nir: Make vec-to-movs handle src/dest aliasing.

It now emits vector MOVs instead of a series of individual MOVs, which
should be useful to any vector backends.  This pushes the problem of
src/dest aliasing of channels on a scalar chip to the backend, but if
there are any vector operations in your shader then you needed to be
handling this already.

Fixes fs-swap-problem with my scalarizing patches.

v2: Rename to insert_mov(), and add a comment about what it does.
v3: Rewrite the comment.

Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v3)

src/glsl/nir/nir_lower_vec_to_movs.c

@@ -32,6 +32,57 @@
  * moves with partial writes.
  */
 
+static bool
+src_matches_dest_reg(nir_dest *dest, nir_src *src)
+{
+   if (dest->is_ssa || src->is_ssa)
+      return false;
+
+   return (dest->reg.reg == src->reg.reg &&
+           dest->reg.base_offset == src->reg.base_offset &&
+           !dest->reg.indirect &&
+           !src->reg.indirect);
+}
+
+/**
+ * For a given starting writemask channel and corresponding source index in
+ * the vec instruction, insert a MOV to the vec instruction's dest of all the
+ * writemask channels that get read from the same src reg.
+ *
+ * Returns the writemask of our MOV, so the parent loop calling this knows
+ * which ones have been processed.
+ */
+static unsigned
+insert_mov(nir_alu_instr *vec, unsigned start_channel,
+            unsigned start_src_idx, void *mem_ctx)
+{
+   unsigned src_idx = start_src_idx;
+   assert(src_idx < nir_op_infos[vec->op].num_inputs);
+
+   nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov);
+   nir_alu_src_copy(&mov->src[0], &vec->src[src_idx], mem_ctx);
+   nir_alu_dest_copy(&mov->dest, &vec->dest, mem_ctx);
+
+   mov->dest.write_mask = (1u << start_channel);
+   mov->src[0].swizzle[start_channel] = vec->src[src_idx].swizzle[0];
+   src_idx++;
+
+   for (unsigned i = start_channel + 1; i < 4; i++) {
+      if (!(vec->dest.write_mask & (1 << i)))
+         continue;
+
+      if (nir_srcs_equal(vec->src[src_idx].src, vec->src[start_src_idx].src)) {
+         mov->dest.write_mask |= (1 << i);
+         mov->src[0].swizzle[i] = vec->src[src_idx].swizzle[0];
+      }
+      src_idx++;
+   }
+
+   nir_instr_insert_before(&vec->instr, &mov->instr);
+
+   return mov->dest.write_mask;
+}
+
 static bool
 lower_vec_to_movs_block(nir_block *block, void *mem_ctx)
 {
@@ -50,22 +101,33 @@ lower_vec_to_movs_block(nir_block *block, void *mem_ctx)
          continue; /* The loop */
       }
 
+      /* Since we insert multiple MOVs, we have to be non-SSA. */
+      assert(!vec->dest.dest.is_ssa);
+
+      unsigned finished_write_mask = 0;
+
+      /* First, emit a MOV for all the src channels that are in the
+       * destination reg, in case other values we're populating in the dest
+       * might overwrite them.
+       */
       for (unsigned i = 0, src_idx = 0; i < 4; i++) {
          if (!(vec->dest.write_mask & (1 << i)))
             continue;
 
-         assert(src_idx < nir_op_infos[vec->op].num_inputs);
+         if (src_matches_dest_reg(&vec->dest.dest, &vec->src[src_idx].src)) {
+            finished_write_mask |= insert_mov(vec, i, src_idx, mem_ctx);
+            break;
+         }
+         src_idx++;
+      }
 
-         nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov);
+      /* Now, emit MOVs for all the other src channels. */
-         nir_alu_src_copy(&mov->src[0], &vec->src[src_idx], mem_ctx);
+      for (unsigned i = 0, src_idx = 0; i < 4; i++) {
+         if (!(vec->dest.write_mask & (1 << i)))
+            continue;
 
-         /* We only care about the one swizzle */
+         if (!(finished_write_mask & (1 << i)))
-         mov->src[0].swizzle[i] = vec->src[src_idx].swizzle[0];
+            finished_write_mask |= insert_mov(vec, i, src_idx, mem_ctx);
-
-         nir_alu_dest_copy(&mov->dest, &vec->dest, mem_ctx);
-         mov->dest.write_mask = (1u << i);
-
-         nir_instr_insert_before(&vec->instr, &mov->instr);
 
          src_idx++;
       }