X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fauxiliary%2Ftgsi%2Ftgsi_exec.c;h=b7569e74d4b6986ddf172eccdfbb707cb229f2e8;hb=48302e9b309c7ce218de2e522c91bdc87e61cdbc;hp=5af0a947947fed528ed5553bd76c2230c22c7ade;hpb=1f713059934c2365a42aadfdca49213b36a7de90;p=mesa.git

diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index 5af0a947947..b7569e74d4b 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -62,6 +62,9 @@
 
 #define FAST_MATH 1
 
+/** for tgsi_full_instruction::Flags */
+#define SOA_DEPENDENCY_FLAG 0x1
+
 #define TILE_TOP_LEFT     0
 #define TILE_TOP_RIGHT    1
 #define TILE_BOTTOM_LEFT  2
@@ -104,6 +107,7 @@
 #define TEMP_HALF_I        TGSI_EXEC_TEMP_HALF_I
 #define TEMP_HALF_C        TGSI_EXEC_TEMP_HALF_C
 #define TEMP_R0            TGSI_EXEC_TEMP_R0
+#define TEMP_P0            TGSI_EXEC_TEMP_P0
 
 #define IS_CHANNEL_ENABLED(INST, CHAN)\
    ((INST).FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN)))
@@ -182,7 +186,7 @@ print_temp(const struct tgsi_exec_machine *mach, uint index)
  *   MOV t3, t2;
  * The second instruction will have the wrong value for t0 if executed as-is.
  */
-static boolean
+boolean
 tgsi_check_soa_dependencies(const struct tgsi_full_instruction *inst)
 {
    uint i, chan;
@@ -207,9 +211,8 @@ tgsi_check_soa_dependencies(const struct tgsi_full_instruction *inst)
          uint channelsWritten = 0x0;
          FOR_EACH_ENABLED_CHANNEL(*inst, chan) {
             /* check if we're reading a channel that's been written */
-            uint swizzle = tgsi_util_get_full_src_register_extswizzle(&inst->FullSrcRegisters[i], chan);
-            if (swizzle <= TGSI_SWIZZLE_W &&
-                (channelsWritten & (1 << swizzle))) {
+            uint swizzle = tgsi_util_get_full_src_register_swizzle(&inst->FullSrcRegisters[i], chan);
+            if (channelsWritten & (1 << swizzle)) {
                return TRUE;
             }
 
@@ -328,19 +331,24 @@ tgsi_exec_machine_bind_shader(
                                    * sizeof(struct tgsi_full_instruction));
             maxInstructions += 10;
          }
-         memcpy(instructions + numInstructions,
-                &parse.FullToken.FullInstruction,
-                sizeof(instructions[0]));
 
-#if 0
          if (tgsi_check_soa_dependencies(&parse.FullToken.FullInstruction)) {
-            debug_printf("SOA dependency in instruction:\n");
-            tgsi_dump_instruction(&parse.FullToken.FullInstruction,
-                                  numInstructions);
+            uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
+            parse.FullToken.FullInstruction.Flags = SOA_DEPENDENCY_FLAG;
+            /* XXX we only handle SOA dependencies properly for MOV/SWZ
+             * at this time!
+             */
+            if (opcode != TGSI_OPCODE_MOV) {
+               debug_printf("Warning: SOA dependency in instruction"
+                            " is not handled:\n");
+               tgsi_dump_instruction(&parse.FullToken.FullInstruction,
+                                     numInstructions);
+            }
          }
-#else
-         (void) tgsi_check_soa_dependencies;
-#endif
+
+         memcpy(instructions + numInstructions,
+                &parse.FullToken.FullInstruction,
+                sizeof(instructions[0]));
 
          numInstructions++;
          break;
@@ -375,15 +383,9 @@ tgsi_exec_machine_create( void )
    if (!mach)
       goto fail;
 
-   mach->Addrs = &mach->Temps[TGSI_EXEC_TEMP_ADDR];
+   memset(mach, 0, sizeof(*mach));
 
