tgsi: reduce x86 reg usage in tgsi_sse generated programs

Pass the tgsi_exec_machine struct in directly and just hold a single
pointer to this struct, rather than keeping one for each of its
internal members.

diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c
index a4f72c40eff1e77ff5b759fa4b5e87d66756f9d2..fb58983e011ffb9bc15fa5347cef468f99b962e5 100644 (file)
--- a/src/gallium/auxiliary/draw/draw_vs_sse.c
+++ b/src/gallium/auxiliary/draw/draw_vs_sse.c
@@ -52,24 +52,12 @@
 
 #define SSE_MAX_VERTICES 4
 
-typedef void (PIPE_CDECL *codegen_function) (
-   const struct tgsi_exec_vector *input, /* 1 */
-   struct tgsi_exec_vector *output, /* 2 */
-   float (*constant)[4],        /* 3 */
-   struct tgsi_exec_vector *temporary, /* 4 */
-   float (*immediates)[4],      /* 5 */
-   const float (*aos_input)[4], /* 6 */
-   uint num_inputs,             /* 7 */
-   uint input_stride,           /* 8 */
-   float (*aos_output)[4],      /* 9 */
-   uint num_outputs,            /* 10 */
-   uint output_stride );        /* 11 */
 
 struct draw_sse_vertex_shader {
    struct draw_vertex_shader base;
    struct x86_function sse2_program;
 
-   codegen_function func;
+   tgsi_sse2_vs_func func;
    
    struct tgsi_exec_machine *machine;
 };
@@ -119,11 +107,9 @@ vs_sse_run_linear( struct draw_vertex_shader *base,
 
       /* run compiled shader
        */
-      shader->func(machine->Inputs,
-                  machine->Outputs,
-                  (float (*)[4])constants,
-                  machine->Temps,
-                  (float (*)[4])shader->base.immediates,
+      shader->func(machine,
+                  constants,
+                  shader->base.immediates,
                    input,
                    base->info.num_inputs,
                    input_stride,
@@ -195,7 +181,7 @@ draw_create_vs_sse(struct draw_context *draw,
                         TRUE )) 
       goto fail;
       
-   vs->func = (codegen_function) x86_get_func( &vs->sse2_program );
+   vs->func = (tgsi_sse2_vs_func) x86_get_func( &vs->sse2_program );
    if (!vs->func) {
       goto fail;
    }

diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index 5084befc4efc0b8cec3665c33214a823de490a66..cfe8ef0ecf166dfffe0d03168ecc048df0ee632d 100644 (file)
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -32,6 +32,7 @@
 #include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "util/u_math.h"
+#include "util/u_memory.h"
 #if defined(PIPE_ARCH_SSE)
 #include "util/u_sse.h"
 #endif
@@ -104,33 +105,43 @@ get_const_base( void )
 }
 
 static struct x86_reg
-get_input_base( void )
+get_machine_base( void )
 {
    return x86_make_reg(
       file_REG32,
       reg_AX );
 }
 
+static struct x86_reg
+get_input_base( void )
+{
+   return x86_make_disp(
+      get_machine_base(),
+      Offset(struct tgsi_exec_machine, Inputs) );
+}
+
 static struct x86_reg
 get_output_base( void )
 {
-   return x86_make_reg(
-      file_REG32,
-      reg_DX );
+   return x86_make_disp(
+      get_machine_base(),
+      Offset(struct tgsi_exec_machine, Outputs) );
 }
 
 static struct x86_reg
 get_temp_base( void )
 {
-   return x86_make_reg(
-      file_REG32,
-      reg_BX );
+   return x86_make_disp(
+      get_machine_base(),
+      Offset(struct tgsi_exec_machine, Temps) );
 }
 
 static struct x86_reg
 get_coef_base( void )
 {
-   return get_output_base();
+   return x86_make_reg(
+      file_REG32,
+      reg_BX );
 }
 
 static struct x86_reg
@@ -138,7 +149,7 @@ get_immediate_base( void )
 {
    return x86_make_reg(
       file_REG32,
-      reg_DI );
+      reg_DX );
 }
 
