[MergeICmps] Fix missing split.

[courbet]

Summary

We were not correctly splitting a blocks for chains of length 1.

This supersedes D92364.

Test case by MaskRay (Fangrui Song).

[jyknight]

llvm/test/Transforms/MergeICmps/X86/gep-references-bb.ll

6–8

Indirection unnecessary; can just put i32 directly in %Triple

10–12

Comment needs to be fixed.

13

This example should've really been transformed into a single memcmp of 12 bytes, not a cmp of the first 4-bytes, and a memcmp of the latter 8. So, there's a missed-optimization bug in this pass, as well. (I guess it's only looking at the direct pointer arg of the GEP, not following through them the chain of GEPs to find the base object.)

That missed-opt shouldn't be fixed at the same time as the miscompile, of course, but the current test is fragile w.r.t. that missing optimization -- if we fix the pass to compile this to a single 12-byte memcmp, we'll no longer have the test case for a size==1 comparison.

I think it would be good to make the test insensitive to that opt by adding another i32 field in the struct, between the first and the second, which gets ignored by the compares -- and update the size to dereferenceable(16). That'll still trigger the misoptimization, but not be transformable into a single memcmp, so this test will still test what happens for the chain of length 1 case -- even after fixing that issue in a future change.

Maybe also include a side-affecting instruction at the top of the function which should _definitely_ not be removed, like store i32 1, i32* @global.

[courbet]

Update comment.

[courbet]

Simplify the test as suggested by James.

[courbet]

Make the test more robust to future optimizations as suggested by James.

[courbet]

...and add a side-effecting instruction at the top.

[courbet]

Thanks for the comments.

[jyknight]

Oh -- if it's not too late, can you fix the commit message before pushing -- it could probably do with a little more explanation.

[courbet]

Improve commit message.

[MaskRay]

As a follow-up, would be good adding some tests where the pointer operand of a GEP references an instruction defined in a basic block. Currently in all tests the pointer operands refer to arguments.

llvm/lib/Transforms/Scalar/MergeICmps.cpp

static BasicBlock *mergeComparisons(ArrayRef<BCECmpBlock> Comparisons,
   IRBuilder<> Builder(BB);
   // Add the GEPs from the first BCECmpBlock.
   Value *const Lhs = Builder.Insert(FirstCmp.Lhs().GEP->clone());
   Value *const Rhs = Builder.Insert(FirstCmp.Rhs().GEP->clone());
 
   Value *IsEqual = nullptr;
   LLVM_DEBUG(dbgs() << "Merging " << Comparisons.size() << " comparisons -> "
                     << BB->getName() << "\n");
-  if (Comparisons.size() == 1) {
-    LLVM_DEBUG(dbgs() << "Only one comparison, updating branches\n");
-    Value *const LhsLoad =
-        Builder.CreateLoad(FirstCmp.Lhs().LoadI->getType(), Lhs);
-    Value *const RhsLoad =
-        Builder.CreateLoad(FirstCmp.Rhs().LoadI->getType(), Rhs);
-    // There are no blocks to merge, just do the comparison.
-    IsEqual = Builder.CreateICmpEQ(LhsLoad, RhsLoad);
-  } else {
-    // If there is one block that requires splitting, we do it now, i.e.
-    // just before we know we will collapse the chain. The instructions
-    // can be executed before any of the instructions in the chain.
-    const auto ToSplit =
-        std::find_if(Comparisons.begin(), Comparisons.end(),
-                     [](const BCECmpBlock &B) { return B.RequireSplit; });
-    if (ToSplit != Comparisons.end()) {
-      LLVM_DEBUG(dbgs() << "Splitting non_BCE work to header\n");
-      ToSplit->split(BB, AA);
-    }
+
 
+  if (Comparisons.size() == 1) {
+    LLVM_DEBUG(dbgs() << "Only one comparison, updating branches\n");
+    Value *const LhsLoad =
+        Builder.CreateLoad(FirstCmp.Lhs().LoadI->getType(), Lhs);
+    Value *const RhsLoad =
+        Builder.CreateLoad(FirstCmp.Rhs().LoadI->getType(), Rhs);
+    // There are no blocks to merge, just do the comparison.
+    IsEqual = Builder.CreateICmpEQ(LhsLoad, RhsLoad);
+  } else {
     const unsigned TotalSizeBits = std::accumulate(
         Comparisons.begin(), Comparisons.end(), 0u,
         [](int Size, const BCECmpBlock &C) { return Size + C.SizeBits(); });
 
