amdgcn: Improve FP division accuracy

amdgcn: Improve FP division accuracy

GCN has a reciprocal-approximation instruction but no
hardware divide. This patch adjusts the open-coded reciprocal
approximation/Newton-Raphson refinement steps to use fused multiply-add
instructions as is necessary to obtain a properly-rounded result, and
adds further refinement steps to correctly round the full division result.

The patterns in question are still guarded by a flag_reciprocal_math
condition, and do not yet support denormals.

Backport from mainline:

2021-01-13  Julian Brown  <julian@codesourcery.com>

gcc/
	* config/gcn/gcn-valu.md (recip<mode>2<exec>, recip<mode>2): Use unspec
	for reciprocal-approximation instructions.
	(div<mode>3): Use fused multiply-accumulate operations for reciprocal
	refinement and division result.
	* config/gcn/gcn.md (UNSPEC_RCP): New unspec constant.

gcc/testsuite/
	* gcc.target/gcn/fpdiv.c: New test.

(cherry picked from commit c8812bac8ee39f73ea881e4f6260acf5590b4190)