-   mach->Samplers = NULL;
-   mach->Consts = NULL;
-   mach->Tokens = NULL;
-   mach->Primitives = NULL;
-   mach->InterpCoefs = NULL;
-   mach->Instructions = NULL;
-   mach->Declarations = NULL;
+   mach->Addrs = &mach->Temps[TGSI_EXEC_TEMP_ADDR];
 
    /* Setup constants. */
    for( i = 0; i < 4; i++ ) {
@@ -1128,10 +1130,10 @@ fetch_src_file_channel(
    union tgsi_exec_channel *chan )
 {
    switch( swizzle ) {
-   case TGSI_EXTSWIZZLE_X:
-   case TGSI_EXTSWIZZLE_Y:
-   case TGSI_EXTSWIZZLE_Z:
-   case TGSI_EXTSWIZZLE_W:
+   case TGSI_SWIZZLE_X:
+   case TGSI_SWIZZLE_Y:
+   case TGSI_SWIZZLE_Z:
+   case TGSI_SWIZZLE_W:
       switch( file ) {
       case TGSI_FILE_CONSTANT:
          assert(mach->Consts);
@@ -1186,6 +1188,17 @@ fetch_src_file_channel(
          chan->u[3] = mach->Addrs[index->i[3]].xyzw[swizzle].u[3];
          break;
 
+      case TGSI_FILE_PREDICATE:
+         assert(index->i[0] < TGSI_EXEC_NUM_PREDS);
+         assert(index->i[1] < TGSI_EXEC_NUM_PREDS);
+         assert(index->i[2] < TGSI_EXEC_NUM_PREDS);
+         assert(index->i[3] < TGSI_EXEC_NUM_PREDS);
+         chan->u[0] = mach->Addrs[0].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Addrs[0].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Addrs[0].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Addrs[0].xyzw[swizzle].u[3];
+         break;
+
       case TGSI_FILE_OUTPUT:
          /* vertex/fragment output vars can be read too */
          chan->u[0] = mach->Outputs[index->i[0]].xyzw[swizzle].u[0];
@@ -1199,14 +1212,6 @@ fetch_src_file_channel(
       }
       break;
 
-   case TGSI_EXTSWIZZLE_ZERO:
-      *chan = mach->Temps[TEMP_0_I].xyzw[TEMP_0_C];
-      break;
-
-   case TGSI_EXTSWIZZLE_ONE:
-      *chan = mach->Temps[TEMP_1_I].xyzw[TEMP_1_C];
-      break;
-
    default:
       assert( 0 );
    }
@@ -1365,7 +1370,7 @@ fetch_source(
        */
    }
 
-   swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+   swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
    fetch_src_file_channel(
       mach,
       reg->SrcRegister.File,
@@ -1473,119 +1478,17 @@ store_dest(
       dst = &mach->Addrs[index].xyzw[chan_index];
       break;
 
+   case TGSI_FILE_PREDICATE:
+      index = reg->DstRegister.Index;
+      assert(index < TGSI_EXEC_NUM_PREDS);
+      dst = &mach->Addrs[index].xyzw[chan_index];
+      break;
+
    default:
       assert( 0 );
       return;
    }
 
