draw: fix vs aos internal/machine state

2 files changed, 63 insertions, 5 deletions

diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index b8e66e8b78c..67761f881d8 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -83,7 +83,7 @@ static struct x86_reg get_reg_ptr(struct aos_compilation *cp,
       return x86_make_disp(ptr, Offset(struct aos_machine, constant[idx]));
 
    case AOS_FILE_INTERNAL:
-      return x86_make_disp(ptr, Offset(struct aos_machine, immediate[idx]));
+      return x86_make_disp(ptr, Offset(struct aos_machine, internal[idx]));
 
    default:
       ERROR(cp, "unknown reg file");
@@ -97,9 +97,63 @@ struct x86_reg aos_get_internal( struct aos_compilation *cp,
 {
    return get_reg_ptr( cp,
                        AOS_FILE_INTERNAL, 
-                       imm + 1 );
+                       imm );
+}
+
+#define X87_CW_EXCEPTION_INV_OP       (1<<0)
+#define X87_CW_EXCEPTION_DENORM_OP    (1<<1)
+#define X87_CW_EXCEPTION_ZERO_DIVIDE  (1<<2)
+#define X87_CW_EXCEPTION_OVERFLOW     (1<<3)
+#define X87_CW_EXCEPTION_UNDERFLOW    (1<<4)
+#define X87_CW_EXCEPTION_PRECISION    (1<<5)
+#define X87_CW_PRECISION_SINGLE       (0<<8)
+#define X87_CW_PRECISION_RESERVED     (1<<8)
+#define X87_CW_PRECISION_DOUBLE       (2<<8)
+#define X87_CW_PRECISION_DOUBLE_EXT   (3<<8)
+#define X87_CW_PRECISION_MASK         (3<<8)
+#define X87_CW_ROUND_NEAREST          (0<<10)
+#define X87_CW_ROUND_DOWN             (1<<10)
+#define X87_CW_ROUND_UP               (2<<10)
+#define X87_CW_ROUND_ZERO             (3<<10)
+#define X87_CW_ROUND_MASK             (3<<10)
+#define X87_CW_INFINITY               (1<<12)
+
+static void init_internals( struct aos_machine *machine )
+{
+   float inv = 1.0f/255.0f;
+   float f255 = 255.0f;
+
+   ASSIGN_4V(machine->internal[IMM_ONES],      1.0f,  1.0f,  1.0f,  1.0f);
+   ASSIGN_4V(machine->internal[IMM_NEGS],     -1.0f, -1.0f, -1.0f, -1.0f);
+   ASSIGN_4V(machine->internal[IMM_IDENTITY],  0.0f,  0.0f,  0.0f,  1.0f);
+   ASSIGN_4V(machine->internal[IMM_INV_255],   inv,   inv,   inv,   inv);
+   ASSIGN_4V(machine->internal[IMM_255],       f255,   f255,   f255,   f255);
+
+
+   machine->fpu_rnd_nearest = (X87_CW_EXCEPTION_INV_OP |
+                               X87_CW_EXCEPTION_DENORM_OP |
+                               X87_CW_EXCEPTION_ZERO_DIVIDE |
+                               X87_CW_EXCEPTION_OVERFLOW |
+                               X87_CW_EXCEPTION_UNDERFLOW |
+                               X87_CW_EXCEPTION_PRECISION |
+                               (1<<6) |
+                               X87_CW_ROUND_NEAREST |
+                               X87_CW_PRECISION_DOUBLE_EXT);
+
+   assert(machine->fpu_rnd_nearest == 0x37f);
+                               
+   machine->fpu_rnd_neg_inf = (X87_CW_EXCEPTION_INV_OP |
+                               X87_CW_EXCEPTION_DENORM_OP |
+                               X87_CW_EXCEPTION_ZERO_DIVIDE |
+                               X87_CW_EXCEPTION_OVERFLOW |
+                               X87_CW_EXCEPTION_UNDERFLOW |
+                               X87_CW_EXCEPTION_PRECISION |
+                               (1<<6) |
+                               X87_CW_ROUND_DOWN |
+                               X87_CW_PRECISION_DOUBLE_EXT);
 }
 
+
 static void spill( struct aos_compilation *cp, unsigned idx )
 {
    if (!cp->xmm[idx].dirty ||
@@ -1736,6 +1790,7 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
       goto fail;
    
    memset(vaos->machine, 0, sizeof(struct aos_machine));
+   init_internals(vaos->machine);
 
    tgsi_dump(vs->state.tokens, 0);
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h
index 16fef6451c4..c2afd4e9a0d 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.h
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.h
@@ -52,10 +52,13 @@ struct x86_function;
 #define MAX_TEMPS      PIPE_MAX_ATTRIBS /* say */
 #define MAX_CONSTANTS  PIPE_MAX_ATTRIBS /* say */
 #define MAX_IMMEDIATES PIPE_MAX_ATTRIBS /* say */
-#define MAX_INTERNALS  4
+#define MAX_INTERNALS  8
 
 #define AOS_FILE_INTERNAL TGSI_FILE_COUNT
 
+#define FPU_RND_NEG    1
+#define FPU_RND_NEAREST 2
+
 /* This is the temporary storage used by all the aos_sse vs varients.
  * Create one per context and reuse by passing a pointer in at
  * vs_varient creation??
@@ -71,8 +74,8 @@ struct aos_machine {
    float scale[4];              /* viewport */
    float translate[4];          /* viewport */
 
-   ushort fpu_round_nearest;
-   ushort fpu_round_neg_inf;
+   ushort fpu_rnd_nearest;
+   ushort fpu_rnd_neg_inf;
    ushort fpu_restore;
    ushort fpucntl;              /* one of FPU_* above */