[clang][codegen] Don't assert on CurLinkModule for silenced diagnostic

[inglorion]

Summary

Fixes PR51564.

Under certain circumstances, it is possible to enter Clang's
BackendConsumer::DiagnosticHandlerImpl without CurLinkModule set.
For diagnostics of kind DK_Linker, this results in an assert.
However, these diagnostics are never actually surfaced, unless
they are errors. Crashing the compiler for diagnostics we don't
actually surface seems not worth it, so this change moves the
assert below the check so that it only triggers for diagnostics
we actually surface.

[inglorion]

There are some possible follow-ups here:

1. Making sure CurLinkModule is set in all (or at least more) cases.
2. Addressing the FIXME about warnings and notes.

I have some partial implementations of those ideas, but I figure we can take this change first while I figure out the other bits.

[xur]

This looks fine to me. But I agree you that a more complete fix would be having CurLinkModule set for all callers.

clang/lib/CodeGen/CodeGenAction.cpp

case llvm::DK_SrcMgr:
     SrcMgrDiagHandler(cast<DiagnosticInfoSrcMgr>(DI));
     return;
   case llvm::DK_StackSize:
     if (StackSizeDiagHandler(cast<DiagnosticInfoStackSize>(DI)))
       return;
     ComputeDiagID(Severity, backend_frame_larger_than, DiagID);
     break;
   case DK_Linker:
-    assert(CurLinkModule);
     // FIXME: stop eating the warnings and notes.
     if (Severity != DS_Error)
       return;
+    assert(CurLinkModule);
     DiagID = diag::err_fe_cannot_link_module;
     break;
   case llvm::DK_OptimizationRemark:
     // Optimization remarks are always handled completely by this
     // handler. There is no generic way of emitting them.
     OptimizationRemarkHandler(cast<OptimizationRemark>(DI));
     return;
   case llvm::DK_OptimizationRemarkMissed:

clang/test/CodeGen/Inputs/linker-diagnostic1.ll

+target triple = "armv4-none-unknown-eabi"
+target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
+
+declare i32 @foo(i32)
+
+define i32 @bar(i32 %x) {
+  %1 = tail call i32 @foo(i32 %x)
+  ret i32 %1
+}

clang/test/CodeGen/linker-diagnostic.ll

+; RUN: mkdir -p %t
+; RUN: opt -module-summary -o %t/foo.o %s
+; RUN: opt -module-summary -o %t/bar.o %S/Inputs/linker-diagnostic1.ll
+; RUN: llvm-lto2 run --thinlto-distributed-indexes -r %t/foo.o,foo,plx -r %t/bar.o,bar,plx \
+; RUN:   -r %t/bar.o,foo, -o %t/foobar.so %t/foo.o %t/bar.o
+; RUN: %clang -c -o %t/lto.bar.o --target=armv6-none-unknown-eabi -O2 \
+; RUN:   -fthinlto-index=%t/bar.o.thinlto.bc %t/bar.o
+
+target triple = "thumbv6-unknown-linux-gnueabihf"
+target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
+
+define i32 @foo(i32 %x) {
+  %1 = add i32 %x, 1
+  ret i32 %1
+}