Add option to force filetype=asm for testing

src/IceTargetLoweringX8632.cpp

Index: src/IceTargetLoweringX8632.cpp
diff --git a/src/IceTargetLoweringX8632.cpp b/src/IceTargetLoweringX8632.cpp
index a09629c477d7f6f4032eeeca09580128c0464a1d..94e7286bed6d8a1f4f92f00e4c8a8b5d5c33233b 100644
--- a/src/IceTargetLoweringX8632.cpp
+++ b/src/IceTargetLoweringX8632.cpp
@@ -205,7 +205,11 @@ void TargetX8632::lowerCall(const InstCall *Instr) {
       break;
     case IceType_i1:
     case IceType_i8:
+      ReturnReg = makeReg(Dest->getType(), Traits::RegisterSet::Reg_al);
+      break;
     case IceType_i16:
+      ReturnReg = makeReg(Dest->getType(), Traits::RegisterSet::Reg_ax);
+      break;
     case IceType_i32:
       ReturnReg = makeReg(Dest->getType(), Traits::RegisterSet::Reg_eax);

[Jim Stichnoth, 2015/12/20 18:42:23]

Instead of the above, I think you could do something like:
  ReturnReg = makeReg(Dest->getType(), Traits::getGprForType(Dest->getType(),
Traits::RegisterSet::Reg_eax));

Also, it would probably be good to change all the Dest->getType() calls in this
method into DestTy=Dest->getType() .  And maybe do the same in the x8664 file
for symmetry.

[sehr, 2016/01/07 18:53:11]

Done.

       break;
@@ -342,6 +346,11 @@ void TargetX8632::lowerRet(const InstRet *Inst) {
       _fld(Src0);
     } else if (isVectorType(Src0->getType())) {
       Reg = legalizeToReg(Src0, Traits::RegisterSet::Reg_xmm0);
+    } else if (Src0->getType() == IceType_i8) {

[Jim Stichnoth, 2015/12/20 18:42:23]

Maybe check against i1 as well, for completeness

[sehr, 2016/01/07 18:53:11]

I converted these all to use the getGprForType scheme.

+      // TODO(sehr,stichnot): I think these need to sign/zero extend.

[Jim Stichnoth, 2015/12/20 18:42:23]

I don't think the x86-32 processor ABI has anything to say on this matter.

So it comes down to consistency between the caller and callee.  Subzero should
be self-consistent in this regard.

But if Subzero output is linked against another translator's output, such as for
bisection debugging, this would be an issue.

Given that returning i1/i8/i16 isn't even valid in a pexe, I think it would be
fine to explicitly zero-extend.  In fact, that could potentially save debugging
time for some future non-PNaCl-based integration when we have long forgotten
this issue.

[sehr, 2016/01/07 18:53:11]

I'm going with the idea that we're doing this for pexes and it's not valid, so I
remove the comment.

+      _mov(Reg, Src0, Traits::RegisterSet::Reg_al);
+    } else if (Src0->getType() == IceType_i16) {
+      _mov(Reg, Src0, Traits::RegisterSet::Reg_ax);
     } else {
       _mov(Reg, Src0, Traits::RegisterSet::Reg_eax);
     }