Skip to content

InstCombine introduces unnecessary and 1 when folding trunc nuw+zext #134908

Open
@scottmcm

Description

@scottmcm

Rust in which we spotted this originally: https://rust.godbolt.org/z/Knr5jdEY5, cc rust-lang/rust#139098 (comment)

Take this input IR:

define noundef range(i8 0, 14) i8 @match0(i8 noundef range(i8 0, 3) %0) unnamed_addr {
start:
  %1 = icmp eq i8 %0, 2
  %_2 = select i1 %1, i64 1, i64 0
  %2 = trunc nuw i64 %_2 to i1
  br i1 %2, label %bb2, label %bb3

bb2:                                              ; preds = %start
  br label %bb4

bb3:                                              ; preds = %start
  %b = trunc nuw i8 %0 to i1
  %3 = zext i1 %b to i8
  br label %bb4

bb4:                                              ; preds = %bb2, %bb3
  %_0.sroa.0.0 = phi i8 [ 13, %bb2 ], [ %3, %bb3 ]
  ret i8 %_0.sroa.0.0
}

Today https://llvm.godbolt.org/z/vnfGGxfbr, --passes=instcombine transforms that to

source_filename = "/app/example.ll"

define noundef range(i8 0, 14) i8 @match0(i8 noundef range(i8 0, 3) %0) unnamed_addr {
  %1 = icmp eq i8 %0, 2
  br i1 %1, label %bb2, label %bb3

bb2:                                              ; preds = %start
  br label %bb4

bb3:                                              ; preds = %start
  %2 = and i8 %0, 1 ; 😢
  br label %bb4

bb4:                                              ; preds = %bb3, %bb2
  %_0.sroa.0.0 = phi i8 [ 13, %bb2 ], [ %2, %bb3 ]
  ret i8 %_0.sroa.0.0
}

However it should not be adding that and i8 %0, 1 -- the trunc nuw means it ought to know it doesn't need to do that.

It should instead be emitting https://alive2.llvm.org/ce/z/oN3dig

define noundef range(i8 0, 14) i8 @tgt(i8 noundef range(i8 0, 3) %0) unnamed_addr {
start:
  %1 = icmp eq i8 %0, 2
  br i1 %1, label %bb2, label %bb3

bb2:                                              ; preds = %start
  br label %bb4

bb3:                                              ; preds = %start
  br label %bb4

bb4:                                              ; preds = %bb3, %bb2
  %_0.sroa.0.0 = phi i8 [ 13, %bb2 ], [ %0, %bb3 ]
  ret i8 %_0.sroa.0.0
}

Note that the correctness for this doesn't depend on the range information; removing that from the parameter it still validates https://alive2.llvm.org/ce/z/_ed2Ch.

(If you run a full -O2 or -O3 then CVP will get rid of the and, but that's not run in -O1. And InstCombine has the information it needs to do it correctly anyway, AFAICT, so it just should.)

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions