[LoopSimplifyCFG] Pay respect to LCSSA when removing dead blocks

[mkazantsev]

Summary

Utility function that we use for blocks deletion always unconditionally removes
one-input Phis. In LoopSimplifyCFG, it can lead to breach of LCSSA form.
This patch alters this function to keep them if needed.

[fhahn]

I think ideally we would just avoid deleting LCSSA PHI nodes (PHI nodes with a single incoming value in loop exit blocks), but maybe it is too much effort with the info available.

include/llvm/Transforms/Utils/BasicBlockUtils.h

44 ↗

(On Diff #183518)

nit: How about KeepUselessPHIs, or KeepSingleValuePHIs, which is a bit more descriptive.

test/Transforms/LoopSimplifyCFG/lcssa.ll

2 ↗

(On Diff #183518)

I think the relevant -verify-loop-lcssa & co are available without asserts as well, so we should be able to drop REQUIRE: asserts.

[mkazantsev]

include/llvm/Transforms/Utils/BasicBlockUtils.h

44 ↗

(On Diff #183518)

I used exactly the same name as existing function removePredecessor has. We can make a follow-up refactoring to change them all, but so far I'd prefer to keep things uniform.

test/Transforms/LoopSimplifyCFG/lcssa.ll

2 ↗

(On Diff #183518)

I didn't check that. We can do it as follow-up if it works.

[mkazantsev]

Ping?

[fedor.sergeev]

LGTM, but please, schedule those followups discussed here.

include/llvm/Transforms/Utils/BasicBlockUtils.h

44 ↗

(On Diff #183518)

I also find this name very confusing...

[mkazantsev]

include/llvm/Transforms/Utils/BasicBlockUtils.h

44 ↗

(On Diff #183518)

I will rename all its occurrences in a follow-up.

[mkazantsev]

I'll split out changes in basic block utils as separate NFC, then submit change in LoopSimplifyCFG separately.

[mkazantsev]

include/llvm/Transforms/Utils/BasicBlockUtils.h

44 ↗

(On Diff #183518)

https://reviews.llvm.org/rL353801

[fedor.sergeev]

In D57231#1394232, @mkazantsev wrote:

I'll split out changes in basic block utils as separate NFC, then submit change in LoopSimplifyCFG separately.

Thanks, your effort is really appreciated!

llvm/trunk/lib/Transforms/Scalar/LoopSimplifyCFG.cpp

for (auto *BB : DeadLoopBlocks) {
                         << "\n");
       if (LI.isLoopHeader(BB)) {
         assert(LI.getLoopFor(BB) != &L && "Attempt to remove current loop!");
         LI.erase(LI.getLoopFor(BB));
       }
       LI.removeBlock(BB);
     }
 
-    DetatchDeadBlocks(DeadLoopBlocks, &DTUpdates);
+    DetatchDeadBlocks(DeadLoopBlocks, &DTUpdates, /*KeepOneInputPHIs*/true);
     DTU.applyUpdates(DTUpdates);
     DTUpdates.clear();
     for (auto *BB : DeadLoopBlocks)
       BB->eraseFromParent();
 
     NumLoopBlocksDeleted += DeadLoopBlocks.size();
   }
 

llvm/trunk/test/Transforms/LoopSimplifyCFG/lcssa.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; XFAIL: *
-; REQUIRES: asserts
 ; RUN: opt -S -enable-loop-simplifycfg-term-folding=true -loop-simplifycfg -verify-loop-info -verify-dom-info -verify-loop-lcssa < %s | FileCheck %s
 ; RUN: opt -S -enable-loop-simplifycfg-term-folding=true -passes='require<domtree>,loop(simplify-cfg)' -verify-loop-info -verify-dom-info -verify-loop-lcssa < %s | FileCheck %s
 ; RUN: opt -S -enable-loop-simplifycfg-term-folding=true -loop-simplifycfg -enable-mssa-loop-dependency=true -verify-memoryssa -verify-loop-info -verify-dom-info -verify-loop-lcssa < %s | FileCheck %s
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 
 define void @c() {
 ; CHECK-LABEL: @c(
while.cond:                                       ; preds = %while.body, %for.cond
   br i1 true, label %while.body, label %for.cond.loopexit
 
 while.body:                                       ; preds = %while.cond
   %inc4 = add nsw i16 %inc41, 1
   br label %while.cond
 }
 
 define void @bar() {
-
-; CHECK-LABEL: @bar
+; CHECK-LABEL: @bar(
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    switch i32 0, label [[BB_SPLIT:%.*]] [
+; CHECK-NEXT:    i32 1, label [[BB10:%.*]]
+; CHECK-NEXT:    ]
+; CHECK:       bb-split:
+; CHECK-NEXT:    br label [[BB1:%.*]]
+; CHECK:       bb1:
+; CHECK-NEXT:    [[TMP:%.*]] = phi i32 [ [[TMP7:%.*]], [[BB6:%.*]] ], [ undef, [[BB_SPLIT]] ]
+; CHECK-NEXT:    switch i32 undef, label [[BB5:%.*]] [
+; CHECK-NEXT:    i32 0, label [[BB6]]
+; CHECK-NEXT:    i32 1, label [[BB8:%.*]]
+; CHECK-NEXT:    ]
+; CHECK:       bb5:
+; CHECK-NEXT:    ret void
+; CHECK:       bb6:
+; CHECK-NEXT:    [[TMP7]] = add i32 undef, 123
+; CHECK-NEXT:    br label [[BB1]]
+; CHECK:       bb8:
+; CHECK-NEXT:    [[TMP9:%.*]] = phi i32 [ [[TMP]], [[BB1]] ]
+; CHECK-NEXT:    [[USE:%.*]] = add i32 [[TMP9]], 1
+; CHECK-NEXT:    ret void
+; CHECK:       bb10:
+; CHECK-NEXT:    ret void
+;
 
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb6, %bb
   %tmp = phi i32 [ %tmp7, %bb6 ], [ undef, %bb ]
   br i1 false, label %bb2, label %bb4
 
 bb2:                                              ; preds = %bb1
   switch i32 undef, label %bb10 [
-    i32 0, label %bb3
-    i32 1, label %bb8
   ]
 
 bb3:                                              ; preds = %bb2
   br label %bb6
 
 bb4:                                              ; preds = %bb1
   switch i32 undef, label %bb5 [
-    i32 0, label %bb6
-    i32 1, label %bb8
   ]
 
 bb5:                                              ; preds = %bb4
   ret void
 
 bb6:                                              ; preds = %bb4, %bb3
   %tmp7 = add i32 undef, 123
   br label %bb1