1061637 - SIMD: optimize 'select' to use blendvps when SSE4.1 is available.

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine: JIT, defect, P5)

Description

[Douglas Crosher [:dougc]]

The blendvps SSE4.1 instruction may be faster than the standard sequence implemented in bug 1060437.

The semantics of blendvps are different to the 'select' operation, it only looks at the top sign bit to decide which input lanes to select for the output. However there are probably many useful cases in which 'select' could be transform into a blend operation, such as when 'select' follows a comparison.

Comment 1

[Douglas Crosher [:dougc]]

Attachment: Optimize 'select' to use blendvps when SSE4.1 is available.

WIP patch.

Comment 2

[Douglas Crosher [:dougc]]

Attachment: Optimize 'select' to use blendvps when SSE4.1 is available and vblendvps when AVX is available.

This is wip that can be used to explore the performance of blendvps. The semantics of blendvps differ from 'select' but it could still be used when we can prove that doing so does not change the semantics of the code, such as when it follows a comparison that sets all bits. It might also need some spec work to define or confirm how this should work. This patch is rebased after the VEX patches in bug 1065339 and also support vblendvps.