Fixes bug on conditional branch where the targets are the same.

tests_lit/llvm2ice_tests/cond-br-same-target.ll

Index: tests_lit/llvm2ice_tests/cond-br-same-target.ll
diff --git a/tests_lit/llvm2ice_tests/cond-br-same-target.ll b/tests_lit/llvm2ice_tests/cond-br-same-target.ll
new file mode 100644
index 0000000000000000000000000000000000000000..77ecae0703e5630e34d71b553f92cc613d1c9428
--- /dev/null
+++ b/tests_lit/llvm2ice_tests/cond-br-same-target.ll
@@ -0,0 +1,32 @@
+; Test conditional branch where targets are the same.
+; Tests issue: https://code.google.com/p/nativeclient/issues/detail?id=4212
+
+; RUN: %p2i -i %s --args -O2 --verbose inst -threads=0 | FileCheck %s

[Jim Stichnoth, 2015/06/25 21:26:29]

You should be able to use "--insts" instead of "--verbose inst -threads=0".  (It
also adds -notranslate, which is actually fine given comments below.)

Also, it would be nice to add an Om1 run, to demonstrate that this is a
fundamental transformation that doesn't depend on the optimization level.

BTW, it occurs to me that doing such a transformation in a high-level ICE
instruction's ctor, is kind of not kosher, because we mean for a Cfg::dump()
called on a freshly parsed function to output pretty much the same as the .ll
file input.  But in this example, we dump an unconditional branch for a
conditional branch input.  Oh well.

[Karl, 2015/06/25 22:33:42]

Done.

+
+define void @f(i32 %foo, i32 %bar) {
+  %c = icmp ult i32 %foo, %bar
+  br i1 %c, label %block, label %block
+block:
+  ret void
+}
+
+; The following shows that the input is parsed correctly.

[Jim Stichnoth, 2015/06/25 21:26:29]

Actually, it shows that conditional has already become unconditional.

[Karl, 2015/06/25 22:33:42]

Correct. The test now shows it.

+
+; CHECK:      ================ Initial CFG ================

[Jim Stichnoth, 2015/06/25 21:26:29]

Probably better not to check for this header, which I don't consider part of the
"stable" dump output.

[Karl, 2015/06/25 22:33:42]

Removed.

+; CHECK-NEXT: define void @f(i32 %foo, i32 %bar) {
+; CHECK-NEXT: __0:
+; CHECK-NEXT:   %c = icmp ult i32 %foo, %bar
+; CHECK-NEXT:   br label %block
+; CHECK-NEXT: block:
+; CHECK-NEXT:   ret void
+; CHECK-NEXT: }
+
+; The following shows the conditional branch is converted to unconditional.

[Jim Stichnoth, 2015/06/25 21:26:30]

Given my earlier comment, I don't think this latter check is needed.  In fact,
the very act of dumping the initial CFG is enough to trigger the assertion
failure.

[Karl, 2015/06/25 22:33:42]

Removed.

+
+; CHECK:      ================ After x86 address mode opt ================
+; CHECK-NEXT: define void @f(i32 %foo, i32 %bar) {
+; CHECK-NEXT: __0:
+; CHECK-NEXT:   br label %block
+; CHECK-NEXT: block:
+; CHECK-NEXT:   ret void
+; CHECK-NEXT: }