[LoopVectorize] Miscompile with masked IV

[nikic]

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-linux-gnu"

; out[i] = x[(i * 5) & 16]
define void @test(ptr noalias %x, ptr noalias %out) {
start:
  br label %loop

loop:
  %iv = phi i64 [ 0, %start ], [ %iv.next, %loop ]
  %iv.next = add nuw nsw i64 %iv, 1
  %iv.mul5 = mul nuw nsw i64 %iv, 5
  %iv.mul5.masked = and i64 %iv.mul5, 15
  %gep = getelementptr inbounds nuw i32, ptr %x, i64 %iv.mul5.masked
  %v = load i32, ptr %gep, align 4
  %out.i = getelementptr inbounds nuw i32, ptr %out, i64 %iv
  store i32 %v, ptr %out.i, align 4
  %exitcond.not = icmp eq i64 %iv.next, 16
  br i1 %exitcond.not, label %exit, label %loop

exit:
  ret void
}

Produces:

define void @test(ptr noalias %x, ptr noalias %out) {
start:
  br i1 false, label %scalar.ph, label %vector.ph

vector.ph:
  br label %vector.body

vector.body:
  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
  %0 = mul nuw nsw i64 %index, 5
  %1 = and i64 %0, 15
  %2 = getelementptr inbounds nuw i32, ptr %x, i64 %1
  %wide.vec = load <20 x i32>, ptr %2, align 4
  %strided.vec = shufflevector <20 x i32> %wide.vec, <20 x i32> poison, <4 x i32> <i32 0, i32 5, i32 10, i32 15>
  %3 = getelementptr inbounds nuw i32, ptr %out, i64 %index
  store <4 x i32> %strided.vec, ptr %3, align 4
  %index.next = add nuw i64 %index, 4
  %4 = icmp eq i64 %index.next, 12
  br i1 %4, label %middle.block, label %vector.body

middle.block:
  br label %scalar.ph

scalar.ph:
  %bc.resume.val = phi i64 [ 12, %middle.block ], [ 0, %start ]
  br label %loop

loop:
  %iv = phi i64 [ %bc.resume.val, %scalar.ph ], [ %iv.next, %loop ]
  %iv.next = add nuw nsw i64 %iv, 1
  %iv.mul5 = mul nuw nsw i64 %iv, 5
  %iv.mul5.masked = and i64 %iv.mul5, 15
  %gep = getelementptr inbounds nuw i32, ptr %x, i64 %iv.mul5.masked
  %v = load i32, ptr %gep, align 4
  %out.i = getelementptr inbounds nuw i32, ptr %out, i64 %iv
  store i32 %v, ptr %out.i, align 4
  %exitcond.not = icmp eq i64 %iv.next, 16
  br i1 %exitcond.not, label %exit, label %loop

exit:
  ret void
}

What's happening here with the load and shufflevector doesn't really make sense. It seems like the masking on the IV ended up being ignored?

[nikic]

cc @fhahn

[fhahn]

Looks like an SCEVExpander issue, will share a fix soon