The creation of Neon Instruction

Dear all,

I am trying to study the Loop Vectorization. Neon supports the vectorization, so I want to get the neon instructions.

I tried to use the following command:

clang -fomit-frame-pointer -ggdb -emit-llvm hello.c -c -arch arm -o hello.bc

opt -mem2reg -loop-rotate -loop-simplify -lcssa -loop-vectorize -force-vector-width=8 hello.bc -o new.bc

llc new.bc -march=arm -mattr=+neon -relocation-model=static -o new.s

But I didn’t get the neon instructions like “vadd, vmul” in the new.s. Is there anything wrong with the command?
(clang3.4 on Mac OS X 10.9)

Looking forward to your answers.

Thank you!

Hanbing LI

Hi,

clang -fomit-frame-pointer -ggdb -emit-llvm hello.c -c -arch arm -o hello.bc
opt -mem2reg -loop-rotate -loop-simplify -lcssa -loop-vectorize
-force-vector-width=8 hello.bc -o new.bc

It's a bit difficult to tell without the example (it could just be
that hello.c isn't vectorisable or needs more opt passes to be
vectorised), but the "-arch arm" is almost certainly wrong. That
targets a very old processor (ARM7TDMI in fact), which doesn't have
any NEON.

I use "-arch armv7s" for most of my tests. That targets the Swift CPU,
which is a modern v7 core with NEON and all the other nice features.

Cheers.

Tim.