util: Include missing header in u_draw.h.

[mesa.git] / src / gallium / auxiliary / tgsi / tgsi_sse2.c

diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index 0820a4137e7d73f282cc20df4e10baf6d653c00d..785a9fb03561758c56b94106fe7284470269802c 100644 (file)
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -1244,16 +1244,9 @@ emit_sub(
       make_xmm( xmm_src ) );
 }
 
-
-
-
-
-
-
 /**
  * Register fetch.
  */
-
 static void
 emit_fetch(
    struct x86_function *func,
@@ -1338,7 +1331,6 @@ emit_fetch(
 /**
  * Register store.
  */
-
 static void
 emit_store(
    struct x86_function *func,
@@ -1455,7 +1447,6 @@ fetch_texel( struct tgsi_sampler **sampler,
 /**
  * High-level instruction translators.
  */
-
 static void
 emit_tex( struct x86_function *func,
           const struct tgsi_full_instruction *inst,
@@ -1507,7 +1498,6 @@ emit_tex( struct x86_function *func,
                get_temp( TEMP_R0, 3 ),
                make_xmm( 3 ) );
 
-   
    if (projected) {
       FETCH( func, *inst, 3, 0, 3 );
 
@@ -1535,7 +1525,6 @@ emit_tex( struct x86_function *func,
    args[0] = get_temp( TEMP_R0, 0 );
    args[1] = get_sampler_ptr( unit );
 
-
    emit_func_call( func,
                    0,
                    args,
@@ -1569,7 +1558,8 @@ emit_kil(
 
    /* This mask stores component bits that were already tested. Note that
     * we test if the value is less than zero, so 1.0 and 0.0 need not to be
-    * tested. */
+    * tested.
+    */
    uniquemask = 0;
 
    FOR_EACH_CHANNEL( chan_index ) {
@@ -1715,22 +1705,26 @@ emit_cmp(
 
 
 /**
- * Check if inst src/dest regs use indirect addressing into temporary
- * register file.
+ * Check if inst src/dest regs use indirect addressing into temporary,
+ * input or output register files.
  */
 static boolean
-indirect_temp_reference(const struct tgsi_full_instruction *inst)
+indirect_reg_reference(const struct tgsi_full_instruction *inst)
 {
    uint i;
    for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
       const struct tgsi_full_src_register *reg = &inst->Src[i];
-      if (reg->Register.File == TGSI_FILE_TEMPORARY &&
+      if ((reg->Register.File == TGSI_FILE_TEMPORARY ||
+           reg->Register.File == TGSI_FILE_INPUT ||
+           reg->Register.File == TGSI_FILE_OUTPUT) &&
           reg->Register.Indirect)
          return TRUE;
    }
    for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
       const struct tgsi_full_dst_register *reg = &inst->Dst[i];
-      if (reg->Register.File == TGSI_FILE_TEMPORARY &&
+      if ((reg->Register.File == TGSI_FILE_TEMPORARY ||
+           reg->Register.File == TGSI_FILE_INPUT ||
+           reg->Register.File == TGSI_FILE_OUTPUT) &&
           reg->Register.Indirect)
          return TRUE;
    }
@@ -1746,7 +1740,7 @@ emit_instruction(
    unsigned chan_index;
 
    /* we can't handle indirect addressing into temp register file yet */
-   if (indirect_temp_reference(inst))
+   if (indirect_reg_reference(inst))
       return FALSE;
 
    switch (inst->Instruction.Opcode) {
@@ -1929,20 +1923,32 @@ emit_instruction(
       break;
 
    case TGSI_OPCODE_MUL:
+      /* do all fetches and adds, storing results in temp regs */
       FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
-         FETCH( func, *inst, 0, 0, chan_index );
-         FETCH( func, *inst, 1, 1, chan_index );
-         emit_mul( func, 0, 1 );
-         STORE( func, *inst, 0, 0, chan_index );
+         int r = chan_index + 1;
+         FETCH( func, *inst, 0, 0, chan_index ); /* load xmm[0] */
+         FETCH( func, *inst, r, 1, chan_index ); /* load xmm[r] */
+         emit_mul( func, r, 0 );   /* xmm[r] = xmm[r] * xmm[0] */
+      }
+      /* do all stores of the temp regs */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         int r = chan_index + 1;
+         STORE( func, *inst, r, 0, chan_index ); /* store xmm[r] */
       }
       break;
 
    case TGSI_OPCODE_ADD:
+      /* do all fetches and adds, storing results in temp regs */
       FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
-         FETCH( func, *inst, 0, 0, chan_index );
-         FETCH( func, *inst, 1, 1, chan_index );
-         emit_add( func, 0, 1 );
-         STORE( func, *inst, 0, 0, chan_index );
+         int r = chan_index + 1;
+         FETCH( func, *inst, 0, 0, chan_index ); /* load xmm[0] */
+         FETCH( func, *inst, r, 1, chan_index ); /* load xmm[r] */
+         emit_add( func, r, 0 );   /* xmm[r] = xmm[r] + xmm[0] */
+      }
+      /* do all stores of the temp regs */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         int r = chan_index + 1;
+         STORE( func, *inst, r, 0, chan_index ); /* store xmm[r] */
       }
       break;
 
@@ -2825,8 +2831,11 @@ static void
 check_soa_dependencies(const struct tgsi_full_instruction *inst)
 {
    switch (inst->Instruction.Opcode) {
+   case TGSI_OPCODE_ADD:
+   case TGSI_OPCODE_MOV:
+   case TGSI_OPCODE_MUL:
    case TGSI_OPCODE_XPD:
-      /* OK */
+      /* OK - these opcodes correctly handle SOA dependencies */
       break;
    default:
       if (tgsi_check_soa_dependencies(inst)) {
@@ -2916,7 +2925,6 @@ tgsi_emit_sse2(
       x86_make_disp( get_machine_base(),
                      Offset( struct tgsi_exec_machine, Samplers ) ) );
 
-
    while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
       tgsi_parse_token( &parse );
 
@@ -3000,4 +3008,3 @@ tgsi_emit_sse2(
 }
 
 #endif /* PIPE_ARCH_X86 */
-