Additions to the instruction set
=====================================
1. SHnADD[O] instructions, which performs a*(2^n)+b
2. Branch on bit instruction (BB)
3. Single precision floating points + lots of other conversion ops
4. Loads and stores with scaled indexed and auto increment addressing modes
New optimizations in the HPPA backend
=====================================
1. An expressions such as a*n+b where n = 2, 4, and 8 now emits SHnADD[O]
instructions.
2. An expression such as
BCC(T.CMP(T.NE,0,T.ANDB(mask,e))
where mask is a power of 2 now emits BB instructions.
SML/NJ: this improves boxity testing code by replacing 3 instructions with 1
instruction.
3. Strength reduction for multiplication with small constants.
SML/NJ: benefits array indexing and multiway branches
4 Addressing modes computation now takes into account of scaled
indexed mode. Unfortunately, due to tagging this optimizations
is not used in SML/NJ.
5 Propagation of annotations
6 Delay slot filling has been added
7 NOTB folding has been implemented. This combines NOTB with other
bit ops such as ANDB, ORB, XORB.
8. COND expressions now generate the composite instruction sequence:
COMCLR,cc r1, r2, t
LDO i(b), t