-   if (inst->InstructionExtNv.CondFlowEnable) {
-      union tgsi_exec_channel *cc = &mach->Temps[TEMP_CC_I].xyzw[TEMP_CC_C];
-      uint swizzle;
-      uint shift;
-      uint mask;
-      uint test;
-
-      /* Only CC0 supported.
-       */
-      assert( inst->InstructionExtNv.CondFlowIndex < 1 );
-
-      switch (chan_index) {
-      case CHAN_X:
-         swizzle = inst->InstructionExtNv.CondSwizzleX;
-         break;
-      case CHAN_Y:
-         swizzle = inst->InstructionExtNv.CondSwizzleY;
-         break;
-      case CHAN_Z:
-         swizzle = inst->InstructionExtNv.CondSwizzleZ;
-         break;
-      case CHAN_W:
-         swizzle = inst->InstructionExtNv.CondSwizzleW;
-         break;
-      default:
-         assert( 0 );
-         return;
-      }
-
-      switch (swizzle) {
-      case TGSI_SWIZZLE_X:
-         shift = TGSI_EXEC_CC_X_SHIFT;
-         mask = TGSI_EXEC_CC_X_MASK;
-         break;
-      case TGSI_SWIZZLE_Y:
-         shift = TGSI_EXEC_CC_Y_SHIFT;
-         mask = TGSI_EXEC_CC_Y_MASK;
-         break;
-      case TGSI_SWIZZLE_Z:
-         shift = TGSI_EXEC_CC_Z_SHIFT;
-         mask = TGSI_EXEC_CC_Z_MASK;
-         break;
-      case TGSI_SWIZZLE_W:
-         shift = TGSI_EXEC_CC_W_SHIFT;
-         mask = TGSI_EXEC_CC_W_MASK;
-         break;
-      default:
-         assert( 0 );
-         return;
-      }
-
-      switch (inst->InstructionExtNv.CondMask) {
-      case TGSI_CC_GT:
-         test = ~(TGSI_EXEC_CC_GT << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_EQ:
-         test = ~(TGSI_EXEC_CC_EQ << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_LT:
-         test = ~(TGSI_EXEC_CC_LT << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_GE:
-         test = ~((TGSI_EXEC_CC_GT | TGSI_EXEC_CC_EQ) << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_LE:
-         test = ~((TGSI_EXEC_CC_LT | TGSI_EXEC_CC_EQ) << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_NE:
-         test = ~((TGSI_EXEC_CC_GT | TGSI_EXEC_CC_LT | TGSI_EXEC_CC_UN) << shift) & mask;
-         for (i = 0; i < QUAD_SIZE; i++)
-            if (cc->u[i] & test)
-               execmask &= ~(1 << i);
-         break;
-
-      case TGSI_CC_TR:
-         break;
-
-      case TGSI_CC_FL:
-         for (i = 0; i < QUAD_SIZE; i++)
-            execmask &= ~(1 << i);
-         break;
-
-      default:
-         assert( 0 );
-         return;
-      }
-   }
-
    switch (inst->Instruction.Saturate) {
    case TGSI_SAT_NONE:
       for (i = 0; i < QUAD_SIZE; i++)
@@ -1620,51 +1523,6 @@ store_dest(
    default:
       assert( 0 );
    }
-
-   if (inst->InstructionExtNv.CondDstUpdate) {
-      union tgsi_exec_channel *cc = &mach->Temps[TEMP_CC_I].xyzw[TEMP_CC_C];
-      uint shift;
-      uint mask;
-
-      /* Only CC0 supported.
-       */
-      assert( inst->InstructionExtNv.CondDstIndex < 1 );
-
-      switch (chan_index) {
-      case CHAN_X:
-         shift = TGSI_EXEC_CC_X_SHIFT;
-         mask = ~TGSI_EXEC_CC_X_MASK;
-         break;
-      case CHAN_Y:
-         shift = TGSI_EXEC_CC_Y_SHIFT;
-         mask = ~TGSI_EXEC_CC_Y_MASK;
-         break;
-      case CHAN_Z:
-         shift = TGSI_EXEC_CC_Z_SHIFT;
-         mask = ~TGSI_EXEC_CC_Z_MASK;
-         break;
-      case CHAN_W:
-         shift = TGSI_EXEC_CC_W_SHIFT;
-         mask = ~TGSI_EXEC_CC_W_MASK;
-         break;
-      default:
-         assert( 0 );
-         return;
-      }
-
-      for (i = 0; i < QUAD_SIZE; i++)
-         if (execmask & (1 << i)) {
-            cc->u[i] &= mask;
-            if (dst->f[i] < 0.0f)
-               cc->u[i] |= TGSI_EXEC_CC_LT << shift;
-            else if (dst->f[i] > 0.0f)
-               cc->u[i] |= TGSI_EXEC_CC_GT << shift;
-            else if (dst->f[i] == 0.0f)
-               cc->u[i] |= TGSI_EXEC_CC_EQ << shift;
-            else
-               cc->u[i] |= TGSI_EXEC_CC_UN << shift;
-         }
-   }
 }
 
