On Mon, Mar 07, 2022 at 10:15:48AM +0800, Hongtao Liu wrote:
> > Note, the predicates on cond_fma* and other FMA variants look all wrong to
> > me, usually the fma instructions require nonimmediate_operand operands,
> > but the cond_* patterns use vector_operand. Besides what this patch
> --------cut from predicate.md---------
> 1142; Return true when OP is operand acceptable for vector memory operand.
> 1143; Only AVX can have misaligned memory operand.
> 1144(define_predicate "vector_memory_operand"
> 1145 (and (match_operand 0 "memory_operand")
> 1146 (ior (match_test "TARGET_AVX")
> 1147 (match_test "MEM_ALIGN (op) >= GET_MODE_ALIGNMENT (mode)"))))
> 1148
> 1149; Return true when OP is register_operand or vector_memory_operand.
> 1150(define_predicate "vector_operand"
> 1151 (ior (match_operand 0 "register_operand")
> 1152 (match_operand 0 "vector_memory_operand")))
> --------cut end------------------------
>
> vector_operand is a subset of nonimmediate_operands, so it's more like
> a potential optimization issue rather than a correctness one?
You're right.
And fma apparently doesn't use something like ix86_fixup_binary_operands*,
allows all 3 inputs as MEMs and only deals with it at RA time.
Jakub
