Vectorization factor limitation in Loop Vectorizer

Project Infrastructure LLVM Dev List Archives
1

Hi Nadav/Devs

I am exploring Loop Vectorizer to vectorize i8 scalar operations into 8xi8 vector operation.

I was expecting the Loop Vectorizer to analyze the profitability for vectorization factor(VF) of 8,

However it is not doing so due to the widest type calculation done for the blocks inside the loop.

May be I am missing something, however, I am curious to know why Loop Vectorizer limits the

profitability check to widest type and not allowing for other narrower type?

Regards,

Shahid

2

Hi Shahid,

Hi Nadav/Devs

I am exploring Loop Vectorizer to vectorize i8 scalar operations into 8xi8 vector operation.

I was expecting the Loop Vectorizer to analyze the profitability for vectorization factor(VF) of 8,
However it is not doing so due to the widest type calculation done for the blocks inside the loop.

May be I am missing something, however, I am curious to know why Loop Vectorizer limits the
profitability check to widest type and not allowing for other narrower type?

The vectorizer stops the search of profitable vectorization factors at the widest type because higher vectorization factors would require the compiler to split the vectorized value into multiple registers. The vectorizer’s cost model first tries to optimize for SIMD instruction utilization. Later, we optimize for ILP by doubling the vectorization factor (we call it “interleave”) and exposing ILP.

Thanks,
Nadav

3

So IMO, if we modify the VF calculation for targets/subtargets using TTI where higher VF is supported

The vectorizer’s scope will become wider.

Did/do you foresee any issue with this?

Thanks,

Shahid

4

Hi shahid,

Having looked into this in the past, we’re not taking into account the fact that integer promotions and truncations will disappear during code generation. The canonicalisation phase likes to create i32 and greater type arithmetic, which creates extensions and promotions . However of course the vectoriser can use i8 arithmetic I vectors, so those should go away.

Often this is the cause of insufficient vectorisation factors being picked, and we should fix it.

James

5

Hi James,

Thanks for the clarification. I hope to come back once I have a fix.

Pls let me know if you have any suggestion.

Thanks,

Shahid