Merge master and fix conflicts

4 files changed, 197 insertions, 27 deletions

diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index e8bd7cda3b..aba7a3f937 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -53,6 +53,7 @@
 #include "pipe/p_compiler.h"
 #include "pipe/p_state.h"
 #include "pipe/p_shader_tokens.h"
+#include "tgsi/tgsi_dump.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
 #include "tgsi_exec.h"
@@ -169,6 +170,56 @@ print_temp(const struct tgsi_exec_machine *mach, uint index)
 #endif
 
 
+/**
+ * Check if there's a potential src/dst register data dependency when
+ * using SOA execution.
+ * Example:
+ *   MOV T, T.yxwz;
+ * This would expand into:
+ *   MOV t0, t1;
+ *   MOV t1, t0;
+ *   MOV t2, t3;
+ *   MOV t3, t2;
+ * The second instruction will have the wrong value for t0 if executed as-is.
+ */
+static boolean
+tgsi_check_soa_dependencies(const struct tgsi_full_instruction *inst)
+{
+   uint i, chan;
+
+   uint writemask = inst->FullDstRegisters[0].DstRegister.WriteMask;
+   if (writemask == TGSI_WRITEMASK_X ||
+       writemask == TGSI_WRITEMASK_Y ||
+       writemask == TGSI_WRITEMASK_Z ||
+       writemask == TGSI_WRITEMASK_W ||
+       writemask == TGSI_WRITEMASK_NONE) {
+      /* no chance of data dependency */
+      return FALSE;
+   }
+
+   /* loop over src regs */
+   for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
+      if ((inst->FullSrcRegisters[i].SrcRegister.File ==
+           inst->FullDstRegisters[0].DstRegister.File) &&
+          (inst->FullSrcRegisters[i].SrcRegister.Index ==
+           inst->FullDstRegisters[0].DstRegister.Index)) {
+         /* loop over dest channels */
+         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))) {
+               return TRUE;
+            }
+
+            channelsWritten |= (1 << chan);
+         }
+      }
+   }
+   return FALSE;
+}
+
 
 /**
  * Initialize machine state by expanding tokens to full instructions,
@@ -280,6 +331,17 @@ tgsi_exec_machine_bind_shader(
          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);
+         }
+#else
+         (void) tgsi_check_soa_dependencies;
+#endif
+
          numInstructions++;
          break;
 
diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.h b/src/gallium/auxiliary/tgsi/tgsi_exec.h
index 0b4b2a6fb6..da22baad3e 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.h
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.h
@@ -34,7 +34,7 @@
 extern "C" {
 #endif
 
-#define MAX_LABELS 1024
+#define MAX_LABELS (4 * 1024)  /**< basically, max instructions */
 
 #define NUM_CHANNELS 4  /* R,G,B,A */
 #define QUAD_SIZE    4  /* 4 pixel/quad */
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index d70bcd03c5..ba2bfdef06 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -1466,6 +1466,31 @@ emit_cmp(
    }
 }
 
+
+/**
+ * Check if inst src/dest regs use indirect addressing into temporary
+ * register file.
+ */
+static boolean
+indirect_temp_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->FullSrcRegisters[i];
+      if (reg->SrcRegister.File == TGSI_FILE_TEMPORARY &&
+          reg->SrcRegister.Indirect)
+         return TRUE;
+   }
+   for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
+      const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[i];
+      if (reg->DstRegister.File == TGSI_FILE_TEMPORARY &&
+          reg->DstRegister.Indirect)
+         return TRUE;
+   }
+   return FALSE;
+}
+
+
 static int
 emit_instruction(
    struct x86_function *func,
@@ -1473,10 +1498,15 @@ emit_instruction(
 {
    unsigned chan_index;
 
