https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65225
Bug ID: 65225
Summary: [AArch64] Various aarch64_rtx_costs improvements
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target: aarch64*
The aarch64 rtx costs walker could be improved. Some particular areas:
- In the MINUS case, the cost of new_op1 is not included at all
- In the PLUS and MINUS case, extends and extended shifts are costed as
alu_arith rather than extend_arith
- In the PLUS and MINUS case if there is a ASHIFT or MULT after an
extend/extended shift, the cost calculated in rtx_mult_cost is as if it was a
plain PLUS/MINUS. If the multiply is by a power of 2, it calculates the cost as
a arith_shift rather than extend_arith. If it was ASHIFT the cost is wrong as
it expects a MULT rather than ASHIFT, so shift immediates which happen to be a
power of 2 get a different cost.
- In the ASHIFT case, if there is an extend the cost of extend should be used
rather than a shift.
ASHIFT 1 should use alu cost rather than shift cost (we emit an add if shift
cost is higher than alu cost)
- In the ASHIFRT/LSHIFTRT case the extend logic is missing.
- In the ABS case for integer, speed cost should be 2 ALU instructions rather
than 4.
- In the and-immediate case look into handling ZERO_EXTRACT.
- The NOT case misses logic for shifts (despite the comment, calculating shift
cost separately is not correct) - this can be costed as ORN with zero similar
to the way NEG works.
