68k Pipeline optimizer

[SamuraiCrow]

When adding the QBE code to the backends, add a simple pass that rearranges instructions differently so that the result of one instruction isn't as frequently dependent on the one immediately before it.

Features

• Increases speed on in-order pipelines like the 68020+ and Amiga-vintage PowerPC.
• Especially vital to super-scalar in-order performance
• Few side effects on non-pipelined architectures like 68000 and TG68k.C

Things to Watch For

• Dependency mapping must consider condition code register dependencies as well
• Register dependencies must also consider loads and stores to memory (heap and stack access)
• Should be last among passes because ordering is affected by other optimizations
• Opcode fusions should not be split-up as they merge into one operation on the Apollo 68080 and NatAmi 68050/68070 and compatibles as documented at the Apollo Fusing page.

[SamuraiCrow]

I'm working on this one for the codegen library anyway. I'm self-assigning this ticket. The backend optimization needed work anyway.

[SamuraiCrow]

It's looking like this optimizer is more involved in development than the QBE macro optimizer. I'll postpone this until that is added.