PNaCl: Implement computed gotos (indirectbr) by expanding out to switches

PNaCl: Implement computed gotos (indirectbr) by expanding out to switches

Add the ExpandIndirectBr pass for expanding out indirectbr and
blockaddress into switch instructions, so that programs using computed
gotos can work with PNaCl's stable ABI.

BUG=https://code.google.com/p/nativeclient/issues/detail?id=3695
TEST=*.ll tests + GCC torture tests

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-llvm.git;a=commit;h=a55919f

[Derek Schuff, 2014-01-06 22:30:22]

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
File lib/Transforms/NaCl/ExpandIndirectBr.cpp (right):

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
lib/Transforms/NaCl/ExpandIndirectBr.cpp:128: // indirectbr, replace them with
dummy values.
Why do this rather than just destroying them?

[Mark Seaborn, 2014-01-06 22:36:01]

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
File lib/Transforms/NaCl/ExpandIndirectBr.cpp (right):

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
lib/Transforms/NaCl/ExpandIndirectBr.cpp:128: // indirectbr, replace them with
dummy values.
On 2014/01/06 22:30:22, Derek Schuff wrote:
> Why do this rather than just destroying them?

Because they might be used, so you can't just destroy them.  The blockaddress
has to evaluate to something.

e.g.
void *foo() {
  return &&label;
label:
  return 0;
}

This might happen in less silly code if the indirectbrs are deleted by dead code
elimination.

[Derek Schuff, 2014-01-07 00:35:01]

lgtm

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
File lib/Transforms/NaCl/ExpandIndirectBr.cpp (right):

https://codereview.chromium.org/120773002/diff/60001/lib/Transforms/NaCl/Expa...
lib/Transforms/NaCl/ExpandIndirectBr.cpp:128: // indirectbr, replace them with
dummy values.
On 2014/01/06 22:36:02, Mark Seaborn wrote:
> On 2014/01/06 22:30:22, Derek Schuff wrote:
> > Why do this rather than just destroying them?
> 
> Because they might be used, so you can't just destroy them.  The blockaddress
> has to evaluate to something.
> 
> e.g.
> void *foo() {
>   return &&label;
> label:
>   return 0;
> }
> 
> This might happen in less silly code if the indirectbrs are deleted by dead
code
> elimination.

Ah, i didn't think there about legitimate uses other than indirectbr

[Mark Seaborn, 2014-01-07 10:35:58]

Committed patchset #2 manually as ra55919f (presubmit successful).

[Mark Seaborn, 2014-01-07 20:09:00]

Hmm, this seems to be causing a test failure on Windows in llvm_regression:

2014-01-07 09:42:22.722221 ******************** TEST 'LLVM ::
Transforms/NaCl/expand-indirectbr.ll' FAILED ********************
...
2014-01-07 09:42:22.722221 Exit Code: 2
2014-01-07 09:42:22.722221 Command Output (stderr):
2014-01-07 09:42:22.722221 --
2014-01-07 09:42:22.722221 Stack dump:
2014-01-07 09:42:22.722221 0.	Program arguments:
/cygdrive/c/b/build/slave/win-pnacl-x86_32/build/native_client/pnacl/build/llvm_x86_32/Release+Asserts/bin/opt
/cygdrive/c/b/build/slave/win-pnacl-x86_32/build/native_client/pnacl/git/llvm/test/Transforms/NaCl/expand-indirectbr.ll
-expand-indirectbr -S 
2014-01-07 09:42:22.722221 1.	Running pass 'Expand out indirectbr and
blockaddress (computed gotos)' on module
'/cygdrive/c/b/build/slave/win-pnacl-x86_32/build/native_client/pnacl/git/llvm/test/Transforms/NaCl/expand-indirectbr.ll'.
2014-01-07 09:42:22.722221 FileCheck error: '-' is empty.

(from
http://build.chromium.org/p/client.nacl.toolchain/builders/win-pnacl-x86_32/b...)

Not sure what is going on here.  run_computed_goto_test in Scons does pass,
though, so hopefully this doesn't block other people at the moment?