[llvm] r359633 - [InstCombine] Limit a vector demanded elts rule which was producing invalid IR.

[Philip Reames via llvm-commits]

Author: reames
Date: Tue Apr 30 16:09:26 2019
New Revision: 359633

URL: http://llvm.org/viewvc/llvm-project?rev=359633&view=rev
Log:
[InstCombine] Limit a vector demanded elts rule which was producing invalid IR.

The demanded elts rules introduced for GEPs in https://reviews.llvm.org/rL356293 replaced vector constants with undefs (by design).  It turns out that the LangRef disallows such cases when indexing structs.  The right fix is probably to relax the langref requirement, and update other passes to expect the result, but for the moment, limit the transform to avoid compiler crashes.

This should fix https://bugs.llvm.org/show_bug.cgi?id=41624.


Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
    llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp?rev=359633&r1=359632&r2=359633&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp Tue Apr 30 16:09:26 2019
@@ -1169,6 +1169,18 @@ Value *InstCombiner::SimplifyDemandedVec
   default: break;
 
   case Instruction::GetElementPtr: {
+    // The LangRef requires that struct geps have all constant indices.  As
+    // such, we can't convert any operand to partial undef.
+    auto mayIndexStructType = [](GetElementPtrInst &GEP) {
+      for (auto I = gep_type_begin(GEP), E = gep_type_end(GEP);
+           I != E; I++)
+        if (I.isStruct())
+          return true;;
+      return false;
+    };
+    if (mayIndexStructType(cast<GetElementPtrInst>(*I)))
+      break;
+    
     // Conservatively track the demanded elements back through any vector
     // operands we may have.  We know there must be at least one, or we
     // wouldn't have a vector result to get here. Note that we intentionally

Modified: llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll?rev=359633&r1=359632&r2=359633&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/vec_demanded_elts.ll Tue Apr 30 16:09:26 2019
@@ -638,3 +638,16 @@ define i32* @gep_demanded_lane_undef(i32
   %ee = extractelement <2 x i32*> %gep, i32 1
   ret i32* %ee
 }
+
+
+;; LangRef has an odd quirk around FCAs which make it illegal to use undef
+;; indices.
+define i32* @PR41624(<2 x { i32, i32 }*> %a) {
+; CHECK-LABEL: @PR41624(
+; CHECK-NEXT:   %w = getelementptr { i32, i32 }, <2 x { i32, i32 }*> %a, <2 x i64> <i64 5, i64 5>, <2 x i32> zeroinitializer
+; CHECK-NEXT:   %r = extractelement <2 x i32*> %w, i32 0
+; CHECK-NEXT:   ret i32* %r
+  %w = getelementptr { i32, i32 }, <2 x { i32, i32 }*> %a, <2 x i64> <i64 5, i64 5>, <2 x i32> zeroinitializer
+  %r = extractelement <2 x i32*> %w, i32 0
+  ret i32* %r
+}