 #define FETCH(VAL,INDEX,CHAN)\
@@ -1687,10 +1545,8 @@ exec_kil(struct tgsi_exec_machine *mach,
    uint kilmask = 0; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
    union tgsi_exec_channel r[1];
 
-   /* 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. */
-   uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+   /* This mask stores component bits that were already tested. */
+   uniquemask = 0;
 
    for (chan_index = 0; chan_index < 4; chan_index++)
    {
@@ -1698,7 +1554,7 @@ exec_kil(struct tgsi_exec_machine *mach,
       uint i;
 
       /* unswizzle channel */
-      swizzle = tgsi_util_get_full_src_register_extswizzle (
+      swizzle = tgsi_util_get_full_src_register_swizzle (
                         &inst->FullSrcRegisters[0],
                         chan_index);
 
@@ -1726,32 +1582,8 @@ exec_kilp(struct tgsi_exec_machine *mach,
 {
    uint kilmask; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
 
-   if (inst->InstructionExtNv.CondFlowEnable) {
-      uint swizzle[4];
-      uint chan_index;
-
-      kilmask = 0x0;
-
-      swizzle[0] = inst->InstructionExtNv.CondSwizzleX;
-      swizzle[1] = inst->InstructionExtNv.CondSwizzleY;
-      swizzle[2] = inst->InstructionExtNv.CondSwizzleZ;
-      swizzle[3] = inst->InstructionExtNv.CondSwizzleW;
-
-      for (chan_index = 0; chan_index < 4; chan_index++)
-      {
-         uint i;
-
-         for (i = 0; i < 4; i++) {
-            /* TODO: evaluate the condition code */
-            if (0)
-               kilmask |= 1 << i;
-         }
-      }
-   }
-   else {
-      /* "unconditional" kil */
-      kilmask = mach->ExecMask;
-   }
+   /* "unconditional" kil */
+   kilmask = mach->ExecMask;
    mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
 }
 
@@ -1979,8 +1811,8 @@ exec_declaration(
             break;
 
          default:
-            eval = NULL;
             assert( 0 );
+            return;
          }
 
          if( mask == TGSI_WRITEMASK_XYZW ) {
@@ -2029,10 +1861,23 @@ exec_instruction(
       break;
 
    case TGSI_OPCODE_MOV:
-   case TGSI_OPCODE_SWZ:
-      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
-         FETCH( &r[0], 0, chan_index );
-         STORE( &r[0], 0, chan_index );
+      if (inst->Flags & SOA_DEPENDENCY_FLAG) {
+         /* Do all fetches into temp regs, then do all stores to avoid
+          * intermediate/accidental clobbering.  This could be done all the
+          * time for MOV but for other instructions we'll need more temps...
+          */
+         FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+            FETCH( &r[chan_index], 0, chan_index );
+         }
+         FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+            STORE( &r[chan_index], 0, chan_index );
+         }
+      }
+      else {
+         FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+            FETCH( &r[0], 0, chan_index );
+            STORE( &r[0], 0, chan_index );
+         }
       }
       break;
 
@@ -2313,16 +2158,6 @@ exec_instruction(
       }
       break;
 
-   case TGSI_OPCODE_CND0:
-      FOR_EACH_ENABLED_CHANNEL(*inst, chan_index) {
-         FETCH(&r[0], 0, chan_index);
-         FETCH(&r[1], 1, chan_index);
-         FETCH(&r[2], 2, chan_index);
-         micro_le(&r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &r[2], &r[0], &r[1]);
-         STORE(&r[0], 0, chan_index);
-      }
-      break;
-
    case TGSI_OPCODE_DP2A:
       FETCH( &r[0], 0, CHAN_X );
       FETCH( &r[1], 1, CHAN_X );
@@ -2750,19 +2585,32 @@ exec_instruction(
       if (mach->ExecMask) {
          /* do the call */
 
