, "FFD0");
     cb_set_pos(cb, 0); call(cb, mem_opnd(64, RSP, 8)); check_bytes(cb, "FF542408");
 

ujit_compile.c

@@ -8,17 +8,12 @@
 #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
 extern st_table *rb_encoded_insn_data;
 
-// See coment for rb_encoded_insn_data in iseq.c
+// See comment for rb_encoded_insn_data in iseq.c
 static void
 addr2insn_bookkeeping(void *code_ptr, int insn)
 {
@@ -29,7 +24,7 @@ addr2insn_bookkeeping(void *code_ptr, int insn)
         st_insert(rb_encoded_insn_data, (st_data_t)code_ptr, encoded_insn_data);
     }
     else {
-        rb_bug("ujit: failed to find info for original instruction while dealing wiht addr2insn");
+        rb_bug("ujit: failed to find info for original instruction while dealing with addr2insn");
     }
 }
 
@@ -51,6 +46,9 @@ ujit_compile_insn(rb_iseq_t *iseq, size_t insn_idx)
     //const char* name = insn_name(insn);
     //printf("%s\n", name);
 
+    // TODO: encode individual instructions, eg
+    // putnil, putobject, pop, dup, getlocal, nilp
+
     if (insn == BIN(pop))
     {
         // Get a pointer to the current write position in the code block