657220 - JM+TI: don't break FP registers before stores

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Brian Hackett [Laid off!]]

Before IC'ed setprops or setelems with an RHS in an FP register, we break the FP register into two normal registers before doing the store.  This is silly and should get fixed.

Comment 1

[Brian Hackett [Laid off!]]

Partial patch, allows ValueRemat to represent values in FP registers and allows setelem_dense and the inlined Array.push to use them, but still breaks FP registers before IC'ed SETPROP or SETELEM.  I tried getting SETPROP to do this but force a stub generation (rather than use the inline path), but doing this slowed things down significantly.  The inline paths would need to be updated to handle this case but getting the baked-in offsets right looks tricky.

http://hg.mozilla.org/projects/jaegermonkey/rev/55438ad5632e

Comment 2

[Brian Hackett [Laid off!]]

Hmm, this busted ARM as the ARM assembler doesn't have functionality to store an FP register to a BaseIndex.  This is easy enough to fix by keeping the old behavior on ARM, but if this is just a gap in the assembler and not the actual instruction set (no clue) then such a fix creates a performance problem that would be hard to fix later.  (This patch speeds up storing doubles to a constant array index by about 60%).

Comment 3

[Jan de Mooij [:jandem]]

JM was removed.