Add support for undef values in ICE IR.

tests_lit/llvm2ice_tests/undef.ll

+; This test checks that undef values are represented as registers.
+
+; RUN: %llvm2ice -O2 --verbose none %s | FileCheck %s
+; RUN: %llvm2ice --verbose none %s | FileCheck --check-prefix=OPTM1 %s
+; RUN: %llvm2ice --verbose none %s | FileCheck --check-prefix=ERRORS %s
+; RUN: %llvm2iceinsts %s | %szdiff %s | FileCheck --check-prefix=DUMP %s
+; RUN: %llvm2iceinsts --pnacl %s | %szdiff %s \
+; RUN:                           | FileCheck --check-prefix=DUMP %s
+
+define i32 @return() {

[Jim Stichnoth, 2014/06/17 06:28:08]

For both of these tests, add i64 and floating-point versions as well.

+entry:
+; CHECK: return:

[jvoung (off chromium), 2014/06/17 15:50:11]

Should we start using CHECK-LABEL more?

http://llvm.org/docs/CommandGuide/FileCheck.html#the-check-label-directive

It's a bit more robust.

[Jim Stichnoth, 2014/06/17 17:51:12]

I agree - most of the tests lack this simply because I was unaware of it.

+; OPTM1: return:
+  ret i32 undef
+; CHECK-NOT: mov eax, dword ptr {{.*}}
+; OPTM1-NOT: mov eax, dword ptr {{.*}}
+; CHECK: ret
+; OPTM1: ret
+}
+
+define i32 @arith(i32 %arg) {
+entry:
+; CHECK: arith:
+  %__1 = add i32 undef, %arg

[Jim Stichnoth, 2014/06/17 06:28:08]

I think it could be worthwhile to test undef as the first operand, versus the
second operand, versus both operands.  Also, use a non-commutative operator like
sub just in case future changes to lowering code try to improve the sequence by
commuting.

+; CHECK: add e{{[a-z]+}}, e{{[a-z]+}}
+  ret i32 %__1
+; CHECK: ret
+}

[jvoung (off chromium), 2014/06/17 15:50:11]

Sometimes function call arguments are "undef". In that case it could use the
uninitialized stack slot instead of the an uninitialized register... but that's
probably not worth avoiding going to a register first, since undef shouldn't be
too common.

[wala, 2014/06/17 20:00:19]

There are a number of cases that could be optimized. For instance, when
immediate values are permitted in the instruction, the undef could be lowered to
an immediate constant rather than a register.

The code was designed to lower undefs to a register because, the current way
operand legalization is implemented in Subzero, legalizing to a register is
always permissible. Legalization to a register is also one of the simplest
approaches to legalization.

I suspect that undefs will be too rare in practice that lowering to a register
will cause a performance hit, although the way that undefs are legalized could
always be refined later.

+
+; ERRORS-NOT: ICE translation error
+; DUMP-NOT: SZ