mirror remaining mmx/xmm dataflow operations

[derekbruening]

From bruen...@google.com on March 22, 2014 11:04:56

split from issue #243 which added the basic shadowing and propagation, and dataflow mirroring for most of the pack, insert, and extract operations

\ TODO remaining dataflow operations to add

repeating the list from above:

• [ ] pack* => can just propagate
• [X] unpck*
• [X] punpck*
• [ ] pshuf*
• [X] shufp*
• [ ] palignr
• [ ] psr*
• [ ] psl*
• [X] movlhp*
• [X] movhlp*
• [X] movhp*
• [X] extractps
• [X] insertps
• [ ] cvt* => we can just propagate, even if sizes don't match, right?
• [ ] phminposuw
• [ ] pcmpestr*
• [ ] blend
• [ ] and/or w/ constant

*\ TODO need to scale shadow for conversions

cvt routines that convert downward like cvtpd2ps

also things that extend: cvt, pmovsx, pmovzx*

*\ TODO conditional moves

blend pblend

*\ TODO opcodes that search the xmm reg and so vary on what they read!

phminposuw: what to propagate to dst depends on actual value in xmm pshufb has a similar property

pcmpestr*: upper bounds determined by immeds. always reads it all? ever short-circuit?

*\ TODO and and or w/ constants need to mark defined probably rarer to have false pos here than w/ gpr b/c it's harder to get consts into xmm

Original issue: http://code.google.com/p/drmemory/issues/detail?id=1484

[derekbruening]

The list is now:

• [ ] pack* => can just propagate
• [X] unpck*
• [X] punpck*
• [ ] pshuf*
• [X] shufp*
• [ ] palignr
• [X] psr*
• [X] psl*
• [X] movlhp*
• [X] movhlp*
• [X] movhp*
• [X] extractps
• [X] insertps
• [ ] cvt* => have to scale shadow if converts downward
• [ ] phminposuw
• [ ] pcmpestr*
• [ ] blend
• [ ] and/or w/ constant

[derekbruening]

Looking at it now, I don't see how pack* can use normal propagation as it shrinks the data.