[llvm-dev] Bug in llvm sink pass

[Xin Tong via llvm-dev]

Hi

I found llvm Sink.cpp pass has a bug. Running opt -basicaa -sink
diamond.ll moves the load in entry BB to BBX. This is incorrect as
there is a store in BB0.

It seems Sink pass is not enabled in the O3 pipeline, Is this the
reason why it is disabled ?

Also, I see sink can not handle sinking across loop either.

*** IR Dump Before Code sinking ***

define i32 @diamond(i32 %a, i32 %b, i32 %c, i32* dereferenceable(32) %d) {

  %1 = mul nsw i32 %c, %b

  %2 = load i32, i32* %d

  %3 = icmp sgt i32 %a, 0

  br i1 %3, label %B0, label %B1

B0:                                               ; preds = %0

  store i32 0, i32* %d

  br label %X

B1:                                               ; preds = %0

  br label %X

X:                                                ; preds = %B1, %B0

%R = sub i32 %1, %2

  ret i32 %R

}

*** IR Dump Before Module Verifier ***

define i32 @diamond(i32 %a, i32 %b, i32 %c, i32* dereferenceable(32) %d) {

  %1 = icmp sgt i32 %a, 0

  br i1 %1, label %B0, label %B1

B0:                                               ; preds = %0

  store i32 0, i32* %d

  br label %X

B1:                                               ; preds = %0

  br label %X

X:                                                ; preds = %B1, %B0

  %2 = mul nsw i32 %c, %b

  %3 = load i32, i32* %d

  %R = sub i32 %2, %3

  ret i32 %R

}


Thanks
-Xin