My previous change to the Thumb-2 movsi patterns caused a codesize regression
with -Os in large functions.  Fix this by using the literal pool offset of the
16-bit literal load so that the literal pool is dumped earlier, reducing the
number of 32-bit literal loads.

Bootstrap & regress OK on arm-none-linux-gnueabihf --with-cpu=cortex-a57

    gcc/
	* config/arm/thumb2.md (thumb2_movsi_insn): Adjust literal offset.
	* config/arm/vfp.md (thumb2_movsi_vfp): Likewise.

From-SVN: r273911