[mlir][LLVMIR] Do not cache Instruction generated on-the-fly

[myhsu]

Summary

We want to avoid caching llvm::Instruction generated by llvm::ConstantExpr::getAsInstruction into instMap. Because if we remove it right away, it's possible that when getAsInstruction is called again, it will create a new instruction that has the same address with the one we just deleted. Inducing a conflicting index in instMap and triggers an assertion in processInstruction.

Note that this patch depends on D124399 simply because it's easier to trigger this bug by putting ConstantExpr inside a ConstantAggregate, which is also how the test case was written.

[rriddle]

mlir/lib/Target/LLVMIR/ConvertFromLLVMIR.cpp

502

Can't we just remove the mapping of i inside of instMap? Wouldn't that make it safe to delete here?

[myhsu]

Taking @rriddle 's advice and simplify the patch

[ftynse]

Nice catch, thanks!

mlir/lib/Target/LLVMIR/ConvertFromLLVMIR.cpp

Value Importer::processConstant(llvm::Constant *c) {
   if (auto *ce = dyn_cast<llvm::ConstantExpr>(c)) {
     llvm::Instruction *i = ce->getAsInstruction();
     OpBuilder::InsertionGuard guard(b);
     b.setInsertionPoint(currentEntryBlock, currentEntryBlock->begin());
     if (failed(processInstruction(i)))
       return nullptr;
     assert(instMap.count(i));
 
+    // If we don't remove entry of `i` here, it's totally possible that the
+    // next time llvm::ConstantExpr::getAsInstruction is called again, which
+    // always allocates a new Instruction, memory address of the newly
+    // created Instruction might be the same as `i`. Making processInstruction
+    // falsely believe that the new Instruction has been processed before
+    // and raised an assertion error.
+    Value value = instMap[i];
+    instMap.erase(i);
     // Remove this zombie LLVM instruction now, leaving us only with the MLIR
     // op.
     i->deleteValue();
-    return instMap[c] = instMap[i];
+    return instMap[c] = value;
   }
   if (auto *ue = dyn_cast<llvm::UndefValue>(c)) {
     Type type = processType(ue->getType());
     if (!type)
       return nullptr;
     return instMap[c] = bEntry.create<UndefOp>(UnknownLoc::get(context), type);
   }
 
   if (isa<llvm::ConstantAggregate>(c) || isa<llvm::ConstantAggregateZero>(c)) {

mlir/test/Target/LLVMIR/Import/incorrect-constexpr-inst-caching.ll

+; RUN: mlir-translate --import-llvm %s | FileCheck %s
+; REQUIRES: asserts
+
+; This test is primarily used to make sure an assertion is not triggered.
+; Thus, we only wrote minimum level of checks.
+
+%my_struct = type {i32, i8*}
+; CHECK: llvm.mlir.constant(3 : i32) : i32
+; CHECK: llvm.mlir.constant(2 : i32) : i32
+; CHECK: llvm.mlir.addressof @str1 : !llvm.ptr<array<5 x i8>>
+; CHECK: llvm.getelementptr
+; CHECK: llvm.mlir.constant(7 : i32) : i32
+; CHECK: llvm.mlir.undef : !llvm.struct<"my_struct", (i32, ptr<i8>)>
+; CHECK: llvm.insertvalue
+; CHECK: llvm.insertvalue
+; CHECK: llvm.mlir.constant(1 : i32) : i32
+; CHECK: llvm.mlir.constant(0 : i32) : i32
+; CHECK: llvm.mlir.addressof @str0 : !llvm.ptr<array<5 x i8>>
+; CHECK: llvm.getelementptr
+; CHECK: llvm.mlir.constant(8 : i32) : i32
+; CHECK: llvm.mlir.undef : !llvm.struct<"my_struct", (i32, ptr<i8>)>
+; CHECK: llvm.insertvalue
+; CHECK: llvm.insertvalue
+; CHECK: llvm.mlir.undef : !llvm.array<2 x struct<"my_struct", (i32, ptr<i8>)>>
+; CHECK: llvm.insertvalue
+; CHECK: llvm.insertvalue
+; CHECK: llvm.return
+@str0 = private unnamed_addr constant [5 x i8] c"aaaa\00"
+@str1 = private unnamed_addr constant [5 x i8] c"bbbb\00"
+@g = global [2 x %my_struct] [%my_struct {i32 8, i8* getelementptr ([5 x i8], [5 x i8]* @str0, i32 0, i32 1)}, %my_struct {i32 7, i8* getelementptr ([5 x i8], [5 x i8]* @str1, i32 2, i32 3)}]
+