[llvm-dev] FAIL: Why does the output of "llc -march=cpp .." fail to compile with clang 3.8

[James Courtier-Dutton via llvm-dev]

It seems that -march=cpp is broken.
the -march=cpp  is very useful for people trying to understand the builder API.
But, if the output of -march=cpp  does not even compile, it obviously
creates code that is broken, so defeats the object of helping people
understand the builder API.

Why isn't -march=cpp  tested at all?

See detail below.

Kind Regards

James



Take a very simple example:

Source code for  test.c:
void munge(int *P) {
  P[0] = P[1] + P[2];
}

compile with:
clang -c -emit-llvm -o test.bc test.c
test that the .bc output looks sensible
llvm-dis-3.8 test.bc
<snip>
; Function Attrs: nounwind uwtable
define void @_Z5mungePi(i32* %P) #0 {
  %1 = alloca i32*, align 8
  store i32* %P, i32** %1, align 8
  %2 = load i32*, i32** %1, align 8
  %3 = getelementptr inbounds i32, i32* %2, i64 1
  %4 = load i32, i32* %3, align 4
<snip>


So, that looks good.
compile to -march=cpp
llc-3.8 -march=cpp -o test.cpp test.bc

So, we should now end up with some CPP code that uses the builder
interface to generate the .bc file.
The problem is, test.cpp  does not even compile.

the -march=cpp  is very useful for people trying to understand the builder API.

But, if the output of -march=cpp  does not even compile, it obviously
creates code that is broken, so defeats the object of helping people
understand the builder API.

Why isn't -march=cpp  tested at all?