mbox series

[0/4] aarch64: Improve codegen for dups and constructors

Message ID 95d2de77-5b68-6d0b-ac99-ac1ca28835e2@arm.com
Headers show
Series aarch64: Improve codegen for dups and constructors | expand

Message

Andre Vieira (lists) Aug. 5, 2022, 12:50 p.m. UTC 
Hi,

This patch series is a work in progress towards getting the compiler to 
generate better code for constructors and dups in both NEON and SVE 
targets.  It first changes the backend to use rtx_vector_builder for 
vector_init's. Then it is followed by some prepraration passes to better 
handle VLA VEC_PERM_EXPRs followed by the addition of a new TARGET_HOOK 
VLA_CONSTRUCTOR that is used to expand VLA VEC_PERM_EXPRs, all based on 
Prathamesh's initial work in this area. As I said before, this is still 
work in progress, though I suspect the first two patches could go in but 
I was trying to get the series ready to post to make sure the first 
patches were in the right shape.

I have put this work on hold right now, but I heard Prathamesh might 
want to pick this up, feel free to use any of this, or discard as you 
see fit.

Andre Vieira (4)
aarch64: Encourage use of GPR input for SIMD inserts
aarch64: Change aarch64_expand_vector_init to use rtx_vector_builder
match.pd: Teach forwprop to handle VLA VEC_PERM_EXPRs with VLS 
CONSTRUCTORs as arguments
[RFC]: VLA Constructors

Kind regards,
Andre