neon: Fix accumulator opcodes

diff --git a/orc/orcprogram-neon.c b/orc/orcprogram-neon.c
index a6a7fed..7dfb726 100644 (file)
--- a/orc/orcprogram-neon.c
+++ b/orc/orcprogram-neon.c
@@ -217,7 +217,7 @@ orc_compiler_neon_init (OrcCompiler *compiler)
 }
 
 void
-orc_neon_load_constants (OrcCompiler *compiler)
+orc_neon_load_constants_outer (OrcCompiler *compiler)
 {
   int i;
   for(i=0;i<ORC_N_VARIABLES;i++){
@@ -251,12 +251,38 @@ orc_neon_load_constants (OrcCompiler *compiler)
         break;
       case ORC_VAR_TYPE_SRC:
       case ORC_VAR_TYPE_DEST:
+        break;
+      case ORC_VAR_TYPE_ACCUMULATOR:
+        orc_neon_emit_loadil (compiler, compiler->vars[i].alloc, 0);
+        break;
+      case ORC_VAR_TYPE_TEMP:
+        break;
+      default:
+        ORC_PROGRAM_ERROR(compiler,"bad vartype");
+        break;
+    }
+  }
+}
+
+void
+orc_neon_load_constants_inner (OrcCompiler *compiler)
+{
+  int i;
+  for(i=0;i<ORC_N_VARIABLES;i++){
+    if (compiler->vars[i].name == NULL) continue;
+
+    switch (compiler->vars[i].vartype) {
+      case ORC_VAR_TYPE_CONST:
+        break;
+      case ORC_VAR_TYPE_PARAM:
+        break;
+      case ORC_VAR_TYPE_SRC:
+      case ORC_VAR_TYPE_DEST:
         orc_arm_emit_load_reg (compiler, 
             compiler->vars[i].ptr_register,
             compiler->exec_reg, ORC_STRUCT_OFFSET(OrcExecutor, arrays[i]));
         break;
       case ORC_VAR_TYPE_ACCUMULATOR:
-        orc_neon_emit_loadil (compiler, compiler->vars[i].alloc, 0);
         break;
       case ORC_VAR_TYPE_TEMP:
         break;
@@ -540,6 +566,8 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
 
   orc_neon_emit_prologue (compiler);
 
+  orc_neon_load_constants_outer (compiler);
+
   if (compiler->program->is_2d) {
     if (compiler->program->constant_m > 0) {
       orc_arm_emit_load_imm (compiler, ORC_ARM_A3, compiler->program->constant_m);
@@ -601,7 +629,7 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
     orc_arm_emit_label (compiler, 7);
   }
 
-  orc_neon_load_constants (compiler);
+  orc_neon_load_constants_inner (compiler);
 
   if (compiler->loop_shift > 0) {
     int save_loop_shift = compiler->loop_shift;
@@ -702,6 +730,7 @@ orc_neon_emit_loop (OrcCompiler *compiler)
   OrcStaticOpcode *opcode;
   OrcRule *rule;
 
+  orc_compiler_append_code(compiler,"# LOOP shift %d\n", compiler->loop_shift);
   for(j=0;j<compiler->n_insns;j++){
     compiler->insn_index = j;
     insn = compiler->insns + j;

diff --git a/orc/orcrules-neon.c b/orc/orcrules-neon.c
index 0454413..c0e24d9 100644 (file)
--- a/orc/orcrules-neon.c
+++ b/orc/orcrules-neon.c
@@ -1157,10 +1157,27 @@ UNARY(convlf,"vcvt.f32.s32",0xf3bb0600, 1)
 static void
 orc_neon_rule_accw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
-  orc_neon_emit_binary (p, "vadd.i16", 0xf2100800,
-      p->vars[insn->dest_args[0]].alloc,
-      p->vars[insn->dest_args[0]].alloc,
-      p->vars[insn->src_args[0]].alloc);
+  unsigned int code;
+
+  if (p->loop_shift < 2) {
+    ORC_ASM_CODE(p,"  vshl.i64 %s, %s, #%d\n",
+        orc_neon_reg_name (p->tmpreg),
+        orc_neon_reg_name (p->vars[insn->src_args[0]].alloc), 48);
+    code = NEON_BINARY(0xf2a00590, p->tmpreg, 0,
+        p->vars[insn->src_args[0]].alloc);
+    code |= (48) << 16;
+    orc_arm_emit (p, code);
+
+    orc_neon_emit_binary (p, "vadd.i16", 0xf2100800,
+        p->vars[insn->dest_args[0]].alloc,
+        p->vars[insn->dest_args[0]].alloc,
+        p->tmpreg);
+  } else {
+    orc_neon_emit_binary (p, "vadd.i16", 0xf2100800,
+        p->vars[insn->dest_args[0]].alloc,
+        p->vars[insn->dest_args[0]].alloc,
+        p->vars[insn->src_args[0]].alloc);
+  }
 }
 
