Create type IDs based on reference counts.

test/NaCl/Bitcode/struct-types.ll

Index: test/NaCl/Bitcode/struct-types.ll
diff --git a/test/NaCl/Bitcode/struct-types.ll b/test/NaCl/Bitcode/struct-types.ll
new file mode 100644
index 0000000000000000000000000000000000000000..39b095f010afaf0ac27956fc9ef877b5219a4751
--- /dev/null
+++ b/test/NaCl/Bitcode/struct-types.ll
@@ -0,0 +1,27 @@
+; Checks if llvm bitcode defines a struct type before the pointer type,
+; even if the struct definintion appears after the pointer type, while
+; pnacl bitcode moves the pointer before the struct.
+; RUN: llvm-as < %s | llvm-bcanalyzer -dump | FileCheck %s -check-prefix=LLVM
+; RUN: llvm-as < %s | pnacl-freeze | pnacl-bcanalyzer -dump | FileCheck %s -check-prefix=PNACL
+
+%1 = type %0*
+%0 = type {i8}

[jvoung (off chromium), 2013/05/15 21:01:36]

How come %0 isn't on line 7 and %1 on line 8?

I guess I was imagining you would have:

%typeA = ...
%typeB = ...
%typeC = type {%typeA *, %typeB *}
%typeD = ...

and have the ranking be:

typeC is most used
typeD is next most used
typeA and typeB are not used directly.

then that would test that you don't recurse into typeC and start putting typeA
and typeB before typeD.

[Karl, 2013/05/20 21:48:51]

I created a more complicated example. Hope this is what you are looking for.

+; LLVM: <TYPE_BLOCK_ID {{.*}}>
+; LLVM:    <STRUCT_NAMED {{.*}}/>
+; LLVM:    <POINTER {{.*}}/>
+; LLVM: </TYPE_BLOCK_ID>
+
+; PNACL: <TYPE_BLOCK_ID {{.*}}>
+; PNACL:    <POINTER {{.*}}/>
+; PNACL:    <STRUCT_NAMED {{.*}}/>
+; PNACL: </TYPE_BLOCK_ID>
+
+define %1 @foo(%1 %a)
+{
+  ret %1 %a;
+}
+
+define %1 @bar(%1 %b)
+{
+  ret %1 %b;
+}