cell: change spe_complement() to take a src and dst reg, like other instructions

src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c

@@ -623,9 +623,9 @@ spe_splat(struct spe_function *p, unsigned rT, unsigned rA)
 
 
 void
-spe_complement(struct spe_function *p, unsigned rT)
+spe_complement(struct spe_function *p, unsigned rT, unsigned rA)
 {
-   spe_nor(p, rT, rT, rT);
+   spe_nor(p, rT, rA, rA);
 }
 
 
@@ -667,7 +667,8 @@ spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word)
    }
 }
 
-/* For each 32-bit float element of rA and rB, choose the smaller of the
+/**
+ * For each 32-bit float element of rA and rB, choose the smaller of the
  * two, compositing them into the rT register.
  * 
  * The Float Compare Greater Than (fcgt) instruction will put 1s into
@@ -683,7 +684,7 @@ spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word)
  * like "x = min(x, a)", we always allocate a new register to be safe.
  */
 void 
-spe_float_min(struct spe_function *p, unsigned int rT, unsigned int rA, unsigned int rB)
+spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
 {
    unsigned int compare_reg = spe_allocate_available_register(p);
    spe_fcgt(p, compare_reg, rA, rB);
@@ -691,7 +692,8 @@ spe_float_min(struct spe_function *p, unsigned int rT, unsigned int rA, unsigned
    spe_release_register(p, compare_reg);
 }
 
-/* For each 32-bit float element of rA and rB, choose the greater of the
+/**
+ * For each 32-bit float element of rA and rB, choose the greater of the
  * two, compositing them into the rT register.
  * 
  * The logic is similar to that of spe_float_min() above; the only
@@ -699,7 +701,7 @@ spe_float_min(struct spe_function *p, unsigned int rT, unsigned int rA, unsigned
  * so that the larger of the two is selected instead of the smaller.
  */
 void 
-spe_float_max(struct spe_function *p, unsigned int rT, unsigned int rA, unsigned int rB)
+spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
 {
    unsigned int compare_reg = spe_allocate_available_register(p);
    spe_fcgt(p, compare_reg, rA, rB);

src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h

@@ -310,9 +310,9 @@ spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui);
 extern void
 spe_splat(struct spe_function *p, unsigned rT, unsigned rA);
 
-/** Complement/invert all bits in rT. */
+/** rT = complement_all_bits(rA). */
 extern void
-spe_complement(struct spe_function *p, unsigned rT);
+spe_complement(struct spe_function *p, unsigned rT, unsigned rA);
 
 /** rT = rA. */
 extern void

src/gallium/drivers/cell/ppu/cell_gen_fp.c

@@ -924,7 +924,7 @@ emit_IF(struct codegen *gen, const struct tgsi_full_instruction *inst)
    /* tmp = (s1_reg == 0) */
    spe_ceqi(gen->f, tmp_reg, s1_reg, 0);
    /* tmp = !tmp */
-   spe_complement(gen->f, tmp_reg);
+   spe_complement(gen->f, tmp_reg, tmp_reg);
    /* exec_mask = exec_mask & tmp */
    spe_and(gen->f, exec_reg, exec_reg, tmp_reg);
 
@@ -944,7 +944,7 @@ emit_ELSE(struct codegen *gen, const struct tgsi_full_instruction *inst)
    spe_comment(gen->f, -4, "ELSE:");
 
    /* exec_mask = !exec_mask */
-   spe_complement(gen->f, exec_reg);
+   spe_complement(gen->f, exec_reg, exec_reg);
 
    return true;
 }

src/gallium/drivers/cell/ppu/cell_gen_fragment.c

@@ -920,7 +920,7 @@ gen_logicop(const struct pipe_blend_state *blend,
          spe_andc(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg);
          break;
       case PIPE_LOGICOP_COPY_INVERTED: /* ~s */
-         spe_complement(f, fragRGBA_reg);
+         spe_complement(f, fragRGBA_reg, fragRGBA_reg);
          break;
       case PIPE_LOGICOP_AND_REVERSE: /* s & ~d */
          /* andc R, A, B computes R = A & ~B */
@@ -941,7 +941,7 @@ gen_logicop(const struct pipe_blend_state *blend,
          break;
       case PIPE_LOGICOP_EQUIV: /* ~(s ^ d) */
          spe_xor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
-         spe_complement(f, fragRGBA_reg);
+         spe_complement(f, fragRGBA_reg, fragRGBA_reg);
          break;
       case PIPE_LOGICOP_NOOP: /* d */
          spe_move(f, fragRGBA_reg, fbRGBA_reg);