 
@@ -2551,7 +2562,7 @@ emit_declaration(
 
 static void aos_to_soa( struct x86_function *func, 
                         uint arg_aos,
-                        uint arg_soa, 
+                        uint arg_machine, 
                         uint arg_num, 
                         uint arg_stride )
 {
@@ -2566,7 +2577,10 @@ static void aos_to_soa( struct x86_function *func,
    x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
 
    x86_mov( func, aos_input,  x86_fn_arg( func, arg_aos ) );
-   x86_mov( func, soa_input,  x86_fn_arg( func, arg_soa ) );
+   x86_mov( func, soa_input,  x86_fn_arg( func, arg_machine ) );
+   x86_lea( func, soa_input,  
+           x86_make_disp( soa_input, 
+                          Offset(struct tgsi_exec_machine, Inputs) ) );
    x86_mov( func, num_inputs, x86_fn_arg( func, arg_num ) );
    x86_mov( func, stride,     x86_fn_arg( func, arg_stride ) );
 
@@ -2608,28 +2622,30 @@ static void aos_to_soa( struct x86_function *func,
    x86_jcc( func, cc_NE, inner_loop );
 
    /* Restore EBX */
-   x86_pop( func, aos_input );
+   x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
 }
 
-static void soa_to_aos( struct x86_function *func, uint aos, uint soa, uint num, uint stride )
+static void soa_to_aos( struct x86_function *func, 
+                       uint arg_aos, 
+                       uint arg_machine, 
+                       uint arg_num, 
+                       uint arg_stride )
 {
-   struct x86_reg soa_output;
-   struct x86_reg aos_output;
-   struct x86_reg num_outputs;
-   struct x86_reg temp;
+   struct x86_reg soa_output = x86_make_reg( file_REG32, reg_AX );
+   struct x86_reg aos_output = x86_make_reg( file_REG32, reg_BX );
+   struct x86_reg num_outputs = x86_make_reg( file_REG32, reg_CX );
+   struct x86_reg temp = x86_make_reg( file_REG32, reg_DX );
    int inner_loop;
 
-   soa_output = x86_make_reg( file_REG32, reg_AX );
-   aos_output = x86_make_reg( file_REG32, reg_BX );
-   num_outputs = x86_make_reg( file_REG32, reg_CX );
-   temp = x86_make_reg( file_REG32, reg_DX );
-
    /* Save EBX */
-   x86_push( func, aos_output );
+   x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
 
-   x86_mov( func, soa_output, x86_fn_arg( func, soa ) );
-   x86_mov( func, aos_output, x86_fn_arg( func, aos ) );
-   x86_mov( func, num_outputs, x86_fn_arg( func, num ) );
+   x86_mov( func, aos_output, x86_fn_arg( func, arg_aos ) );
+   x86_mov( func, soa_output, x86_fn_arg( func, arg_machine ) );
+   x86_lea( func, soa_output, 
+           x86_make_disp( soa_output, 
+                          Offset(struct tgsi_exec_machine, Outputs) ) );
+   x86_mov( func, num_outputs, x86_fn_arg( func, arg_num ) );
 
    /* do */
    inner_loop = x86_get_label( func );
@@ -2646,7 +2662,7 @@ static void soa_to_aos( struct x86_function *func, uint aos, uint soa, uint num,
       sse_unpcklps( func, make_xmm( 3 ), make_xmm( 4 ) );
       sse_unpckhps( func, make_xmm( 5 ), make_xmm( 4 ) );
 
-      x86_mov( func, temp, x86_fn_arg( func, stride ) );
+      x86_mov( func, temp, x86_fn_arg( func, arg_stride ) );
       x86_push( func, aos_output );
       sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
       sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
@@ -2670,20 +2686,13 @@ static void soa_to_aos( struct x86_function *func, uint aos, uint soa, uint num,
    x86_jcc( func, cc_NE, inner_loop );
 
    /* Restore EBX */
-   x86_pop( func, aos_output );
+   x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
 }
 
