mmx: convert from sse

diff --git a/orc/orcprogram-mmx.c b/orc/orcprogram-mmx.c
index 68e7975..258e0ba 100644 (file)
--- a/orc/orcprogram-mmx.c
+++ b/orc/orcprogram-mmx.c
@@ -694,7 +694,7 @@ orc_emit_split_3_regions (OrcCompiler *compiler)
   orc_x86_emit_jmp (compiler, 7);
 
   /* else, iterations are all unaligned: n1=n, n2=0, n3=0 */
-  orc_x86_emit_label (compiler, 6, 0);
+  orc_x86_emit_label (compiler, 6);
 
   orc_x86_emit_mov_memoffset_reg (compiler, 4,
       (int)ORC_STRUCT_OFFSET(OrcExecutor,n), compiler->exec_reg, X86_EAX);
@@ -706,7 +706,7 @@ orc_emit_split_3_regions (OrcCompiler *compiler)
   orc_x86_emit_mov_reg_memoffset (compiler, 4, X86_EAX,
       (int)ORC_STRUCT_OFFSET(OrcExecutor,counter3), compiler->exec_reg);
 
-  orc_x86_emit_label (compiler, 7, 0);
+  orc_x86_emit_label (compiler, 7);
 }
 
 static void
@@ -787,6 +787,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
     compiler->asm_code = NULL;
     compiler->asm_code_len = 0;
     memset (compiler->labels, 0, sizeof (compiler->labels));
+    memset (compiler->labels_int, 0, sizeof (compiler->labels_int));
     compiler->n_fixups = 0;
     compiler->n_output_insns = 0;
   }
@@ -822,7 +823,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
           compiler->exec_reg);
     }
 
-    orc_x86_emit_label (compiler, LABEL_OUTER_LOOP, 0);
+    orc_x86_emit_label (compiler, LABEL_OUTER_LOOP);
   }
 
   if (compiler->program->constant_n > 0 &&
@@ -901,14 +902,14 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
             (int)ORC_STRUCT_OFFSET(OrcExecutor,counter1), compiler->exec_reg);
         orc_x86_emit_je (compiler, LABEL_STEP_UP(compiler->loop_shift));
         orc_mmx_emit_loop (compiler, 0, 1<<compiler->loop_shift);
-        orc_x86_emit_label (compiler, LABEL_STEP_UP(compiler->loop_shift), 0);
+        orc_x86_emit_label (compiler, LABEL_STEP_UP(compiler->loop_shift));
       }
 
       compiler->loop_shift = save_loop_shift;
       compiler->vars[align_var].is_aligned = TRUE;
     }
 
-    orc_x86_emit_label (compiler, LABEL_REGION1_SKIP, 0);
+    orc_x86_emit_label (compiler, LABEL_REGION1_SKIP);
 
     orc_x86_emit_cmp_imm_memoffset (compiler, 4, 0,
         (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2), compiler->exec_reg);
@@ -921,7 +922,8 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
     }
 
     ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
-    orc_x86_emit_label (compiler, LABEL_INNER_LOOP_START, 0);
+    orc_x86_emit_align (compiler, 4);
+    orc_x86_emit_label (compiler, LABEL_INNER_LOOP_START);
     ui_max = 1<<compiler->unroll_shift;
     for(ui=0;ui<ui_max;ui++) {
       compiler->offset = ui<<compiler->loop_shift;
@@ -937,7 +939,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
           compiler->exec_reg);
     }
     orc_x86_emit_jne (compiler, LABEL_INNER_LOOP_START);
-    orc_x86_emit_label (compiler, LABEL_REGION2_SKIP, 0);
+    orc_x86_emit_label (compiler, LABEL_REGION2_SKIP);
 
     if (emit_region3) {
       int save_loop_shift;
@@ -954,7 +956,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
             (int)ORC_STRUCT_OFFSET(OrcExecutor,counter3), compiler->exec_reg);
         orc_x86_emit_je (compiler, LABEL_STEP_DOWN(compiler->loop_shift));
         orc_mmx_emit_loop (compiler, 0, 1<<compiler->loop_shift);
-        orc_x86_emit_label (compiler, LABEL_STEP_DOWN(compiler->loop_shift), 0);
+        orc_x86_emit_label (compiler, LABEL_STEP_DOWN(compiler->loop_shift));
       }
 
       compiler->loop_shift = save_loop_shift;
@@ -968,7 +970,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
         (int)ORC_STRUCT_OFFSET(OrcExecutor,params[ORC_VAR_A2]),
         compiler->exec_reg);
     orc_x86_emit_jne (compiler, LABEL_OUTER_LOOP);
-    orc_x86_emit_label (compiler, LABEL_OUTER_LOOP_SKIP, 0);
+    orc_x86_emit_label (compiler, LABEL_OUTER_LOOP_SKIP);
   }
 
   mmx_save_accumulators (compiler);
@@ -982,6 +984,7 @@ orc_compiler_mmx_assemble (OrcCompiler *compiler)
 #endif
   orc_x86_emit_epilogue (compiler);
 
