Improved register allocation: allow four 'float' vars or temporaries to share a single register. Clean-up needed.

3 files changed, 137 insertions, 64 deletions

diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c
index 82e8c0b1582..c3167e36fb3 100644
--- a/src/mesa/shader/slang/slang_emit.c
+++ b/src/mesa/shader/slang/slang_emit.c
@@ -40,7 +40,7 @@
 
 
 #define PEEPHOLE_OPTIMIZATIONS 1
-#define ANNOTATE 0
+#define ANNOTATE 1
 
 
 /**
@@ -336,11 +336,14 @@ static void
 alloc_temp_storage(slang_var_table *vt, slang_ir_node *n, GLint size)
 {
    GLint indx;
+   GLuint swizzle;
    assert(!n->Var);
    assert(!n->Store);
    assert(size > 0);
-   indx = _slang_alloc_temp(vt, size);
+   indx = _slang_alloc_temp(vt, size, &swizzle);
    n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, indx, size);
+   if (n->Store)
+      n->Store->Swizzle = swizzle;
 }
 
 
@@ -352,8 +355,8 @@ static void
 free_temp_storage(slang_var_table *vt, slang_ir_node *n)
 {
    if (n->Store->File == PROGRAM_TEMPORARY && n->Store->Index >= 0) {
-      if (_slang_is_temp(vt, n->Store->Index)) {
-         _slang_free_temp(vt, n->Store->Index, n->Store->Size);
+      if (_slang_is_temp(vt, n->Store->Index, n->Store->Swizzle)) {
+         _slang_free_temp(vt, n->Store->Index, n->Store->Size, n->Store->Swizzle);
          /* XXX free(store)? */
          n->Store->Index = -1;
          n->Store->Size = -1;
@@ -381,7 +384,15 @@ storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st,
    assert(st->File != PROGRAM_UNDEFINED);
    assert(st->Size >= 1);
    assert(st->Size <= 4);
-   dst->WriteMask = defaultWritemask[st->Size - 1] & writemask;
+   if (st->Size == 1) {
+      GLuint comp = GET_SWZ(st->Swizzle, 0);
+      assert(comp < 4);
+      assert(writemask & WRITEMASK_X);
+      dst->WriteMask = WRITEMASK_X << comp;
+   }
+   else {
+      dst->WriteMask = defaultWritemask[st->Size - 1] & writemask;
+   }
 }
 
 
@@ -803,13 +814,15 @@ emit_move(slang_var_table *vt, slang_ir_node *n, struct gl_program *prog)
    emit(vt, n->Children[0], prog);
 
