py/asmthumb: Generate proper sequences for large register offsets.

This commit lets the Thumb native emitter generate a proper opcode
sequence when calculating an indexed register offset for load/store
operations with said offset beight both greater than 65535 and not
able to be represented as a shifted 8-bit bitmask.

The original code would assume the scaled index would always fit in 16
bits and silently discard upper bits of the offset.  Now an optimised
constant loading sequence is emitted instead, and the final offset is
also stored in the correct register in all cases.

Signed-off-by: Alessandro Gatti <[email protected]>

Author: agatti

py/asmthumb.c

@@ -450,12 +450,12 @@ static void asm_thumb_add_reg_reg_offset(asm_thumb_t *as, uint reg_dest, uint re
             asm_thumb_lsl_rlo_rlo_i5(as, reg_dest, reg_dest, offset_shift);
             asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_base);
         } else if (reg_dest != reg_base) {
-            asm_thumb_mov_rlo_i16(as, reg_dest, offset << offset_shift);
-            asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_dest);
+            asm_thumb_mov_reg_i32_optimised(as, reg_dest, offset << offset_shift);
+            asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_base);
         } else {
             uint reg_other = reg_dest ^ 7;
             asm_thumb_op16(as, OP_PUSH_RLIST((1 << reg_other)));
-            asm_thumb_mov_rlo_i16(as, reg_other, offset << offset_shift);
+            asm_thumb_mov_reg_i32_optimised(as, reg_other, offset << offset_shift);
             asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_other);
             asm_thumb_op16(as, OP_POP_RLIST((1 << reg_other)));
         }