From: David Schleef <ds@schleef.org>
Date: Sun, 1 Aug 2010 22:41:36 +0000 (-0700)
Subject: Better handling of invariants
X-Git-Tag: orc-0.4.7~98
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=28134046e7484ca3432fd7f824a8bf0dd3c246e6;p=platform%2Fupstream%2Forc.git

Better handling of invariants
---

diff --git a/orc/orccompiler.c b/orc/orccompiler.c
index 90dadbd..c5a6855 100644
--- a/orc/orccompiler.c
+++ b/orc/orccompiler.c
@@ -212,10 +212,11 @@ orc_program_compile_full (OrcProgram *program, OrcTarget *target,
     ORC_LOG("variables");
     for(i=0;i<ORC_N_VARIABLES;i++){
       if (program->vars[i].size > 0) {
-        ORC_LOG("%d: %s %d %d", i,
+        ORC_LOG("%d: %s size %d type %d alloc %d", i,
             program->vars[i].name,
             program->vars[i].size,
-            program->vars[i].vartype);
+            program->vars[i].vartype,
+            program->vars[i].alloc);
       }
     }
     ORC_LOG("instructions");
@@ -259,6 +260,32 @@ orc_program_compile_full (OrcProgram *program, OrcTarget *target,
   orc_compiler_global_reg_alloc (compiler);
 
   orc_compiler_rewrite_vars2 (compiler);
+#if 0
+  {
+    ORC_ERROR("variables");
+    for(i=0;i<ORC_N_VARIABLES;i++){
+      if (compiler->vars[i].size > 0) {
+        ORC_ERROR("%d: %s size %d type %d alloc %d [%d,%d]", i,
+            compiler->vars[i].name,
+            compiler->vars[i].size,
+            compiler->vars[i].vartype,
+            compiler->vars[i].alloc,
+            compiler->vars[i].first_use,
+            compiler->vars[i].last_use);
+      }
+    }
+    ORC_ERROR("instructions");
+    for(i=0;i<compiler->n_insns;i++){
+      ORC_ERROR("%d: %s %d %d %d %d", i,
+          compiler->insns[i].opcode->name,
+          compiler->insns[i].dest_args[0],
+          compiler->insns[i].dest_args[1],
+          compiler->insns[i].src_args[0],
+          compiler->insns[i].src_args[1]);
+    }
+  }
+#endif
+
   if (compiler->error) goto error;
 
   ORC_INFO("allocating code memory");
@@ -467,6 +494,10 @@ orc_compiler_rewrite_vars (OrcCompiler *compiler)
   int var;
   int actual_var;
 
+  for(j=0;j<ORC_N_VARIABLES;j++){
+    if (compiler->vars[j].alloc) continue;
+    compiler->vars[j].last_use = -1;
+  }
   for(j=0;j<compiler->n_insns;j++){
     insn = compiler->insns + j;
     opcode = insn->opcode;
@@ -479,6 +510,12 @@ orc_compiler_rewrite_vars (OrcCompiler *compiler)
       if (compiler->vars[var].vartype == ORC_VAR_TYPE_DEST) {
         compiler->vars[var].load_dest = TRUE;
       }
+      if (compiler->vars[var].vartype == ORC_VAR_TYPE_SRC ||
+          compiler->vars[var].vartype == ORC_VAR_TYPE_DEST ||
+          compiler->vars[var].vartype == ORC_VAR_TYPE_CONST ||
+          compiler->vars[var].vartype == ORC_VAR_TYPE_PARAM) {
+        continue;
+      }
 
       actual_var = var;
       if (compiler->vars[var].replaced) {
@@ -502,6 +539,9 @@ orc_compiler_rewrite_vars (OrcCompiler *compiler)
 
       var = insn->dest_args[k];
 
+      if (compiler->vars[var].vartype == ORC_VAR_TYPE_DEST) {
+        continue;
+      }
       if (compiler->vars[var].vartype == ORC_VAR_TYPE_SRC) {
         ORC_COMPILER_ERROR(compiler,"using src var as dest");
         compiler->result = ORC_COMPILE_RESULT_UNKNOWN_PARSE;
@@ -562,7 +602,6 @@ orc_compiler_global_reg_alloc (OrcCompiler *compiler)
   int i;
   OrcVariable *var;
 
-
   for(i=0;i<ORC_N_VARIABLES;i++){
     var = compiler->vars + i;
     if (var->name == NULL) continue;
@@ -598,6 +637,20 @@ orc_compiler_global_reg_alloc (OrcCompiler *compiler)
     if (compiler->error) break;
   }
 
