[PATCH] D112268: [LegalizeTypes][RISCV][PowerPC] Expand CTLZ/CTTZ/CTPOP instead of promoting if they'll be expanded later.

[Thomas Preud'homme via Phabricator via llvm-commits]

thopre added a comment.

In D112268#3138471 <https://reviews.llvm.org/D112268#3138471>, @craig.topper wrote:

> In D112268#3137708 <https://reviews.llvm.org/D112268#3137708>, @thopre wrote:
>
>> Hi @craig.topper ,
>>
>> This caused a regression for us for i16 cttz with is_undef_zero true because cttz is expanded to a sequence using ctpop which loses the fact that any output is ok for the 0 case.
>>
>>   t17: i16 = sub t9, Constant:i16<1>
>>   t19: i16 = xor t9, Constant:i16<-1>
>>   t20: i16 = and t19, t17
>>   t21: i16 = ctpop t20
>>
>> So when this gets promoted, a mask is inserted to not have the result of ctpop change when `t9` is 0.
>>
>>   t15: i32 = extract_vector_elt t2, Constant:i32<0>
>>   t25: i32 = xor t15, Constant:i32<-1>
>>   t23: i32 = sub t15, Constant:i32<1>
>>   t26: i32 = and t25, t23
>>   t28: i32 = and t26, Constant:i32<65535>
>>   t29: i32 = ctpop t28
>>
>> Prior to this patch, we'd get:
>>
>>   t19: i32 = sub t15, Constant:i32<1>
>>   t21: i32 = xor t15, Constant:i32<-1>
>>   t22: i32 = and t21, t19
>>   t23: i32 = ctpop t22
>>
>> I thought about inserting some llvm.assume saying the value is non zero when expanding but those work at IR level. Any pointer on how to solve this? Best regards.
>
> I take it i32 ctpop is legal for your target or the i16 ctpop would have gotten expanded too?

Correct.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D112268/new/

https://reviews.llvm.org/D112268