-         /* push the Cond, Loop, Cont stacks */
+         /* First, record the depths of the execution stacks.
+          * This is important for deeply nested/looped return statements.
+          * We have to unwind the stacks by the correct amount.  For a
+          * real code generator, we could determine the number of entries
+          * to pop off each stack with simple static analysis and avoid
+          * implementing this data structure at run time.
+          */
+         mach->CallStack[mach->CallStackTop].CondStackTop = mach->CondStackTop;
+         mach->CallStack[mach->CallStackTop].LoopStackTop = mach->LoopStackTop;
+         mach->CallStack[mach->CallStackTop].ContStackTop = mach->ContStackTop;
+         /* note that PC was already incremented above */
+         mach->CallStack[mach->CallStackTop].ReturnAddr = *pc;
+
+         mach->CallStackTop++;
+
+         /* Second, push the Cond, Loop, Cont, Func stacks */
          assert(mach->CondStackTop < TGSI_EXEC_MAX_COND_NESTING);
          mach->CondStack[mach->CondStackTop++] = mach->CondMask;
          assert(mach->LoopStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
          mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
          assert(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
          mach->ContStack[mach->ContStackTop++] = mach->ContMask;
-
          assert(mach->FuncStackTop < TGSI_EXEC_MAX_CALL_NESTING);
          mach->FuncStack[mach->FuncStackTop++] = mach->FuncMask;
 
-         /* note that PC was already incremented above */
-         mach->CallStack[mach->CallStackTop++] = *pc;
+         /* Finally, jump to the subroutine */
          *pc = inst->InstructionExtLabel.Label;
       }
       break;
@@ -2779,18 +2627,24 @@ exec_instruction(
             *pc = -1;
             return;
          }
-         *pc = mach->CallStack[--mach->CallStackTop];
 
-         /* pop the Cond, Loop, Cont stacks */
-         assert(mach->CondStackTop > 0);
-         mach->CondMask = mach->CondStack[--mach->CondStackTop];
-         assert(mach->LoopStackTop > 0);
-         mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
-         assert(mach->ContStackTop > 0);
-         mach->ContMask = mach->ContStack[--mach->ContStackTop];
+         assert(mach->CallStackTop > 0);
+         mach->CallStackTop--;
+
+         mach->CondStackTop = mach->CallStack[mach->CallStackTop].CondStackTop;
+         mach->CondMask = mach->CondStack[mach->CondStackTop];
+
+         mach->LoopStackTop = mach->CallStack[mach->CallStackTop].LoopStackTop;
+         mach->LoopMask = mach->LoopStack[mach->LoopStackTop];
+
+         mach->ContStackTop = mach->CallStack[mach->CallStackTop].ContStackTop;
+         mach->ContMask = mach->ContStack[mach->ContStackTop];
+
          assert(mach->FuncStackTop > 0);
          mach->FuncMask = mach->FuncStack[--mach->FuncStackTop];
 
+         *pc = mach->CallStack[mach->CallStackTop].ReturnAddr;
+
          UPDATE_EXEC_MASK(mach);
       }
       break;
@@ -3088,6 +2942,12 @@ exec_instruction(
       break;
 
    case TGSI_OPCODE_BGNFOR:
+      assert(mach->LoopCounterStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+      for (chan_index = 0; chan_index < 3; chan_index++) {
+         FETCH( &mach->LoopCounterStack[mach->LoopCounterStackTop].xyzw[chan_index], 0, chan_index );
+      }
+      STORE( &mach->LoopCounterStack[mach->LoopCounterStackTop].xyzw[CHAN_Y], 0, CHAN_X );
+      ++mach->LoopCounterStackTop;
       /* fall-through (for now) */
    case TGSI_OPCODE_BGNLOOP:
       /* push LoopMask and ContMasks */
@@ -3095,10 +2955,58 @@ exec_instruction(
       mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
       assert(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
       mach->ContStack[mach->ContStackTop++] = mach->ContMask;
+      assert(mach->LoopLabelStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+      mach->LoopLabelStack[mach->LoopLabelStackTop++] = *pc - 1;
       break;
 