+  for(i=0;i<compiler->n_insns;i++){
+    OrcInstruction *insn = compiler->insns + i;
+    OrcStaticOpcode *opcode = insn->opcode;
+
+    if (opcode->flags & ORC_STATIC_OPCODE_INVARIANT) {
+      var = compiler->vars + insn->dest_args[0];
+
+      var->first_use = -1;
+      var->last_use = -1;
+      var->alloc = orc_compiler_allocate_register (compiler, TRUE);
+      compiler->insn_flags[i] |= ORC_INSN_FLAG_INVARIANT;
+    }
+  }
+
   if (compiler->alloc_loop_counter && !compiler->error) {
     compiler->loop_counter = orc_compiler_allocate_register (compiler, FALSE);
     /* FIXME massive hack */
@@ -622,6 +675,8 @@ orc_compiler_rewrite_vars2 (OrcCompiler *compiler)
      *  - src1 must be last_use
      *  - only one dest
      */
+    if (compiler->insn_flags[j] & ORC_INSN_FLAG_INVARIANT) continue;
+
     if (!(compiler->insns[j].opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR)
         && compiler->insns[j].opcode->dest_size[1] == 0) {
       int src1 = compiler->insns[j].src_args[0];
@@ -651,6 +706,7 @@ orc_compiler_rewrite_vars2 (OrcCompiler *compiler)
 
     for(i=0;i<ORC_N_VARIABLES;i++){
       if (compiler->vars[i].name == NULL) continue;
+      if (compiler->vars[i].last_use == -1) continue;
       if (compiler->vars[i].first_use == j) {
         if (compiler->vars[i].alloc) continue;
         k = orc_compiler_allocate_register (compiler, TRUE);
diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index a71ab16..eb97fbe 100644
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -23,6 +23,7 @@ void orc_compiler_sse_init (OrcCompiler *compiler);
 unsigned int orc_compiler_sse_get_default_flags (void);
 void orc_compiler_sse_assemble (OrcCompiler *compiler);
 void orc_compiler_sse_register_rules (OrcTarget *target);
+void orc_sse_emit_invariants (OrcCompiler *compiler);
 
 
 void orc_compiler_rewrite_vars (OrcCompiler *compiler);
@@ -303,6 +304,9 @@ sse_load_constants_outer (OrcCompiler *compiler)
         break;
     }
   }
+
+  orc_sse_emit_invariants (compiler);
+
 }
 
 void
@@ -322,8 +326,6 @@ sse_load_constants_inner (OrcCompiler *compiler)
           orc_x86_emit_mov_memoffset_reg (compiler, compiler->is_64bit ? 8 : 4,
               (int)ORC_STRUCT_OFFSET(OrcExecutor, arrays[i]), compiler->exec_reg,
               compiler->vars[i].ptr_register);
-        } else {
-          ORC_COMPILER_ERROR(compiler,"unimplemented");
         }
         break;
       case ORC_VAR_TYPE_ACCUMULATOR:
@@ -687,23 +689,13 @@ orc_sse_emit_loop (OrcCompiler *compiler, int offset, int update)
     insn = compiler->insns + j;
     opcode = insn->opcode;
 
-    ORC_ASM_CODE(compiler,"# %d: %s\n", j, insn->opcode->name);
+    if (compiler->insn_flags[j] & ORC_INSN_FLAG_INVARIANT) continue;
 
-#if 0
-    /* set up args */
-    for(k=0;k<opcode->n_src + opcode->n_dest;k++){
-      args[k] = compiler->vars + insn->args[k];
-      ORC_ASM_CODE(compiler," %d", args[k]->alloc);
-      if (args[k]->is_chained) {
-        ORC_ASM_CODE(compiler," (chained)");
-      }
-    }
-    ORC_ASM_CODE(compiler,"\n");
-#endif
+    ORC_ASM_CODE(compiler,"# %d: %s\n", j, insn->opcode->name);
 
     rule = insn->rule;
     if (rule && rule->emit) {
-      if (!(insn->opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR) &&
+      if (!(insn->opcode->flags & (ORC_STATIC_OPCODE_ACCUMULATOR|ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_STORE)) &&
           compiler->vars[insn->dest_args[0]].alloc !=
           compiler->vars[insn->src_args[0]].alloc) {
         orc_x86_emit_mov_sse_reg_reg (compiler,
@@ -736,3 +728,28 @@ orc_sse_emit_loop (OrcCompiler *compiler, int offset, int update)
   }
 }
 
+void
+orc_sse_emit_invariants (OrcCompiler *compiler)
+{
+  int j;
+  OrcInstruction *insn;
+  OrcStaticOpcode *opcode;
+  OrcRule *rule;
+
+  for(j=0;j<compiler->n_insns;j++){
+    insn = compiler->insns + j;
+    opcode = insn->opcode;
+
+    if (!(compiler->insn_flags[j] & ORC_INSN_FLAG_INVARIANT)) continue;
+
+    ORC_ASM_CODE(compiler,"# %d: %s\n", j, insn->opcode->name);
+
+    rule = insn->rule;
+    if (rule && rule->emit) {
+      rule->emit (compiler, rule->emit_user, insn);
+    } else {
+      ORC_COMPILER_ERROR(compiler,"No rule for: %s", opcode->name);
+    }
+  }
+}
+
diff --git a/orc/orcprogram.h b/orc/orcprogram.h
index 1cff1f8..ae5e558 100644
--- a/orc/orcprogram.h
+++ b/orc/orcprogram.h
@@ -363,6 +363,7 @@ struct _OrcCompiler {
 
   OrcInstruction insns[ORC_N_INSNS];
   int n_insns;
+  int insn_flags[ORC_N_INSNS];
 