 /**
  * Translate a TGSI vertex/fragment shader to SSE2 code.
  * Slightly different things are done for vertex vs. fragment shaders.
  *
- * Note that fragment shaders are responsible for interpolating shader
- * inputs. Because on x86 we have only 4 GP registers, and here we
- * have 5 shader arguments (input, output, const, temp and coef), the
- * code is split into two phases -- DECLARATION and INSTRUCTION phase.
- * GP register holding the output argument is aliased with the coeff
- * argument, as outputs are not needed in the DECLARATION phase.
- *
  * \param tokens  the TGSI input shader
  * \param func  the output SSE code/function
  * \param immediates  buffer to place immediates, later passed to SSE func
@@ -2697,7 +2706,6 @@ tgsi_emit_sse2(
    boolean do_swizzles )
 {
    struct tgsi_parse_context parse;
-   boolean instruction_phase = FALSE;
    unsigned ok = 1;
    uint num_immediates = 0;
 
@@ -2709,74 +2717,42 @@ tgsi_emit_sse2(
 
    /* Can't just use EDI, EBX without save/restoring them:
     */
-   x86_push(
-      func,
-      get_immediate_base() );
-
-   x86_push(
-      func,
-      get_temp_base() );
-
+   x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
+   x86_push( func, x86_make_reg( file_REG32, reg_DI ) );
 
    /*
     * Different function args for vertex/fragment shaders:
     */
-   if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
-      /* DECLARATION phase, do not load output argument. */
-      x86_mov(
-         func,
-         get_input_base(),
-         x86_fn_arg( func, 1 ) );
-      /* skipping outputs argument here */
-      x86_mov(
-         func,
-         get_const_base(),
-         x86_fn_arg( func, 3 ) );
-      x86_mov(
-         func,
-         get_temp_base(),
-         x86_fn_arg( func, 4 ) );
-      x86_mov(
-         func,
-         get_coef_base(),
-         x86_fn_arg( func, 5 ) );
-      x86_mov(
-         func,
-         get_immediate_base(),
-         x86_fn_arg( func, 6 ) );
-   }
-   else {
-      assert(parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX);
-
+   if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
       if (do_swizzles)
          aos_to_soa( func, 
-                     6,         /* aos_input */
-                     1,         /* machine->input */
-                     7,         /* num_inputs */
-                     8 );       /* input_stride */
+                     4,         /* aos_input */
+                     1,         /* machine */
+                     5,         /* num_inputs */
+                     6 );       /* input_stride */
+   }
 
+   x86_mov(
+      func,
+      get_machine_base(),
+      x86_fn_arg( func, 1 ) );
+   x86_mov(
+      func,
+      get_const_base(),
+      x86_fn_arg( func, 2 ) );
+   x86_mov(
+      func,
+      get_immediate_base(),
+      x86_fn_arg( func, 3 ) );
+
+   if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
       x86_mov(
-         func,
-         get_input_base(),
-         x86_fn_arg( func, 1 ) );
-      x86_mov(
-         func,
-         get_output_base(),
-         x86_fn_arg( func, 2 ) );
-      x86_mov(
-         func,
-         get_const_base(),
-         x86_fn_arg( func, 3 ) );
-      x86_mov(
-         func,
-         get_temp_base(),
-         x86_fn_arg( func, 4 ) );
-      x86_mov(
-         func,
-         get_immediate_base(),
-         x86_fn_arg( func, 5 ) );
+        func,
+        get_coef_base(),
+        x86_fn_arg( func, 4 ) );
    }
 
+
    while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
       tgsi_parse_token( &parse );
 
@@ -2790,17 +2766,6 @@ tgsi_emit_sse2(
          break;
 
       case TGSI_TOKEN_TYPE_INSTRUCTION:
-         if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
-            if( !instruction_phase ) {
-               /* INSTRUCTION phase, overwrite coeff with output. */
-               instruction_phase = TRUE;
-               x86_mov(
-                  func,
-                  get_output_base(),
-                  x86_fn_arg( func, 2 ) );
-            }
-         }
-
          ok = emit_instruction(
             func,
             &parse.FullToken.FullInstruction );
@@ -2844,18 +2809,17 @@ tgsi_emit_sse2(
 
    if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
       if (do_swizzles)
-         soa_to_aos( func, 9, 2, 10, 11 );
+         soa_to_aos( func, 
+                    7,         /* aos_output */
+                    1,         /* machine */
+                    8,         /* num_outputs */
+                    9 );       /* output_stride */
    }
 