    case TGSI_OPCODE_ENDFOR:
-      /* fall-through (for now at least) */
+      assert(mach->LoopCounterStackTop > 0);
+      micro_sub( &mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X], 
+                 &mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X],
+                 &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+      /* update LoopMask */
+      if( mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X].f[0] <= 0) {
+         mach->LoopMask &= ~0x1;
+      }
+      if( mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X].f[1] <= 0 ) {
+         mach->LoopMask &= ~0x2;
+      }
+      if( mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X].f[2] <= 0 ) {
+         mach->LoopMask &= ~0x4;
+      }
+      if( mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_X].f[3] <= 0 ) {
+         mach->LoopMask &= ~0x8;
+      }
+      micro_add( &mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_Y], 
+                 &mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_Y], 
+                 &mach->LoopCounterStack[mach->LoopCounterStackTop - 1].xyzw[CHAN_Z]);
+      assert(mach->LoopLabelStackTop > 0);
+      inst = mach->Instructions + mach->LoopLabelStack[mach->LoopLabelStackTop - 1];
+      STORE( &mach->LoopCounterStack[mach->LoopCounterStackTop].xyzw[CHAN_Y], 0, CHAN_X );
+      /* Restore ContMask, but don't pop */
+      assert(mach->ContStackTop > 0);
+      mach->ContMask = mach->ContStack[mach->ContStackTop - 1];
+      UPDATE_EXEC_MASK(mach);
+      if (mach->ExecMask) {
+         /* repeat loop: jump to instruction just past BGNLOOP */
+         assert(mach->LoopLabelStackTop > 0);
+         *pc = mach->LoopLabelStack[mach->LoopLabelStackTop - 1] + 1;
+      }
+      else {
+         /* exit loop: pop LoopMask */
+         assert(mach->LoopStackTop > 0);
+         mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
+         /* pop ContMask */
+         assert(mach->ContStackTop > 0);
+         mach->ContMask = mach->ContStack[--mach->ContStackTop];
+         assert(mach->LoopLabelStackTop > 0);
+         --mach->LoopLabelStackTop;
+         assert(mach->LoopCounterStackTop > 0);
+         --mach->LoopCounterStackTop;
+      }
+      UPDATE_EXEC_MASK(mach);
+      break;
+      
    case TGSI_OPCODE_ENDLOOP:
       /* Restore ContMask, but don't pop */
       assert(mach->ContStackTop > 0);
@@ -3106,7 +3014,8 @@ exec_instruction(
       UPDATE_EXEC_MASK(mach);
       if (mach->ExecMask) {
          /* repeat loop: jump to instruction just past BGNLOOP */
-         *pc = inst->InstructionExtLabel.Label + 1;
+         assert(mach->LoopLabelStackTop > 0);
+         *pc = mach->LoopLabelStack[mach->LoopLabelStackTop - 1] + 1;
       }
       else {
          /* exit loop: pop LoopMask */
@@ -3115,6 +3024,8 @@ exec_instruction(
          /* pop ContMask */
          assert(mach->ContStackTop > 0);
          mach->ContMask = mach->ContStack[--mach->ContStackTop];
+         assert(mach->LoopLabelStackTop > 0);
+         --mach->LoopLabelStackTop;
       }
       UPDATE_EXEC_MASK(mach);
       break;
@@ -3141,22 +3052,6 @@ exec_instruction(
       /* no-op */
       break;
 
-   case TGSI_OPCODE_NOISE1:
-      assert( 0 );
-      break;
-
-   case TGSI_OPCODE_NOISE2:
-      assert( 0 );
-      break;
-
-   case TGSI_OPCODE_NOISE3:
-      assert( 0 );
-      break;
-
-   case TGSI_OPCODE_NOISE4:
-      assert( 0 );
-      break;
-
    case TGSI_OPCODE_NOP:
       break;
 
@@ -3182,7 +3077,6 @@ tgsi_exec_machine_run( struct tgsi_exec_machine *mach )
    mach->FuncMask = 0xf;
    mach->ExecMask = 0xf;
 
-   mach->CondStackTop = 0; /* temporarily subvert this assertion */
    assert(mach->CondStackTop == 0);
    assert(mach->LoopStackTop == 0);
    assert(mach->ContStackTop == 0);