[PATCH] D119965: [LICM][PhaseOrder] Don't speculate in LICM until after running loop rotate

Mon Apr 4 14:24:27 PDT 2022

Carrot added a comment.

>> I'm thinking of the following change, so we can remove PHI instruction without introduce any extra instructions on any path.
>>
>>   BB1:
>>      br %cond, label %BB2, label %BB3
>>   
>>   BB2:
>>      ...
>>     %161 = or i64 %24, 1
>>     ...
>>      br label BBX
>>   
>>   BB3:
>>      ...
>>     %179 = or i64 %24, 1
>>     ...
>>      br label BBX
>>   
>>   BBX:
>>      // our interesting bb
>>     %245 = phi i64 [ %161, %BB2 ], [ %179, %BB3 ]
>>     ...
>>     %296 = mul nsw i64 %245, %5, !dbg !2155
>>     %297 = getelementptr inbounds float, float* %4, i64 %296, !dbg !2156
>>     ...
>>   
>>   ==>
>>   
>>   BB1:
>>      ...
>>      // New instruction is inserted here
>>      %245 = or i64 %24, 1
>>      br %cond, label %BB2, label %BB3
>>   
>>   BB2:
>>      ...
>>     ...
>>      br label BBX
>>   
>>   BB3:
>>      ...
>>     // Use of %245.
>>     ...
>>      br label BBX
>>   
>>   BBX:
>>     // our interesting bb
>>     // PHI instruction is deleted.
>>     ...
>>     %296 = mul nsw i64 %245, %5, !dbg !2155
>>     %297 = getelementptr inbounds float, float* %4, i64 %296, !dbg !2156
>>     ...
>
> That kinda sounds like something for GVNHoist, which i think is currently still disabled due to some miscompilations?

I tried GVNHoist, it works as expected, moves  the "or" instruction to BB1.

But BB1 is in a loop, later LICM moves the "or" instruction to the loop preheader. Unfortunately another identical "or" instruction exists in one of its predecessors, I encountered the same problem as before GVNHoist, but this time no GVNHoist can hoist the "or" instruction, and later GVNPass creates a PHI for it.

Can I add another GVNHoist pass after loop optimizations?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D119965/new/

https://reviews.llvm.org/D119965