[llvm] [InstCombinePHI] Enhance PHI CSE to remove redundant phis (PR #163453)

[via llvm-commits]

CongzheUalberta wrote:

> > target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
> > target triple = "x86_64-unknown-fuchsia"
> > define i8 @f(ptr %__p, i1 %0) {
> > entry:
> > br label %for.cond
> > for.cond:                                         ; preds = %for.inc, %entry
> > %__p.pn = phi ptr [ %__p, %entry ], [ %__p.addr.0, %for.inc ]
> > %__p.addr.0 = getelementptr i8, ptr %__p.pn, i64 1
> > br i1 false, label %cleanup17, label %for.body
> > for.body:                                         ; preds = %for.cond
> > %1 = load i8, ptr %__p.pn, align 1
> > br i1 false, label %for.inc, label %if.else
> > if.else:                                          ; preds = %for.body
> > %incdec.ptr11 = getelementptr i8, ptr %__p.pn, i64 2
> > %spec.select = select i1 %0, ptr %__p.addr.0, ptr %incdec.ptr11
> > br label %cleanup17
> > for.inc:                                          ; preds = %for.body
> > br label %for.cond
> > cleanup17:                                        ; preds = %if.else, %for.cond
> > %__p.addr.3 = phi ptr [ %spec.select, %if.else ], [ %__p.addr.0, %for.cond ]
> > %2 = load i8, ptr %__p.addr.3, align 1
> > ret i8 %2
> > }
> 
> > PN = %lsr.iv = phi ptr [ %scevgep1, %for.inc ], [ %scevgep, %entry ]
> > IdenticalPN = %sunk_phi5 = phi ptr
> 
> It is caused by CGP:AddressingModeMatcher. It calls simplifyInstruction before filling the incoming values of new PHI nodes. I will post a fix tonight (UTC+8).

Thanks for looking into it! I've also looked into the crash and I reached the same conclusion as yours. I'm just waiting for all my local tests to finish before replying here.

It seems that the following update would fix the crash:
from
` if (PN.getNumIncomingValues() != 2) return false;` 
to 
` if (PN.getNumIncomingValues() != 2 || IdenticalPN.getNumIncomingValues() != 2) return false;` 
Or if you have something more appropriate in mind, I'd be glad to see it landed.

Thanks for your willingness to post a fix, I'd very much appreciate it!

https://github.com/llvm/llvm-project/pull/163453