Subzero: Be more strict about i1 calculations.

src/IceInstX8632.cpp

Index: src/IceInstX8632.cpp
diff --git a/src/IceInstX8632.cpp b/src/IceInstX8632.cpp
index 1454778f39907fa6f3a4de86ccb53c1b24d6db46..961f2fca9bac8939f062a4cba3677aa730f15c62 100644
--- a/src/IceInstX8632.cpp
+++ b/src/IceInstX8632.cpp
@@ -593,7 +593,8 @@ template <> void InstX8632Pblendvb::emit(const Cfg *Func) const {
 template <> void InstX8632Imul::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 2);
-  if (getDest()->getType() == IceType_i8) {
+  if (getDest()->getType() == IceType_i1 ||

[jvoung (off chromium), 2014/09/05 23:07:15]

I don't think arithmetic is allowed on i1, only and/or/xor bitwise ops:

"add, sub, mul, shl, udiv, sdiv, urem, srem, lshr, ashr

These arithmetic operations are disallowed on values of type i1."

See: test/Transforms/NaCl/promote-i1-ops.ll

How are you seeing these?

[Jim Stichnoth, 2014/09/05 23:46:48]

Interesting, I didn't realize we were simplifying those away, so, awesome! :)

I was hand-constructing the i1 tests for this CL.  The problem I actually saw,
that started it all, was bitcode that did an xor of an i1 with the true
constant, and Subzero lowered it as an xor with the immediate -1 (0xff) instead
of 1, which caused problems with a subsequent 8-bit compare against 0.

I'll take another look and see what parts are really needed given the PNaCl
simplfications.

+      getDest()->getType() == IceType_i8) {
     // The 8-bit version of imul only allows the form "imul r/m8".
     Variable *Src0 = llvm::dyn_cast<Variable>(getSrc(0));
     (void)Src0;