Use pragmas to work around MSVC x86_32 debug miscompile bug

[rnk]

Summary

Halide users reported this here: https://llvm.org/pr46176
I reported the issue to MSVC here:
https://developercommunity.visualstudio.com/content/problem/1179643/msvc-copies-overaligned-non-trivially-copyable-par.html

This codepath is apparently not covered by LLVM's unit tests, so I added
coverage in a unit test.

If we want to support this configuration going forward, it means that is
in general not safe to pass a SmallVector<T, N> by value if alignof(T)
is greater than 4. This doesn't appear to come up often because passing
a SmallVector by value is inefficient and not idiomatic: it copies the
inline storage. In this case, the SmallVector<LLT,4> is captured by
value by a lambda, and the lambda is passed by value into std::function,
and that's how we hit the bug.

[asbirlea]

Thank you so much Reid for taking the time to look into this, discovering the msvc issue and proposing this work-around!

llvm/unittests/CodeGen/GlobalISel/LegalizerInfoTest.cpp

413

nit: fix clang-tidy warnings
s/s1/S1
s/p0/P0
s/builder/Builder

[craig.topper]

llvm/lib/CodeGen/GlobalISel/LegalityPredicates.cpp

13

This comment says "Disable optimizations", but the one below says "We have to disable runtime checks in order to enable optimizations". Are we enabling or disabling optimizations?

[rnk]

llvm/lib/CodeGen/GlobalISel/LegalityPredicates.cpp

13

Bleh, the late breaking comment update was wrong. We are enabling optimizations. I'll fix it.

You can observe the interaction between runtime checks and optimizations on the command line with cl -RTC1 -O2, it will produce an error. Using pragma optimize on by itself did nothing without producing a warning. =P

llvm/lib/CodeGen/GlobalISel/LegalityPredicates.cpp

 //===- lib/CodeGen/GlobalISel/LegalizerPredicates.cpp - Predicates --------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // A library of predicate factories to use for LegalityPredicate.
 //
 //===----------------------------------------------------------------------===//
 
+// Disable optimizations to work around MSVC debug mode bug in 32-bit:
+// https://developercommunity.visualstudio.com/content/problem/1179643/msvc-copies-overaligned-non-trivially-copyable-par.html
+// FIXME: Remove this when the issue is closed.
+#if defined(_MSC_VER) && !defined(__clang__) && defined(_M_IX86)
+// We have to disable runtime checks in order to enable optimizations. This is
+// done for the entire file because the problem is actually observed in STL
+// template functions.
+#pragma runtime_checks("", off)
+#pragma optimize("gs", on)
+#endif
+
 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
 
 using namespace llvm;
 
 LegalityPredicate LegalityPredicates::typeIs(unsigned TypeIdx, LLT Type) {
   return
       [=](const LegalityQuery &Query) { return Query.Types[TypeIdx] == Type; };
 }

llvm/unittests/CodeGen/GlobalISel/LegalizerInfoTest.cpp

EXPECT_ACTION(Legal, 0, LLT(),
                                 LegalityQuery::MemDesc{32,
                                     MaxAlignInBits, AtomicOrdering::NotAtomic}));
     EXPECT_ACTION(Unsupported, 0, LLT(),
                   LegalityQuery(G_LOAD, {s32, p0},
                                 LegalityQuery::MemDesc{
                                   32, 8, AtomicOrdering::NotAtomic }));
   }
 }
+
+// This code sequence doesn't do anything, but it covers a previously uncovered
+// codepath that used to crash in MSVC x86_32 debug mode.
+TEST(LegalizerInfoTest, MSVCDebugMiscompile) {
+  const LLT S1 = LLT::scalar(1);
+  const LLT P0 = LLT::pointer(0, 32);
+  LegalizerInfo LI;
+  auto Builder = LI.getActionDefinitionsBuilder(TargetOpcode::G_PTRTOINT);
+  (void)Builder.legalForCartesianProduct({S1}, {P0});
+}