"Instruction does not dominate all uses" after Complex Deinterleaving Pass

[danilaml]

For the following IR:

target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-ni:1-p2:32:8:8:32-ni:2"
target triple = "aarch64-none-linux-gnu"

define void @foo() #0 {
bb:
  br label %bb173

bb173:                                            ; preds = %bb173, %bb
  %phi177 = phi <2 x i32> [ %add190, %bb173 ], [ zeroinitializer, %bb ]
  %phi178 = phi <2 x i32> [ %add187, %bb173 ], [ zeroinitializer, %bb ]
  %add185 = add <2 x i32> %phi178, <i32 1, i32 1>
  %add186 = add <2 x i32> %phi177, <i32 1, i32 1>
  %shufflevector = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %add187 = add <2 x i32> %add185, %shufflevector
  %shufflevector189 = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %add190 = add <2 x i32> %add186, %shufflevector189
  br i1 poison, label %bb193, label %bb173

bb193:                                            ; preds = %bb173
  %add194 = or <2 x i32> %add190, %add187
  store volatile i32 0, ptr null, align 4
  unreachable
}

attributes #0 = { "target-cpu"="neoverse-v1" }

Complex deinterleaving produced broken code:

identifyNode on   %add187 = add <2 x i32> %add185, %shufflevector /   %add190 = add <2 x i32> %add186, %shufflevector189
identifyNode on   %add185 = add <2 x i32> %phi178, <i32 1, i32 1> /   %add186 = add <2 x i32> %phi177, <i32 1, i32 1>
identifyNode on   %phi178 = phi <2 x i32> [ %add187, %bb173 ], [ zeroinitializer, %bb ] /   %phi177 = phi <2 x i32> [ %add190, %bb173 ], [ zeroinitializer, %bb ]
identifyNode on <2 x i32> <i32 1, i32 1> / <2 x i32> <i32 1, i32 1>
identifyNode on   %shufflevector = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer /   %shufflevector189 = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
Identified reduction starting from instructions:   %add187 = add <2 x i32> %add185, %shufflevector /   %add190 = add <2 x i32> %add186, %shufflevector189
Verifying parent property of node %bb173
Verifying parent property of node %bb
*** IR Dump After Complex Deinterleaving Pass (complex-deinterleaving) ***
define void @foo() #0 {
bb:
  %0 = call <4 x i32> @llvm.experimental.vector.interleave2.v4i32(<2 x i32> zeroinitializer, <2 x i32> zeroinitializer)
  br label %bb173

bb173:                                            ; preds = %bb173, %bb
  %1 = phi <4 x i32> [ %0, %bb ], [ %4, %bb173 ]
  %shufflevector = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %2 = call <4 x i32> @llvm.experimental.vector.interleave2.v4i32(<2 x i32> <i32 1, i32 1>, <2 x i32> <i32 1, i32 1>)
  %3 = add <4 x i32> %1, %2
  %4 = add <4 x i32> %3, %5
  %shufflevector189 = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %5 = call <4 x i32> @llvm.experimental.vector.interleave2.v4i32(<2 x i32> %shufflevector, <2 x i32> %shufflevector189)
  br i1 true, label %bb193, label %bb173

bb193:                                            ; preds = %bb173
  %6 = call { <2 x i32>, <2 x i32> } @llvm.experimental.vector.deinterleave2.v4i32(<4 x i32> %4)
  %7 = extractvalue { <2 x i32>, <2 x i32> } %6, 0
  %8 = extractvalue { <2 x i32>, <2 x i32> } %6, 1
  %add194 = or <2 x i32> %8, %7
  store volatile i32 0, ptr null, align 4
  unreachable
}
... eventually
Instruction does not dominate all uses!
  %4 = call <4 x i32> @llvm.experimental.vector.interleave2.v4i32(<2 x i32> %shufflevector, <2 x i32> %shufflevector189)
  %3 = add <4 x i32> %2, %4
in function foo
LLVM ERROR: Broken function found, compilation aborted!

Godbolt link https://godbolt.org/z/7zbbePPd6

Bisect pointed to c15557d by @igogo-x86

[danilaml]

Looks like it's because and extra builder is created for insertion here:

llvm-project/llvm/lib/CodeGen/ComplexDeinterleavingPass.cpp

Line 1952 in c15557d

Instruction *InsertPoint = (I->comesBefore(R) ? R : I)->getNextNode();

While the method that replaced operands uses the main builder to insert the main function. Don't know if the fix is as simple as changing the insertion point of the main builder (or perhaps recalculating it based on the operands).

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[igogo-x86]

Thanks, I am on it

[danilaml]

@igogo-x86 in case you've missed it, I've submitted a potential fix for review here: https://reviews.llvm.org/D159123

[igogo-x86]

I have a fix here - https://reviews.llvm.org/D159209