gallium: call tgsi_set_exec_mask() and use exec mask in SSE ARL code

This prevents vertex shaders from referencing invalid memory locations when
the shader is operating on less than four vertices or fragments.

4 files changed, 54 insertions, 4 deletions

diff --git a/src/gallium/auxiliary/draw/draw_vs_exec.c b/src/gallium/auxiliary/draw/draw_vs_exec.c
index 44563803f90..82d27d44934 100644
--- a/src/gallium/auxiliary/draw/draw_vs_exec.c
+++ b/src/gallium/auxiliary/draw/draw_vs_exec.c
@@ -120,6 +120,12 @@ vs_exec_run_linear( struct draw_vertex_shader *shader,
 	 input = (const float (*)[4])((const char *)input + input_stride);
       } 
 
+      tgsi_set_exec_mask(machine,
+                         1,
+                         max_vertices > 1,
+                         max_vertices > 2,
+                         max_vertices > 3);
+
       /* run interpreter */
       tgsi_exec_machine_run( machine );
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c
index 0efabd9de8b..77ba5152f9f 100644
--- a/src/gallium/auxiliary/draw/draw_vs_sse.c
+++ b/src/gallium/auxiliary/draw/draw_vs_sse.c
@@ -99,9 +99,23 @@ vs_sse_run_linear( struct draw_vertex_shader *base,
    struct tgsi_exec_machine *machine = shader->machine;
    unsigned int i;
 
+   /* By default, execute all channels.  XXX move this inside the loop
+    * below when we support shader conditionals/loops.
+    */
+   tgsi_set_exec_mask(machine, 1, 1, 1, 1);
+
    for (i = 0; i < count; i += MAX_TGSI_VERTICES) {
       unsigned int max_vertices = MIN2(MAX_TGSI_VERTICES, count - i);
 
+      if (max_vertices < 4) {
+         /* disable the unused execution channels */
+         tgsi_set_exec_mask(machine,
+                            1,
+                            max_vertices > 1,
+                            max_vertices > 2,
+                            0);
+      }
+
       /* run compiled shader
        */
       shader->func(machine->Inputs,
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index c115956c5d7..4d59106dbfa 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -69,6 +69,9 @@
 
 #define TEMP_R0   TGSI_EXEC_TEMP_R0
 #define TEMP_ADDR TGSI_EXEC_TEMP_ADDR
+#define TEMP_EXEC_MASK_I TGSI_EXEC_MASK_I
+#define TEMP_EXEC_MASK_C TGSI_EXEC_MASK_C
+
 
 /**
  * X86 utility functions.
@@ -230,6 +233,9 @@ emit_const(
    int indirectIndex )
 {
    if (indirect) {
+      /* 'vec' is the offset from the address register's value.
+       * We're loading CONST[ADDR+vec] into an xmm register.
+       */
       struct x86_reg r0 = get_input_base();
       struct x86_reg r1 = get_output_base();
       uint i;
@@ -240,18 +246,40 @@ emit_const(
       x86_push( func, r0 );
       x86_push( func, r1 );
 
+      /*
+       * Loop over the four pixels or vertices in the quad.
+       * Get the value of the address (offset) register for pixel/vertex[i],
+       * add it to the src offset and index into the constant buffer.
+       * Note that we're working on SOA data.
+       * If any of the pixel/vertex execution channels are unused their
+       * values will be garbage.  It's very important that we don't use
+       * those garbage values as indexes into the constant buffer since
+       * that'll cause segfaults.
+       * The solution is to bitwise-AND the offset with the execution mask
+       * register whose values are either 0 or ~0.
+       * The caller must setup the execution mask register to indicate
+       * which channels are valid/alive before running the shader.
+       * The execution mask will also figure into loops and conditionals
+       * someday.
+       */
       for (i = 0; i < QUAD_SIZE; i++) {
-         x86_lea( func, r0, get_const( vec, chan ) );
+         /* r1 = address register[i] */
          x86_mov( func, r1, x86_make_disp( get_temp( TEMP_ADDR, CHAN_X ), i * 4 ) );
+         /* r0 = execution mask[i] */
+         x86_mov( func, r0, x86_make_disp( get_temp( TEMP_EXEC_MASK_I, TEMP_EXEC_MASK_C ), i * 4 ) );
+         /* r1 = r1 & r0 */
+         x86_and( func, r1, r0 );
+         /* r0 = 'vec', the offset */
+         x86_lea( func, r0, get_const( vec, chan ) );
 
-         /* Quick hack to multiply by 16 -- need to add SHL to rtasm.
+         /* Quick hack to multiply r1 by 16 -- need to add SHL to rtasm.
           */
          x86_add( func, r1, r1 );
          x86_add( func, r1, r1 );
          x86_add( func, r1, r1 );
          x86_add( func, r1, r1 );
 
-         x86_add( func, r0, r1 );
+         x86_add( func, r0, r1 );  /* r0 = r0 + r1 */
          x86_mov( func, r1, x86_deref( r0 ) );
          x86_mov( func, x86_make_disp( get_temp( TEMP_R0, CHAN_X ), i * 4 ), r1 );
       }
@@ -265,6 +293,7 @@ emit_const(
          get_temp( TEMP_R0, CHAN_X ) );
    }
    else {
+      /* 'vec' is the index into the src register file, such as TEMP[vec] */
       assert( vec >= 0 );
 
       sse_movss(
diff --git a/src/gallium/drivers/softpipe/sp_fs_sse.c b/src/gallium/drivers/softpipe/sp_fs_sse.c
index 496ed43df26..50eb2c07bcb 100644
--- a/src/gallium/drivers/softpipe/sp_fs_sse.c
+++ b/src/gallium/drivers/softpipe/sp_fs_sse.c
@@ -92,7 +92,8 @@ fs_sse_run( const struct sp_fragment_shader *base,
 		       machine->Temps);
 
    /* init kill mask */
-   machine->Temps[TGSI_EXEC_TEMP_KILMASK_I].xyzw[TGSI_EXEC_TEMP_KILMASK_C].u[0] = 0x0;
+   tgsi_set_kill_mask(machine, 0x0);
+   tgsi_set_exec_mask(machine, 1, 1, 1, 1);
 
    shader->func( machine->Inputs,
 		 machine->Outputs,