On 12 June 2015 at 21:43, Jim Wilson <jim.wil...@linaro.org> wrote: > We have 5 patterns that can emit the movi instruction. These patterns > map it to 4 different type attributes. The mov<mode>_aarch64 pattern > uses mov_imm. The movdi_aarch64 pattern uses fmov. The movtf_aarch64 > pattern uses fconstd. And the two aarch64_simd_mov<mode> patterns for > VD and VQ use neon_move. Bitwise identical instructions should always > map to the same attribute type, so we need to change these patterns to > agree on the right attribute. movi is an integer simd instruction, so > neon_move is the only choice that makes sense. The following patch > corrects the first 3 patterns to use neon_move like the last two. > > We could optionally create a new type attribute, e.g. neon_move_imm. > I can do that if people think it would be better. > > This patch was tested with a make bootstrap and make check on an APM > box running Ubuntu 14.04. > > FYI This patch overlaps with my movtf-zero patch which is still > waiting review, but the overlap is trivial to resolve so this should > not be a problem.
The movdi* and movtf* hunks look fine. The mov<mode>_aarch64 pattern calls aarch64_output_scalar_simd_mov_immediate which can emit either mvni, movi, with or without msl and lsl. In the case of the plain movi neon_move looks sensible, the other possbile outputs should ideally be represented by logic_immediate and logic_shift_immediate. Using neon_move is a step in the right direction. OK /Marcus