 #if PEEPHOLE_OPTIMIZATIONS
-   if (inst && _slang_is_temp(vt, n->Children[1]->Store->Index)) {
+   if (inst && _slang_is_temp(vt, n->Children[1]->Store->Index,
+                              n->Children[1]->Store->Swizzle)) {
       /* Peephole optimization:
        * Just modify the RHS to put its result into the dest of this
        * MOVE operation.  Then, this MOVE is a no-op.
        */
       _slang_free_temp(vt, n->Children[1]->Store->Index,
-                       n->Children[1]->Store->Size);
+                       n->Children[1]->Store->Size,
+                       n->Children[1]->Store->Swizzle);
       *n->Children[1]->Store = *n->Children[0]->Store;
       /* fixup the prev (RHS) instruction */
       assert(n->Children[0]->Store->Index >= 0);
@@ -852,12 +865,7 @@ emit_move(slang_var_table *vt, slang_ir_node *n, struct gl_program *prog)
          inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
                                                 srcAnnot, NULL);
       }
-      /* XXX is this test correct? */
-      if (_slang_is_temp(vt, n->Children[1]->Store->Index)) {
-         _slang_free_temp(vt, n->Children[1]->Store->Index,
-                          n->Children[1]->Store->Size);
-      }
-      /*inst->Comment = _mesa_strdup("IR_MOVE");*/
+      free_temp_storage(vt, n->Children[1]);
       assert(!n->Store);
       n->Store = n->Children[0]->Store; /*XXX new */
       return inst;
@@ -883,13 +891,17 @@ emit_cond(slang_var_table *vt, slang_ir_node *n, struct gl_program *prog)
       /* This'll happen for things like "if (i) ..." where no code
        * is normally generated for the expression "i".
        * Generate a move instruction just to set condition codes.
+       * Note: must use full 4-component vector since all four
+       * condition codes must be set identically.
        */
-      alloc_temp_storage(vt, n, 1);
+      alloc_temp_storage(vt, n, 4);
       inst = new_instruction(prog, OPCODE_MOV);
       inst->CondUpdate = GL_TRUE;
       storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
       storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
-      _slang_free_temp(vt, n->Store->Index, n->Store->Size);
+      _slang_free_temp(vt, n->Store->Index, n->Store->Size,
+                       n->Store->Swizzle);
+      inst->Comment = _mesa_strdup("COND expr");
       return inst; /* XXX or null? */
    }
 }
@@ -928,12 +940,16 @@ emit(slang_var_table *vt, slang_ir_node *n, struct gl_program *prog)
       assert(n->Store->Index < 0);
       if (!n->Var || n->Var->isTemp) {
          /* a nameless/temporary variable, will be freed after first use */
-         n->Store->Index = _slang_alloc_temp(vt, n->Store->Size);
+         GLuint swizzle;
+         n->Store->Index = _slang_alloc_temp(vt, n->Store->Size, &swizzle);
+         n->Store->Swizzle = swizzle;
       }
       else {
          /* a regular variable */
+         GLuint swizzle;
          _slang_add_variable(vt, n->Var);
-         n->Store->Index = _slang_alloc_var(vt, n->Store->Size);
+         n->Store->Index = _slang_alloc_var(vt, n->Store->Size, &swizzle);
+         n->Store->Swizzle = swizzle;
          /*
          printf("IR_VAR_DECL %s %d store %p\n",
                 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
diff --git a/src/mesa/shader/slang/slang_vartable.c b/src/mesa/shader/slang/slang_vartable.c
index 6d2b4369d94..cadefda1bca 100644
--- a/src/mesa/shader/slang/slang_vartable.c
+++ b/src/mesa/shader/slang/slang_vartable.c
@@ -4,6 +4,7 @@
 #include "slang_compile_variable.h"
 #include "slang_vartable.h"
 #include "slang_ir.h"
+#include "prog_instruction.h"
 
 
 static int dbg = 0;
@@ -23,7 +24,8 @@ struct slang_var_table_
    int num_entries;
    slang_variable **vars;  /* array [num_entries] */
 
-   TempState temps[MAX_PROGRAM_TEMPS];
+   TempState temps[MAX_PROGRAM_TEMPS * 4];
+   int size[MAX_PROGRAM_TEMPS];
 
    struct slang_var_table_ *parent;
 };
@@ -46,6 +48,7 @@ _slang_push_var_table(slang_var_table *parent)
       if (parent) {
          /* copy the info indicating which temp regs are in use */
          memcpy(t->temps, parent->temps, sizeof(t->temps));
+         memcpy(t->size, parent->size, sizeof(t->size));
       }
       if (dbg) printf("Pushing level %d\n", t->level);
    }
@@ -67,12 +70,22 @@ _slang_pop_var_table(slang_var_table *t)
    /* free the storage allocated for each variable */
    for (i = 0; i < t->num_entries; i++) {
       slang_ir_storage *store = (slang_ir_storage *) t->vars[i]->aux;
-      GLint j, sz4 = (store->Size + 3) / 4;
-      if (dbg) printf("  Free var %s, size %d\n",
-                      (char*) t->vars[i]->a_name, store->Size);
-      for (j = 0; j < sz4; j++) {
-         assert(t->temps[store->Index + j] == VAR);
-         t->temps[store->Index + j] = FREE;
+      GLint j;
+      const GLuint sz = store->Size;
+      GLuint comp;
+      if (dbg) printf("  Free var %s, size %d at %d\n",
+                      (char*) t->vars[i]->a_name, store->Size,
+                      store->Index);
+
+      if (sz == 1)
+         comp = GET_SWZ(store->Swizzle, 0);
+      else
+         comp = 0;
+
+      assert(store->Index >= 0);
+      for (j = 0; j < sz; j++) {
+         assert(t->temps[store->Index * 4 + j + comp] == VAR);
+         t->temps[store->Index * 4 + j + comp] = FREE;
       }
       store->Index = -1;
    }
@@ -80,9 +93,9 @@ _slang_pop_var_table(slang_var_table *t)
       /* just verify that any remaining allocations in this scope 
        * were for temps
        */
-      for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
+      for (i = 0; i < MAX_PROGRAM_TEMPS * 4; i++) {
          if (t->temps[i] && !t->parent->temps[i]) {
-            if (dbg) printf("  Free reg %d\n", i);
+            if (dbg) printf("  Free reg %d\n", i/4);
             assert(t->temps[i] == TEMP);
          }
       }
@@ -131,16 +144,22 @@ _slang_find_variable(const slang_var_table *t, slang_atom name)
 }
 
