OSDN Git Service

Revert r176154 in favor of a better approach.
authorBill Wendling <isanbard@gmail.com>
Fri, 8 Mar 2013 02:21:08 +0000 (02:21 +0000)
committerBill Wendling <isanbard@gmail.com>
Fri, 8 Mar 2013 02:21:08 +0000 (02:21 +0000)
commite36b47e17be0e3289890f33fe8aaecfc231d83a0
tree8fbc56e26546ca20a2a89b4bf166581d90d3f820
parent4434a4f8bdb817b29b59db88d682b722c5d76824
Revert r176154 in favor of a better approach.

Code generation makes some basic assumptions about the IR it's been given. In
particular, if there is only one 'invoke' in the function, then that invoke
won't be going away. However, with the advent of the `llvm.donothing' intrinsic,
those invokes may go away. If all of them go away, the landing pad no longer has
any users. This confuses the back-end, which asserts.

This happens with SjLj exceptions, because that's the model that modifies the IR
based on there being invokes, etc. in the function.

Remove any invokes of `llvm.donothing' during SjLj EH preparation. This will
give us a CFG that the back-end won't be confused about. If all of the invokes
in a function are removed, then the SjLj EH prepare pass won't insert the bogus
code the relies upon the invokes being there.
<rdar://problem/13228754&13316637>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@176677 91177308-0d34-0410-b5e6-96231b3b80d8
lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
lib/CodeGen/SjLjEHPrepare.cpp
test/CodeGen/ARM/invoke-donothing-assert.ll