 static void
@@ -1255,23 +1272,50 @@ static void
 orc_neon_rule_accsadubl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   orc_uint32 x;
-  
-  x = 0xf3800700;
-  ORC_ASM_CODE(p,"  vabdl.u8 %s, %s, %s\n",
-      orc_neon_reg_name_quad (p->tmpreg),
-      orc_neon_reg_name (p->vars[insn->src_args[0]].alloc),
-      orc_neon_reg_name (p->vars[insn->src_args[1]].alloc));
-  x |= (p->tmpreg&0xf)<<12;
-  x |= ((p->tmpreg>>4)&0x1)<<22;
-  x |= (p->vars[insn->src_args[0]].alloc&0xf)<<16;
-  x |= ((p->vars[insn->src_args[0]].alloc>>4)&0x1)<<7;
-  x |= (p->vars[insn->src_args[1]].alloc&0xf)<<0;
-  x |= ((p->vars[insn->src_args[1]].alloc>>4)&0x1)<<5;
-  orc_arm_emit (p, x);
-
-  orc_neon_emit_unary (p, "vpadal.u16", 0xf3b40680,
-      p->vars[insn->dest_args[0]].alloc,
-      p->tmpreg);
+  unsigned int code;
+
+  if (p->loop_shift < 2) {
+    x = 0xf3800700;
+    ORC_ASM_CODE(p,"  vabdl.u8 %s, %s, %s\n",
+        orc_neon_reg_name_quad (p->tmpreg),
+        orc_neon_reg_name (p->vars[insn->src_args[0]].alloc),
+        orc_neon_reg_name (p->vars[insn->src_args[1]].alloc));
+    x |= (p->tmpreg&0xf)<<12;
+    x |= ((p->tmpreg>>4)&0x1)<<22;
+    x |= (p->vars[insn->src_args[0]].alloc&0xf)<<16;
+    x |= ((p->vars[insn->src_args[0]].alloc>>4)&0x1)<<7;
+    x |= (p->vars[insn->src_args[1]].alloc&0xf)<<0;
+    x |= ((p->vars[insn->src_args[1]].alloc>>4)&0x1)<<5;
+    orc_arm_emit (p, x);
+
+    ORC_ASM_CODE(p,"  vshl.i64 %s, %s, #%d\n",
+        orc_neon_reg_name (p->tmpreg),
+        orc_neon_reg_name (p->tmpreg), 64 - (16<<p->loop_shift));
+    code = NEON_BINARY(0xf2a00590, p->tmpreg, 0, p->tmpreg);
+    code |= (64 - (16<<p->loop_shift)) << 16;
+    orc_arm_emit (p, code);
+
+    orc_neon_emit_unary (p, "vpadal.u16", 0xf3b40680,
+        p->vars[insn->dest_args[0]].alloc,
+        p->tmpreg);
+  } else {
+    x = 0xf3800700;
+    ORC_ASM_CODE(p,"  vabdl.u8 %s, %s, %s\n",
+        orc_neon_reg_name_quad (p->tmpreg),
+        orc_neon_reg_name (p->vars[insn->src_args[0]].alloc),
+        orc_neon_reg_name (p->vars[insn->src_args[1]].alloc));
+    x |= (p->tmpreg&0xf)<<12;
+    x |= ((p->tmpreg>>4)&0x1)<<22;
+    x |= (p->vars[insn->src_args[0]].alloc&0xf)<<16;
+    x |= ((p->vars[insn->src_args[0]].alloc>>4)&0x1)<<7;
+    x |= (p->vars[insn->src_args[1]].alloc&0xf)<<0;
+    x |= ((p->vars[insn->src_args[1]].alloc>>4)&0x1)<<5;
+    orc_arm_emit (p, x);
+
+    orc_neon_emit_unary (p, "vpadal.u16", 0xf3b40680,
+        p->vars[insn->dest_args[0]].alloc,
+        p->tmpreg);
+  }
 }
 
 static void