Optimization: Lower load and store operations by pruning types

[Robbepop]

Currently, the wasmi_ir crate supports these load operations among others:

• {I,F}{32,64}Load
• {I,F}{32,64}LoadAt
• {I,F}{32,64}LoadOffset16

As well as the following store operations among others:

• {F,I}{32,64}Store
• I{32,64}StoreImm16*
• {F,I}{32,64}StoreOffset16
• I{32,64}StoreOffset16Imm16*
• {F,I}{32,64}StoreAt
• I{32,64}StoreAtImm16*

(*): The Imm16 postfixed variants cannot be further generlized. The same applies to truncation-stores and extension-loads.

Since Wasmi bytecode is executed without type safety or type checking we could easily lower all of these instructions down to load and store instructions that are just loading and storing a fixed number of bytes instead:

• Load{32,64}
• Load{32,64}At
• Load{32,64}Offset16
• Store{32,64}
• Store{32,64}Offset16
• Store{32,64}At

Effects

• Replace 24 instructions by just 12 new ones.
• Simplify Wasmi executor.
• Reduce pressure on the CPU's instruction cache.