 
+/**
+ * Allocation helper.
+ * \param size  var size in floats
+ * \return  position for var, measured in floats
+ */
 static GLint
 alloc_reg(slang_var_table *t, GLint size, GLboolean isTemp)
 {
-   const GLuint sz4 = (size + 3) / 4;
+   /* if size == 1, allocate anywhere, else, pos must be multiple of 4 */
+   const GLuint step = (size == 1) ? 1 : 4;
    GLuint i, j;
    assert(size > 0); /* number of floats */
 
-   for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
+   for (i = 0; i < MAX_PROGRAM_TEMPS - size; i += step) {
       GLuint found = 0;
-      for (j = 0; j < sz4; j++) {
+      for (j = 0; j < size; j++) {
          if (i + j < MAX_PROGRAM_TEMPS && !t->temps[i + j]) {
             found++;
          }
@@ -148,10 +167,14 @@ alloc_reg(slang_var_table *t, GLint size, GLboolean isTemp)
             break;
          }
       }
-      if (found == sz4) {
-         /* found block of size/4 free regs */
-         for (j = 0; j < sz4; j++)
+      if (found == size) {
+         /* found block of size free regs */
+         if (size > 1)
+            assert(i % 4 == 0);
+         for (j = 0; j < size; j++)
             t->temps[i + j] = isTemp ? TEMP : VAR;
+         printf("t->size[%d] = %d\n", i, size);
+         t->size[i] = size;
          return i;
       }
    }
@@ -161,61 +184,98 @@ alloc_reg(slang_var_table *t, GLint size, GLboolean isTemp)
 
 /**
  * Allocate temp register(s) for storing a variable.
+ * \param size  size needed, in floats
+ * \param swizzle  returns swizzle mask for accessing var in register
+ * \return  register allocated, or -1
  */
 GLint
-_slang_alloc_var(slang_var_table *t, GLint size)
+_slang_alloc_var(slang_var_table *t, GLint size, GLuint *swizzle)
 {
    int i = alloc_reg(t, size, GL_FALSE);
-   if (dbg) printf("Alloc var %d (level %d)\n", i, t->level);
-   return i;
-}
-
-
-void
-_slang_reserve_var(slang_var_table *t, GLint r, GLint size)
-{
-   const GLint sz4 = (size + 3) / 4;
-   GLint i;
-   for (i = 0; i < sz4; i++) {
-      t->temps[r + i] = VAR;
+   if (i < 0)
+      return -1;
+
+   if (size == 1) {
+      GLuint comp = i % 4;
+      *swizzle = MAKE_SWIZZLE4(comp, comp, comp, comp);
+      char swz = "xyzw"[comp];
+      if (dbg) printf("Alloc var sz %d at %d.%c (level %d)\n", size, i/4, swz, t->level);
    }
+   else {
+      *swizzle = SWIZZLE_NOOP;
+      if (dbg) printf("Alloc var sz %d at %d.xyzw (level %d)\n", size, i/4, t->level);
+   }
+   return i / 4;
 }
 
