On Fri, Mar 28, 2014 at 11:11:58AM +0000, pins...@gmail.com wrote: > > On Mar 28, 2014, at 2:12 AM, James Greenhalgh <james.greenha...@arm.com> > > wrote: > > Hi, > > > > There is no way to perform scalar addition in the vector register file, > > but with the RTX costs in place we start rewriting (x << 1) to (x + x) > > on almost all cores. The code which makes this decision has no idea that we > > will end up doing this (it happens well before reload) and so we end up with > > very ugly code generation in the case where addition was selected, but > > we are operating in vector registers. > > > > This patch relies on the same gimmick we are already using to allow > > shifts on 32-bit scalars in the vector register file - Use a vector 32x2 > > operation instead, knowing that we can safely ignore the top bits. > > > > This restores some normality to scalar_shift_1.c, however the test > > that we generate a left shift by one is clearly bogus, so remove that. > > > > This patch is pretty ugly, but it does generate superficially better > > looking code for this testcase. > > > > Tested on aarch64-none-elf with no issues. > > > > OK for stage 1? > > It seems we should also discourage the neon alternatives as there might be > extra movement between the two register sets which we don't want.
I see your point, but we've tried to avoid doing that elsewhere in the AArch64 backend. Our argument has been that strictly speaking, it isn't that the alternative is expensive, it is the movement between the register sets. We do model that elsewhere, and the register allocator should already be trying to avoid unneccesary moves between register classes. If those mechanisms are broken, we should fix them - in that case fixing this by discouraging valid alternatives would seem to be gaffer-taping over the real problem. Thanks, James > > Thanks, > Andrew > > > > > Thanks, > > James > > > > --- > > gcc/ > > > > 2014-03-27 James Greenhalgh <james.greenha...@arm.com> > > > > * config/aarch64/aarch64.md (*addsi3_aarch64): Add alternative in > > vector registers. > > > > gcc/testsuite/ > > 2014-03-27 James Greenhalgh <james.greenha...@arm.com> > > > > * gcc.target/aarch64/scalar_shift_1.c: Fix expected assembler. > > <0001-AArch64-Implement-ADD-in-vector-registers-for-32-bit.patch> >
