411687 - TT: optimize LIR_cmov

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Edwin Smith]

emit code that uses the condition instruction to choose one of the many MOVcc variants, instead of computing a boolean and testing for eq/ne 0.  (ie do like LIR_xf/xt do for conditional branches).

Comment 1

[Edwin Smith]

dbox could benefit from a quad-sized cmov

Comment 2

[Steven Johnson]

if dbox is the only place that's likely to benefit, might be simpler/better to just write a "normalizenan" primitive

Comment 3

[Steven Johnson]

Attachment: Patch

Redid LIR_cmov support to take advantage of the full suite of conditional moves rather than just MREQ. Combined with a little restructuring of the assembler code (to reduce redundant register moves) this removes several instructions from the typical cmov trace. As a result I rewrote the boxtype superword to use choose (rather than bit twiddling) as it produces slightly shorter code.

Note that the ARM implementation is untested but (theoretically) should work. Thumb doesn't support conditional moves so uses bit-twiddling as a fallback.

Comment 4

[Steven Johnson]

pushed as changeset:   284:abd80bd1eef4

Comment 5

[Chris Peyer]

Closing out all TT: transfer bugs that are resolved fixed.