On Thu, 2019-04-11 at 09:59 +0100, Richard Earnshaw (lists) wrote: > > > > > 2018-04-10 Steve Ellcey <sell...@marvell.com> > > > > PR rtl-optimization/87763 > > * config/aarch64/aarch64-protos.h > > (aarch64_masks_and_shift_for_bfi_p): > > New prototype. > > * config/aarch64/aarch64.c (aarch64_masks_and_shift_for_bfi_p): > > New function. > > * config/aarch64/aarch64.md (*aarch64_bfi<GPI:mode>5_shift): > > New instruction. > > (*aarch64_bfi<GPI:mode>5_shift_alt): Ditto. > > (*aarch64_bfi<GPI:mode>4_noand): Ditto. > > (*aarch64_bfi<GPI:mode>4_noand_alt): Ditto. > > (*aarch64_bfi<GPI:mode>4_noshift): Ditto. > > > > You've removed the ..._noshift_alt variant. That wasn't my intention, > so perhaps you misunderstood what I was trying to say. > > The two versions are both needed, since the register tie is not > orthogonal to the constants used in the masks and canonicalization will > not generate a consistent ordering of the operands.
I started doing this and then convinced myself (perhaps incorrectly) that I didn't need the alt version. Operands 1 and 3 are registers and Operands 2 and 4 are constants, so the only difference is in the call to aarch64_masks_and_shift_for_bfi_p. Given that argument 2 to this call is 0 this call should be the equivelent of ((x & MASK1) | (y & MASK2)) and that should mean that: aarch64_masks_and_shift_for_bfi_p(X,0,Y) == aarch64_masks_and_shift_for_bfi_p(Y,0,X) Maybe I am wrong about that? I will do some expirements. My testing did not find any cases in the testsuite where not having the _alt version resulted in a bfi not being generated. Steve Ellcey sell...@marvell.com > >
