implemented JIT_OP_MEMCPY for x86

diff --git a/ChangeLog b/ChangeLog
index 3f9f4f99d275f4fb45d4b3238a479e645942ee9c..f4c4d94f5f7c6d90026e7c50a2ef4c4c95eddab4 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,7 +1,7 @@
 2005-12-22  Aleksey Demakov  <ademakov@gmail.com>
 
-       * jit/jit-rules-x86.sel: implement JIT_OP_MEMSET rule optimized
-       for small constant size blocks.
+       * jit/jit-rules-x86.sel: implement JIT_OP_MEMSET and JIT_OP_MEMCPY
+       rules optimized for small constant size blocks.
 
 2005-12-20  Aleksey Demakov  <ademakov@gmail.com>
 

diff --git a/jit/jit-rules-x86.sel b/jit/jit-rules-x86.sel
index 96ac5327361c608b3f42c0cec2a85e891373ddaa..96289314b563e0a2ed2e380cbbe2ba25bc5dc5fa 100644 (file)
--- a/jit/jit-rules-x86.sel
+++ b/jit/jit-rules-x86.sel
@@ -3031,8 +3031,117 @@ JIT_OP_STORE_ELEMENT_NFLOAT: manual
 
 JIT_OP_MEMCPY: manual
        [] -> {
-               /* TODO */
-               TODO();
+               unsigned char *inst;
+               int reg, reg2, reg3;
+               int regi, save_reg3;
+               int disp;
+
+               if(insn->value2->is_constant && insn->value2->address <= 0)
+               {
+               }
+               else if(insn->value2->is_constant && insn->value2->address <= 32)
+               {
+                       reg = _jit_regs_load_value
+                               (gen, insn->dest, 0,
+                                (insn->flags & (JIT_INSN_DEST_NEXT_USE | JIT_INSN_DEST_LIVE)));
+                       reg2 = _jit_regs_load_value
+                               (gen, insn->value1, 0,
+                                (insn->flags & (JIT_INSN_VALUE1_NEXT_USE | JIT_INSN_VALUE1_LIVE)));
+
+                       reg3 = -1;
+                       save_reg3 = 0;
+                       for(regi = 0; regi < 4; regi++)
+                       {
+                               if(regi != reg && regi != reg2)
+                               {
+                                       if(gen->contents[regi].num_values == 0 &&
+                                          gen->contents[regi].used_for_temp == 0)
+                                       {
+                                               reg3 = regi;
+                                               break;
+                                       }
+                                       if(reg3 == -1)
+                                       {
+                                               reg3 = regi;
+                                       }
+                               }
+                       }
+                       if(gen->contents[reg3].num_values > 0 ||
+                          gen->contents[reg3].used_for_temp)
+                       {
+                               save_reg3 = 1;
+                       }
+
+                       inst = gen->posn.ptr;
+                       if(!jit_cache_check_for_n(&(gen->posn), 256))
+                       {
+                               jit_cache_mark_full(&(gen->posn));
+                               return;
+                       }
+
+                       reg = _jit_reg_info[reg].cpu_reg;
+                       reg2 = _jit_reg_info[reg2].cpu_reg;
+                       reg3 = _jit_reg_info[reg3].cpu_reg;
+
+                               if(save_reg3)
+                       {
+                               x86_push_reg(inst, reg3);
+                       }
+
+                       disp = 0;
+                       while(insn->value2->address >= (disp + 4))
+                       {
+                               x86_mov_reg_membase(inst, reg3, reg2, disp, 4);
+                               x86_mov_membase_reg(inst, reg, disp, reg3, 4);
+                               disp += 4;
+                       }
+                       if(insn->value2->address >= (disp + 2))
+                       {
+                               x86_mov_reg_membase(inst, reg3, reg2, disp, 2);
+                               x86_mov_membase_reg(inst, reg, disp, reg3, 2);
+                               disp += 2;
+                       }
+                       if(insn->value2->address > disp)
+                       {
+                               x86_mov_reg_membase(inst, reg3, reg2, disp, 1);
+                               x86_mov_membase_reg(inst, reg, disp, reg3, 1);
+                       }
+
+                       if(save_reg3)
+                       {
+                               x86_pop_reg(inst, reg3);
+                       }
+
+                       gen->posn.ptr = inst;
+               }
+               else
+               {
+                       reg = _jit_regs_load_value
+                               (gen, insn->dest, 0,
+                                (insn->flags & (JIT_INSN_DEST_NEXT_USE | JIT_INSN_DEST_LIVE)));
+                       reg2 = _jit_regs_load_value
+                               (gen, insn->value1, 0,
+                                (insn->flags & (JIT_INSN_VALUE1_NEXT_USE | JIT_INSN_VALUE1_LIVE)));
+                       reg3 = _jit_regs_load_value
+                               (gen, insn->value2, 0,
+                                (insn->flags & (JIT_INSN_VALUE2_NEXT_USE | JIT_INSN_VALUE2_LIVE)));
+                       _jit_regs_spill_all(gen);
+
+                       inst = gen->posn.ptr;
+                       if(!jit_cache_check_for_n(&(gen->posn), 32))
+                       {
+                               jit_cache_mark_full(&(gen->posn));
+                               return;
+                       }
+
+                       x86_push_reg(inst, _jit_reg_info[reg3].cpu_reg);
+                       x86_push_reg(inst, _jit_reg_info[reg2].cpu_reg);
+                       x86_push_reg(inst, _jit_reg_info[reg].cpu_reg);
+                       x86_call_code(inst, jit_memcpy);
+                       x86_alu_reg_imm(inst, X86_ADD, X86_ESP, 3 * sizeof(void *));
+
+                       gen->posn.ptr = inst;
+               }
        }
 
 JIT_OP_MEMMOVE: manual