The 11/17/2017 22:06, James Greenhalgh wrote: > On Tue, Nov 14, 2017 at 03:54:56PM +0000, Tamar Christina wrote: > > Hi All, > > > > Dot Product is intended to only be available for Armv8.2-a and newer. > > While this restriction is reflected in the intrinsics, the patterns > > themselves were missing the Armv8.2-a bit. > > > > This means that using -march=armv8.1-a+dotprod incorrectly got the > > auto-vectorizer to generate dot product instructions. > > > > Regtested on aarch64-none-elf and no issues. > > > > Ok for trunk? > > What is the design here? That -march=armv8.1-a+dotprod is valid, but that > the +dotprod is silently ignored? Is that really less surprising than > a hard error?
As far as I am aware, AArch64 unlike ARM does not provide a facility to reject invalid combinations of architecture and extensions. The argument parsers don't specify any relationship between architectures and extensions. This is why -march=armv8-a+fp16 is also accepted silently. This patch merely makes it not generate code that can't possibly run. > > How does this interact with -march=native ? If I remember right, we don't > get any good indication from the kernel that we are Armv8.2-A rather than > Armv8.1-A? Maybe I'm misremembering. As far as I am aware, if you have a kernel which has asimddp support, and you compile with -march=armv8.1-a then it's equivalent to specifying -march=armv8.1-a+dotprod, as in, the bit will be enabled, but the feature won't, as it's gated on Armv8.2-a (with this patch). > > Thanks, > James > > > > > Thanks, > > Tamar > > > > gcc/ > > 2017-11-14 Tamar Christina <tamar.christ...@arm.com> > > > > * config/aarch64/aarch64.h (TARGET_DOTPROD): Add AARCH64_ISA_V8_2. > > > > gcc/testsuite/ > > 2017-11-14 Tamar Christina <tamar.christ...@arm.com> > > > > * gcc.target/aarch64/advsimd-intrinsics/vect-dot-s8_1.c: New. > > * gcc.target/aarch64/advsimd-intrinsics/vect-dot-s8_2.c: New. --