 
+
 /**
  * Allocate temp register(s) for storing an unnamed intermediate value.
  */
 GLint
-_slang_alloc_temp(slang_var_table *t, GLint size)
+_slang_alloc_temp(slang_var_table *t, GLint size, GLuint *swizzle)
 {
    int i = alloc_reg(t, size, GL_TRUE);
-   if (dbg) printf("Alloc temp %d (level %d)\n", i, t->level);
-   return i;
+   if (i < 0)
+      return -1;
+
+   if (size == 1) {
+      GLuint comp = i % 4;
+      assert(comp < 4);
+      int swz = "xyzw"[comp];
+      *swizzle = MAKE_SWIZZLE4(comp, comp, comp, comp);
+      if (dbg) printf("Alloc temp sz %d at %d.%c (level %d)\n",
+                      size, i/4, swz, t->level);
+   }
+   else {
+      *swizzle = SWIZZLE_NOOP;
+      if (dbg) printf("Alloc temp sz %d at %d.xyzw (level %d)\n",
+                      size, i/4, t->level);
+   }
+   return i / 4;
 }
 
 
 void
-_slang_free_temp(slang_var_table *t, GLint r, GLint size)
+_slang_free_temp(slang_var_table *t, GLint r, GLint size, GLuint swizzle)
 {
-   const GLuint sz4 = (size + 3) / 4;
    GLuint i;
    assert(size > 0);
    assert(r >= 0);
-   assert(r < MAX_PROGRAM_TEMPS);
-   if (dbg) printf("Free temp %d (level %d)\n", r, t->level);
-   for (i = 0; i < sz4; i++) {
-      assert(t->temps[r + i] == TEMP);
-      t->temps[r + i] = FREE;
+   assert(r + size <= MAX_PROGRAM_TEMPS);
+   if (dbg) printf("Free temp sz %d at %d (level %d)\n", size, r, t->level);
+   if (size == 1) {
+      GLuint comp = GET_SWZ(swizzle, 0);
+      assert(swizzle == MAKE_SWIZZLE4(comp, comp, comp, comp));
+      assert(comp < 4);
+      assert(t->size[r * 4 + comp] == 1);
+      assert(t->temps[r * 4 + comp] == TEMP);
+      t->temps[r * 4 + comp] = FREE;
+   }
+   else {
+      assert(swizzle == SWIZZLE_NOOP);
+      assert(t->size[r*4] == size);
+      for (i = 0; i < size; i++) {
+         assert(t->temps[r * 4 + i] == TEMP);
+         t->temps[r * 4 + i] = FREE;
+      }
    }
 }
 
 
 GLboolean
-_slang_is_temp(slang_var_table *t, GLint r)
+_slang_is_temp(slang_var_table *t, GLint r, GLuint swizzle)
 {
    assert(r >= 0);
    assert(r < MAX_PROGRAM_TEMPS);
-   if (t->temps[r] == TEMP)
+   GLuint comp;
+   if (swizzle == SWIZZLE_NOOP)
+      comp = 0;
+   else
+      comp = GET_SWZ(swizzle, 0);
+
+   if (t->temps[r * 4 + comp] == TEMP)
       return GL_TRUE;
    else
       return GL_FALSE;
diff --git a/src/mesa/shader/slang/slang_vartable.h b/src/mesa/shader/slang/slang_vartable.h
index c8e37c94aa5..86fa5d4a4a1 100644
--- a/src/mesa/shader/slang/slang_vartable.h
+++ b/src/mesa/shader/slang/slang_vartable.h
@@ -20,19 +20,16 @@ extern struct slang_variable_ *
 _slang_find_variable(const slang_var_table *t, slang_atom name);
 
 extern GLint
-_slang_alloc_var(slang_var_table *t, GLint size);
-
-extern void
-_slang_reserve_var(slang_var_table *t, GLint r, GLint size);
+_slang_alloc_var(slang_var_table *t, GLint size, GLuint *swizzle);
 
 extern GLint
-_slang_alloc_temp(slang_var_table *t, GLint size);
+_slang_alloc_temp(slang_var_table *t, GLint size, GLuint *swizzle);
 
 extern void
-_slang_free_temp(slang_var_table *t, GLint r, GLint size);
+_slang_free_temp(slang_var_table *t, GLint r, GLint size, GLuint swizzle);
 
 extern GLboolean
-_slang_is_temp(slang_var_table *t, GLint r);
+_slang_is_temp(slang_var_table *t, GLint r, GLuint swizzle);
 
 
 #endif /* SLANG_VARTABLE_H */