+  orc_x86_calculate_offsets (compiler);
   orc_x86_output_insns (compiler);
 
   orc_x86_do_fixups (compiler);

diff --git a/orc/orcrules-mmx.c b/orc/orcrules-mmx.c
index a98189f..923405b 100644 (file)
--- a/orc/orcrules-mmx.c
+++ b/orc/orcrules-mmx.c
@@ -1582,8 +1582,8 @@ mmx_rule_mulhsl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   for(i=0;i<(1<<p->insn_shift);i++) {
     orc_x86_emit_mov_memoffset_reg (p, 4, offset + 4*i, p->exec_reg, X86_EAX);
-    orc_x86_emit_cpuinsn_load_memoffset (p, ORC_X86_imul_rm, 4, 0,
-        offset + 16 + 4*i, p->exec_reg, -1);
+    orc_x86_emit_cpuinsn_memoffset (p, ORC_X86_imul_rm, 4,
+        offset + 16 + 4*i, p->exec_reg);
     orc_x86_emit_mov_reg_memoffset (p, 4, X86_EDX, offset + 4*i, p->exec_reg);
   }
 
@@ -1628,6 +1628,38 @@ mmx_rule_mulslq (OrcCompiler *p, void *user, OrcInstruction *insn)
 
 #ifndef MMX
 static void
+mmx_rule_mulslq_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int i;
+  int regsize = p->is_64bit ? 8 : 4;
+  int offset = ORC_STRUCT_OFFSET(OrcExecutor,arrays[ORC_VAR_T1]);
+
+  orc_x86_emit_mov_mmx_memoffset (p, 8, p->vars[insn->src_args[0]].alloc,
+      offset, p->exec_reg, FALSE, FALSE);
+  orc_x86_emit_mov_mmx_memoffset (p, 8, p->vars[insn->src_args[1]].alloc,
+      offset + 8, p->exec_reg, FALSE, FALSE);
+  orc_x86_emit_mov_reg_memoffset (p, regsize, X86_EAX, offset + 32,
+      p->exec_reg);
+  orc_x86_emit_mov_reg_memoffset (p, regsize, X86_EDX, offset + 40,
+      p->exec_reg);
+
+  for(i=0;i<(1<<p->insn_shift);i++) {
+    orc_x86_emit_mov_memoffset_reg (p, 4, offset + 4*i, p->exec_reg, X86_EAX);
+    orc_x86_emit_cpuinsn_memoffset (p, ORC_X86_imul_rm, 4,
+        offset + 8 + 4*i, p->exec_reg);
+    orc_x86_emit_mov_reg_memoffset (p, 4, X86_EAX, offset + 16 + 8*i, p->exec_reg);
+    orc_x86_emit_mov_reg_memoffset (p, 4, X86_EDX, offset + 16 + 8*i + 4, p->exec_reg);
+  }
+
+  orc_x86_emit_mov_memoffset_mmx (p, 16, offset + 16, p->exec_reg,
+      p->vars[insn->dest_args[0]].alloc, FALSE);
+  orc_x86_emit_mov_memoffset_reg (p, 8, offset + 32, p->exec_reg, X86_EAX);
+  orc_x86_emit_mov_memoffset_reg (p, 8, offset + 40, p->exec_reg, X86_EDX);
+}
+#endif
+
+#ifndef MMX
+static void
 mmx_rule_mululq (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   int src = p->vars[insn->src_args[1]].alloc;
@@ -2889,6 +2921,7 @@ orc_compiler_mmx_register_rules (OrcTarget *target)
 #ifndef MMX
   orc_rule_register (rule_set, "mulhsl", mmx_rule_mulhsl_slow, NULL);
   orc_rule_register (rule_set, "mulhul", mmx_rule_mulhul, NULL);
+  orc_rule_register (rule_set, "mulslq", mmx_rule_mulslq_slow, NULL);
 #endif
   orc_rule_register (rule_set, "mullb", mmx_rule_mullb, NULL);
   orc_rule_register (rule_set, "mulhsb", mmx_rule_mulhsb, NULL);