[PATCH] D31965: [SLP] Enable 64-bit wide vectorization for Cyclone

Matt Arsenault via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Wed May 3 12:30:38 PDT 2017


arsenm added inline comments.


================
Comment at: include/llvm/Analysis/TargetTransformInfoImpl.h:306
 
+  unsigned getMinVectorRegisterBitWidth() { return 128; }
+
----------------
anemet wrote:
> sdardis wrote:
> > rengolin wrote:
> > > Is this value really the best default to all targets?
> > My quick survey of vector register widths suggests this is double the minimum.
> > 
> > SPARC's VIS extension uses the double precision floating point register set (64 bits wide) , as does Intel's MMX, MIPS' MIPS-3D (though currently unimplemented in LLVM).
> > 
> > The S/390 vector registers appear to be 128 bits, like the basic Intel SSE, MIPS MSA, ARM NEON, PowerPC Altivec.
> This does not change change the default from SLP.  It just brings it to TTI so that targets can change it as they see fit after careful benchmarking (it will need careful benchmarking!).  It is *not* the goal of this patch to find a new proper default across targets.
I had D32714 because AMDGPU wants 32, but this is probably better


https://reviews.llvm.org/D31965





More information about the llvm-commits mailing list