   OrcVariable vars[ORC_N_VARIABLES];
   int n_temp_vars;
@@ -410,6 +411,8 @@ struct _OrcCompiler {
   int offset;
 };
 
+#define ORC_INSN_FLAG_INVARIANT 1
+
 #define ORC_SRC_ARG(p,i,n) ((p)->vars[(i)->src_args[(n)]].alloc)
 #define ORC_DEST_ARG(p,i,n) ((p)->vars[(i)->dest_args[(n)]].alloc)
 #define ORC_SRC_TYPE(p,i,n) ((p)->vars[(i)->src_args[(n)]].vartype)
diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index 922cb96..14c9290 100644
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -25,25 +25,14 @@ sse_rule_loadpX (OrcCompiler *compiler, void *user, OrcInstruction *insn)
   if (src->vartype == ORC_VAR_TYPE_PARAM) {
     reg = dest->alloc;
 
-    switch (src->size) {
-      case 1:
-        orc_x86_emit_mov_memoffset_sse (compiler, 4,
-            (int)ORC_STRUCT_OFFSET(OrcExecutor, params[insn->src_args[0]]),
-            compiler->exec_reg, reg, FALSE);
-        orc_sse_emit_punpcklbw (compiler, reg, reg);
-        orc_sse_emit_pshuflw (compiler, 0, reg, reg);
-        break;
-      case 2:
-        orc_x86_emit_mov_memoffset_sse (compiler, 4,
-            (int)ORC_STRUCT_OFFSET(OrcExecutor, params[insn->src_args[0]]),
-            compiler->exec_reg, reg, FALSE);
-        orc_sse_emit_pshuflw (compiler, 0, reg, reg);
-        break;
-      case 4:
-        orc_x86_emit_mov_memoffset_sse (compiler, 4,
-            (int)ORC_STRUCT_OFFSET(OrcExecutor, params[insn->src_args[0]]),
-            compiler->exec_reg, reg, FALSE);
-        break;
+    orc_x86_emit_mov_memoffset_sse (compiler, 4,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor, params[insn->src_args[0]]),
+        compiler->exec_reg, reg, FALSE);
+    if (src->size == 1) {
+      orc_sse_emit_punpcklbw (compiler, reg, reg);
+    }
+    if (src->size <= 2) {
+      orc_sse_emit_pshuflw (compiler, 0, reg, reg);
     }
     orc_sse_emit_pshufd (compiler, 0, reg, reg);
   } else if (src->vartype == ORC_VAR_TYPE_CONST) {
@@ -299,14 +288,32 @@ sse_rule_accsadubl (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src2 = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
+  int tmp2 = X86_XMM7;
 
-  if (p->loop_shift < 2) {
-    ORC_COMPILER_ERROR(p, "accsadubl SSE rule fails with loop_shift < 2");
-    p->result = ORC_COMPILE_RESULT_UNKNOWN_COMPILE;
+  if (p->loop_shift == 0) {
+    orc_sse_emit_movdqa (p, src1, tmp);
+    orc_sse_emit_pslldq (p, 15, tmp);
+    orc_sse_emit_movdqa (p, src2, tmp2);
+    orc_sse_emit_pslldq (p, 15, tmp2);
+    orc_sse_emit_psadbw (p, tmp2, tmp);
+  } else if (p->loop_shift == 1) {
+    orc_sse_emit_movdqa (p, src1, tmp);
+    orc_sse_emit_pslldq (p, 14, tmp);
+    orc_sse_emit_movdqa (p, src2, tmp2);
+    orc_sse_emit_pslldq (p, 14, tmp2);
+    orc_sse_emit_psadbw (p, tmp2, tmp);
+  } else if (p->loop_shift == 2) {
+    orc_sse_emit_movdqa (p, src1, tmp);
+    orc_sse_emit_psadbw (p, src2, tmp);
+    orc_sse_emit_pslldq (p, 12, tmp);
+  } else if (p->loop_shift == 3) {
+    orc_sse_emit_movdqa (p, src1, tmp);
+    orc_sse_emit_psadbw (p, src2, tmp);
+    orc_sse_emit_pslldq (p, 8, tmp);
+  } else {
+    orc_sse_emit_movdqa (p, src1, tmp);
+    orc_sse_emit_psadbw (p, src2, tmp);
   }
-
-  orc_sse_emit_movdqa (p, src1, tmp);
-  orc_sse_emit_psadbw (p, src2, tmp);
   orc_sse_emit_paddd (p, tmp, dest);
 }
 
diff --git a/testsuite/test_accsadubl.c b/testsuite/test_accsadubl.c
index b64656b..1388045 100644
--- a/testsuite/test_accsadubl.c
+++ b/testsuite/test_accsadubl.c
@@ -70,7 +70,7 @@ orc_sad_u8 (orc_uint8 *s1, orc_uint8 *s2, int n)
       return 0;
     }
 
-    printf("%s\n", orc_program_get_asm_code (p));
+    //printf("%s\n", orc_program_get_asm_code (p));
   }
 
   ex = orc_executor_new (p);