SSE lane splitting: a series of 8 small pull requests (#9742 to #9762), how would you like them reviewed?

[ai-hakzarov]

Hello Ghidra team,

I opened eight small pull requests between 8 and 10 October, all about SSE code that the decompiler renders as opaque pcodeops or as unsplit vector registers. The contributing guide suggests opening a dialogue, so here is one place for the whole series. I'd like to know how you want it reviewed.

Why: on a large commercial x86-64 binary built around packed-float math, 61 of about 1,370 functions in one subsystem decompile as vector noise on current master. On our fork these fixes, plus four not yet submitted, make all 61 readable.

The series (each is one commit, applies alone on master, and adds a datatest to certification.manifest):

x86 SLEIGH (@GhidorahRex):

• x86: SQRTPS, DIVPS and MOVMSKPS as lane-wise p-code instead of opaque pcodeops #9742 SQRTPS, DIVPS, MOVMSKPS as lane-wise p-code
• x86: PSLLDQ as a single 128-bit shift, matching PSRLDQ #9743 PSLLDQ as one 128-bit shift, matching PSRLDQ

Decompiler lane splitting (@caheckman), smallest first:

• Decompiler: lane split follows a left shift forward #9744 lane split follows a left shift forward (1 line)
• Decompiler: lane split follows a zero extension forward #9762 lane split follows a zero extension forward
• Decompiler: lane division runs again as laned storage appears #9755 lane division runs again as laned storage appears
• Decompiler: lane split reassembles a value consumed whole by a call #9745 lane split reassembles a value consumed whole by a call
• Decompiler: a piece or zero extension of a laned register that is not a lane size suggests the lane size that tiles it #9759 a piece or zero extension that is not a lane size suggests the lane size that tiles it
• Decompiler: the default lane size guess reads the lanes its storage already uses #9761 the default lane size guess reads the lanes its storage already uses

Questions:

1. Would you rather review the six decompiler changes as one pull request? I'm happy to squash them, or to keep them separate as the guide asks.
2. Is a different approach for any of these better aligned with where lane splitting is going internally? I'll rework them.

I'll hold the remaining four fixes until these have been looked at, so I don't add to the queue. Thanks for your time.