Merge fixed alloca stack adjustments into the prolog

src/IceTargetLowering.cpp

Index: src/IceTargetLowering.cpp
diff --git a/src/IceTargetLowering.cpp b/src/IceTargetLowering.cpp
index 0ee0afa2301929fa5e2596a709f0db2bd03e3886..5679b107591ac88224392069be6e275ce34280b4 100644
--- a/src/IceTargetLowering.cpp
+++ b/src/IceTargetLowering.cpp
@@ -363,7 +363,9 @@ void TargetLowering::getVarStackSlotParams(
   VarList SpilledVariables;
   for (Variable *Var : Variables) {
     if (Var->hasReg()) {
-      RegsUsed[Var->getRegNum()] = true;
+      // Rematerializable uses of esp/ebp do not count as uses for spilling.

[John, 2015/11/16 14:00:02]

I would rather say

... uses of stack/frame pointer do not ...

for this is an abstract, base class, and esp/ebp are target specific.

[Jim Stichnoth, 2015/11/16 14:47:57]

I don't understand what this change does.

The purpose of the Var->hasReg() check is to record the complete set of physical
registers used in the function, so that the prolog/epilog generation code can
filter them against the set of callee-save registers to get the set of registers
to save/restore.  It doesn't seem like this conditional should have any effect
on the outcome.  (But I haven't tried it, so I could be wrong...)

Also, it's better to say something like "stack/frame pointer" instead of
"esp/ebp" in this target-neutral file.

[sehr, 2015/11/16 18:42:24]

Well, it doesn't work to remove it, as I discovered.  The issue is as follows. 
This code keeps track of references to variables before (I think) register
allocating anything.  It's looking for physical register references to ebp and
noting that those would cause ebp to need to be spilled and restored.  With
rematerialization I create uses of ebp that are to the binding established
within the call frame, not outside it, so I don't want the spilled version. 
Make sense?

[sehr, 2015/11/16 18:42:24]

Done.

[Jim Stichnoth, 2015/11/16 22:47:18]

I spent some quality time with the code to understand what breaks when the
condition is removed.  In the end I agree that this is the right thing to do,
but I would change the comment to something like:

Don't consider a rematerializable variable to be an actual register use
(specifically of the frame pointer).  Otherwise, the prolog may decide to save
the frame pointer twice - once because of the explicit need for a frame pointer,
and once because of an active use of a callee-save register.

+      if( !Var->isRematerializable())

[Jim Stichnoth, 2015/11/16 14:47:57]

Please run clang-format...

[sehr, 2015/11/16 18:42:24]

Done.

+        RegsUsed[Var->getRegNum()] = true;
       continue;
     }
     // An argument either does not need a stack slot (if passed in a register)