agx: Implement nir_op_pack_64_2x32_split

This maps to a collect where the dest size is 64 and the src size is 32.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19971>

src/asahi/compiler/agx_compile.c

@@ -1088,6 +1088,12 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
       return agx_convert_to(b, dst, agx_immediate(mode), s0, AGX_ROUND_RTE);
    }
 
+   case nir_op_pack_64_2x32_split:
+   {
+      agx_index idx[] = { s0, s1 };
+      return agx_emit_collect_to(b, dst, 2, idx);
+   }
+
    /* Split a 64-bit word into 32-bit parts. Do not use null destinations to
     * let us CSE (and coalesce) the splits when both x and y are split.
     */

src/asahi/compiler/agx_register_allocate.c

@@ -61,7 +61,7 @@ agx_write_registers(agx_instr *I, unsigned d)
    case AGX_OPCODE_LDCF:
       return 6;
    case AGX_OPCODE_COLLECT:
-      return I->nr_srcs * size;
+      return I->nr_srcs * agx_size_align_16(I->src[0].size);
    default:
       return size;
    }
@@ -178,6 +178,10 @@ pick_regs(struct ra_ctx *rctx, agx_instr *I, unsigned d)
          if (base >= rctx->bound || (base + count) > rctx->bound)
             continue;
 
+         /* Unaligned destinations can happen when dest size > src size */
+         if (base % align)
+            continue;
+
          if (!BITSET_TEST_RANGE(rctx->used_regs, base, base + count - 1))
             return base;
       }
@@ -424,7 +428,7 @@ agx_ra(agx_context *ctx)
       if (ins->op == AGX_OPCODE_COLLECT) {
          assert(ins->dest[0].type == AGX_INDEX_REGISTER);
          unsigned base = ins->dest[0].value;
-         unsigned width = agx_size_align_16(ins->dest[0].size);
+         unsigned width = agx_size_align_16(ins->src[0].size);
 
          struct agx_copy *copies = alloca(sizeof(copies[0]) * ins->nr_srcs);
          unsigned n = 0;
@@ -432,7 +436,7 @@ agx_ra(agx_context *ctx)
          /* Move the sources */
          agx_foreach_src(ins, i) {
             if (agx_is_null(ins->src[i])) continue;
-            assert(ins->src[i].size == ins->dest[0].size);
+            assert(ins->src[i].size == ins->src[0].size);
 
             copies[n++] = (struct agx_copy) {
                .dest = base + (i * width),