+   /* we can't handle indirect addressing into temp register file yet */
+   if (indirect_temp_reference(inst))
+      return FALSE;
+
    switch (inst->Instruction.Opcode) {
    case TGSI_OPCODE_ARL:
       FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
          FETCH( func, *inst, 0, 0, chan_index );
+         emit_flr(func, 0, 0);
          emit_f2it( func, 0 );
          STORE( func, *inst, 0, 0, chan_index );
       }
@@ -1553,7 +1583,7 @@ emit_instruction(
                func,
                make_xmm( 2 ),
                make_xmm( 0 ),
-               cc_LessThanEqual );
+               cc_LessThan );
             sse_andps(
                func,
                make_xmm( 2 ),
@@ -2177,32 +2207,83 @@ emit_instruction(
       /* 3 or 4-component normalization */
       {
          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
-         /* note: cannot use xmm regs 2/3 here (see emit_rsqrt() above) */
-         FETCH( func, *inst, 4, 0, CHAN_X );    /* xmm4 = src[0].x */
-         FETCH( func, *inst, 5, 0, CHAN_Y );    /* xmm5 = src[0].y */
-         FETCH( func, *inst, 6, 0, CHAN_Z );    /* xmm6 = src[0].z */
-         if (dims == 4) {
-            FETCH( func, *inst, 7, 0, CHAN_W ); /* xmm7 = src[0].w */
-         }
-         emit_MOV( func, 0, 4 );                /* xmm0 = xmm3 */
-         emit_mul( func, 0, 4 );                /* xmm0 *= xmm3 */
-         emit_MOV( func, 1, 5 );                /* xmm1 = xmm4 */
-         emit_mul( func, 1, 5 );                /* xmm1 *= xmm4 */
-         emit_add( func, 0, 1 );                /* xmm0 += xmm1 */
-         emit_MOV( func, 1, 6 );                /* xmm1 = xmm5 */
-         emit_mul( func, 1, 6 );                /* xmm1 *= xmm5 */
-         emit_add( func, 0, 1 );                /* xmm0 += xmm1 */
-         if (dims == 4) {
-            emit_MOV( func, 1, 7 );             /* xmm1 = xmm7 */
-            emit_mul( func, 1, 7 );             /* xmm1 *= xmm7 */
-            emit_add( func, 0, 0 );             /* xmm0 += xmm1 */
-         }
-         emit_rsqrt( func, 1, 0 );              /* xmm1 = 1/sqrt(xmm0) */
-         FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
-            if (chan_index < dims) {
-               emit_mul( func, 4+chan_index, 1); /* xmm[4+ch] *= xmm1 */
-               STORE( func, *inst, 4+chan_index, 0, chan_index );
+
+         if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) ||
+             IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) ||
+             IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z) ||
+             (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 4)) {
+
+            /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
+
+            /* xmm4 = src.x */
+            /* xmm0 = src.x * src.x */
+            FETCH(func, *inst, 0, 0, CHAN_X);
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
+               emit_MOV(func, 4, 0);
+            }
+            emit_mul(func, 0, 0);
+
+            /* xmm5 = src.y */
+            /* xmm0 = xmm0 + src.y * src.y */
+            FETCH(func, *inst, 1, 0, CHAN_Y);
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
+               emit_MOV(func, 5, 1);
+            }
+            emit_mul(func, 1, 1);
+            emit_add(func, 0, 1);
+
+            /* xmm6 = src.z */
+            /* xmm0 = xmm0 + src.z * src.z */
+            FETCH(func, *inst, 1, 0, CHAN_Z);
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
+               emit_MOV(func, 6, 1);
+            }
+            emit_mul(func, 1, 1);
+            emit_add(func, 0, 1);
+
+            if (dims == 4) {
+               /* xmm7 = src.w */
+               /* xmm0 = xmm0 + src.w * src.w */
+               FETCH(func, *inst, 1, 0, CHAN_W);
+               if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) {
+                  emit_MOV(func, 7, 1);
+               }
+               emit_mul(func, 1, 1);
+               emit_add(func, 0, 1);
             }
+
+            /* xmm1 = 1 / sqrt(xmm0) */
+            emit_rsqrt(func, 1, 0);
+
+            /* dst.x = xmm1 * src.x */
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
+               emit_mul(func, 4, 1);
+               STORE(func, *inst, 4, 0, CHAN_X);
+            }
+
+            /* dst.y = xmm1 * src.y */
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
+               emit_mul(func, 5, 1);
+               STORE(func, *inst, 5, 0, CHAN_Y);
+            }
+
+            /* dst.z = xmm1 * src.z */
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
+               emit_mul(func, 6, 1);
+               STORE(func, *inst, 6, 0, CHAN_Z);
+            }
+
+            /* dst.w = xmm1 * src.w */
+            if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) && dims == 4) {
+               emit_mul(func, 7, 1);
+               STORE(func, *inst, 7, 0, CHAN_W);
+            }
+         }
+
+         /* dst0.w = 1.0 */
+         if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 3) {
+            emit_tempf(func, 0, TEMP_ONE_I, TEMP_ONE_C);
+            STORE(func, *inst, 0, 0, CHAN_W);
          }
       }
       break;
diff --git a/src/gallium/auxiliary/tgsi/tgsi_transform.c b/src/gallium/auxiliary/tgsi/tgsi_transform.c
index 062c1be938..bc9c18fd4a 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_transform.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_transform.c
@@ -198,3 +198,30 @@ tgsi_transform_shader(const struct tgsi_token *tokens_in,
 
    return ctx->ti;
 }
+
+
+#include "tgsi_text.h"
+
+extern int tgsi_transform_foo( struct tgsi_token *tokens_out,
+                               uint max_tokens_out );
+
+/* This function exists only so that tgsi_text_translate() doesn't get
+ * magic-ed out of the libtgsi.a archive by the build system.  Don't
+ * remove unless you know this has been fixed - check on mingw/scons
+ * builds as well.
+ */
+int
+tgsi_transform_foo( struct tgsi_token *tokens_out,
+                    uint max_tokens_out )
+{
+   const char *text = 
+      "FRAG1.1\n"
+      "DCL IN[0], COLOR, CONSTANT\n"
+      "DCL OUT[0], COLOR\n"
+      "  0: MOV OUT[0], IN[0]\n"
+      "  1: END";
+        
+   return tgsi_text_translate( text,
+                               tokens_out,
+                               max_tokens_out );
+}