     // Create memcmp() == 0.
     const auto &DL = Phi.getModule()->getDataLayout();
     Value *const MemCmpCall = emitMemCmp(
         Lhs, Rhs,

llvm/test/Transforms/MergeICmps/X86/gep-references-bb.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -mergeicmps -verify-dom-info | FileCheck %s
+target triple = "x86_64"
+
+%Triple = type { %Elem0, %Elem1, %Elem2 }
+%Elem0 = type { i32 }
+%Elem1 = type { i32 }
+%Elem2 = type { i32 }
+
+;; %gep does not reference an argument. Disable the optimization because
+;; otherwise the newly created gep may reference undef (after the basic block
+;; defining the pointer operand is deleted).
+define i1 @bug(%Triple* nonnull dereferenceable(12) %lhs, %Triple* nonnull dereferenceable(12) %rhs) {
+; CHECK-LABEL: @bug(
+; CHECK-NEXT:  bb02:
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr [[TRIPLE:%.*]], %Triple* [[RHS:%.*]], i64 0, i32 0
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[TRIPLE]], %Triple* [[LHS:%.*]], i64 0, i32 0, i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[TRIPLE]], %Triple* [[RHS]], i64 0, i32 0, i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[TMP0]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = load i32, i32* [[TMP1]], align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[TMP2]], [[TMP3]]
+; CHECK-NEXT:    br i1 [[TMP4]], label %"bb1+bb2", label [[FINAL:%.*]]
+; CHECK:       "bb1+bb2":
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[TRIPLE]], %Triple* [[LHS]], i64 0, i32 1, i32 0
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[ELEM0:%.*]], %Elem0* [[GEP]], i64 1, i32 0
+; CHECK-NEXT:    [[CSTR:%.*]] = bitcast i32* [[TMP5]] to i8*
+; CHECK-NEXT:    [[CSTR1:%.*]] = bitcast i32* [[TMP6]] to i8*
+; CHECK-NEXT:    [[MEMCMP:%.*]] = call i32 @memcmp(i8* [[CSTR]], i8* [[CSTR1]], i64 8)
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i32 [[MEMCMP]], 0
+; CHECK-NEXT:    br label [[FINAL]]
+; CHECK:       final:
+; CHECK-NEXT:    [[RET:%.*]] = phi i1 [ [[TMP7]], %"bb1+bb2" ], [ false, [[BB02:%.*]] ]
+; CHECK-NEXT:    ret i1 [[RET]]
+;
+bb0:
+  %gep = getelementptr %Triple, %Triple* %rhs, i64 0, i32 0
+  %l0_addr = getelementptr inbounds %Triple, %Triple* %lhs, i64 0, i32 0, i32 0
+  %l0 = load i32, i32* %l0_addr, align 4
+  %r0_addr = getelementptr inbounds %Triple, %Triple* %rhs, i64 0, i32 0, i32 0
+  %r0 = load i32, i32* %r0_addr, align 4
+  %cmp0 = icmp eq i32 %l0, %r0
+  br i1 %cmp0, label %bb1, label %final
+
+bb1:                                           ; preds = %bb0
+  %l1_addr = getelementptr inbounds %Triple, %Triple* %lhs, i64 0, i32 1, i32 0
+  %l1 = load i32, i32* %l1_addr, align 4
+  %r1_addr = getelementptr inbounds %Elem0, %Elem0* %gep, i64 1, i32 0
+  %r1 = load i32, i32* %r1_addr, align 4
+  %cmp1 = icmp eq i32 %l1, %r1
+  br i1 %cmp1, label %bb2, label %final
+
+bb2:                                           ; preds = %bb1
+  %l2_addr = getelementptr inbounds %Triple, %Triple* %lhs, i64 0, i32 2, i32 0
+  %l2 = load i32, i32* %l2_addr, align 4
+  %r2_addr = getelementptr inbounds %Elem0, %Elem0* %gep, i64 2, i32 0
+  %r2 = load i32, i32* %r2_addr, align 4
+  %cmp2 = icmp eq i32 %l2, %r2
+  br label %final
+
+final:                                            ; preds = %bb2, %bb1, %bb0
+  %ret = phi i1 [ false, %bb0 ], [ false, %bb1 ], [ %cmp2, %bb2 ]
+  ret i1 %ret
+}
+