    /* Can't just use EBX, EDI without save/restoring them:
     */
-   x86_pop(
-      func,
-      get_temp_base() );
-
-   x86_pop(
-      func,
-      get_immediate_base() );
+   x86_pop( func, x86_make_reg( file_REG32, reg_DI ) );
+   x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
 
    emit_ret( func );
 

diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.h b/src/gallium/auxiliary/tgsi/tgsi_sse2.h
index af838b2a25b8882c9b104f2ba3077067eabae516..d81ee3d00ec90e6f0faffb1ee924ea9353fb27fc 100644 (file)
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.h
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.h
@@ -34,6 +34,7 @@ extern "C" {
 
 struct tgsi_token;
 struct x86_function;
+struct tgsi_interp_coef;
 
 unsigned
 tgsi_emit_sse2(
@@ -42,6 +43,33 @@ tgsi_emit_sse2(
    float (*immediates)[4],
    boolean do_swizzles );
 
+
+/* This is the function prototype generated when do_swizzles is false
+ * -- effectively for fragment shaders.
+ */
+typedef void (PIPE_CDECL *tgsi_sse2_fs_function) (
+   struct tgsi_exec_machine *machine, /* 1 */
+   const float (*constant)[4],             /* 2 */
+   const float (*immediate)[4],                    /* 3 */
+   const struct tgsi_interp_coef *coef     /* 4 */
+   );
+
+
+/* This is the function prototype generated when do_swizzles is true
+ * -- effectively for vertex shaders.
+ */
+typedef void (PIPE_CDECL *tgsi_sse2_vs_func) (
+   struct tgsi_exec_machine *machine, /* 1 */
+   const float (*constant)[4],        /* 2 */
+   const float (*immediate)[4],       /* 3 */
+   const float (*aos_input)[4], /* 4 */
+   uint num_inputs,             /* 5 */
+   uint input_stride,           /* 6 */
+   float (*aos_output)[4],      /* 7 */
+   uint num_outputs,            /* 8 */
+   uint output_stride );        /* 9 */
+
+
 #if defined __cplusplus
 }
 #endif

diff --git a/src/gallium/drivers/softpipe/sp_fs_sse.c b/src/gallium/drivers/softpipe/sp_fs_sse.c
index 31c3ca21c51937129f4f0c16e2844d0dc399693b..f9362efcb71f9657f16f58129ac98bba4631efd1 100644 (file)
--- a/src/gallium/drivers/softpipe/sp_fs_sse.c
+++ b/src/gallium/drivers/softpipe/sp_fs_sse.c
@@ -45,17 +45,6 @@
 
 #include "rtasm/rtasm_x86sse.h"
 
-/* Surely this should be defined somewhere in a tgsi header:
- */
-typedef void (PIPE_CDECL *codegen_function)(
-   const struct tgsi_exec_vector *input,
-   struct tgsi_exec_vector *output,
-   const float (*constant)[4],
-   struct tgsi_exec_vector *temporary,
-   const struct tgsi_interp_coef *coef,
-   float (*immediates)[4]
-   //, const struct tgsi_exec_vector *quadPos
- );
 
 
 /**
@@ -65,7 +54,7 @@ struct sp_sse_fragment_shader
 {
    struct sp_fragment_shader base;
    struct x86_function sse2_program;
-   codegen_function func;
+   tgsi_sse2_fs_function func;
    float immediates[TGSI_EXEC_NUM_IMMEDIATES][4];
 };
 
@@ -107,12 +96,10 @@ fs_sse_run( const struct sp_fragment_shader *base,
    tgsi_set_kill_mask(machine, 0x0);
    tgsi_set_exec_mask(machine, 1, 1, 1, 1);
 
-   shader->func( machine->Inputs,
-                machine->Outputs,
+   shader->func( machine,
                 machine->Consts,
-                machine->Temps,
-                machine->InterpCoefs,
-                 shader->immediates
+                 (const float (*)[4])shader->immediates,
+                machine->InterpCoefs
                 //      , &machine->QuadPos
       );
 
@@ -151,7 +138,7 @@ softpipe_create_fs_sse(struct softpipe_context *softpipe,
       return NULL;
    }
 
-   shader->func = (codegen_function) x86_get_func( &shader->sse2_program );
+   shader->func = (tgsi_sse2_fs_function) x86_get_func( &shader->sse2_program );
    if (!shader->func) {
       x86_release_func( &shader->sse2_program );
       FREE(shader);