This adds an extra word to LoadInst and StoreInst. Since
alignments must be a power of two, I suggest representing them as
a shift amount (e.g, an alignment of 16 is stored as 4, 1<<4 ==
16). Given this, we don't need a large number of bits to support
very large shift amounts. Given this, you can store it in the
"SubClassData" member of Value (which is 16 bits wide), to avoid
using the word. Load and StoreInst currently store a bit in that
word. The only trick here is that we want power-of-two-plus one,
so that "0" represents an alignment of zero, not an alignment of 1.
Yeah, I thought about that. It's not a bit problem, I'll encode the
alignment the same way it's encoded in the BC which has this same
issue.
sounds good.
Please don't #include MathExtras.h in Instructions.h
No prob. I'll have to move the accessor methods to the .cpp file
and include MathExtras there to use the Log2_32 function, though.
Ok. The "getter" should be inlined (it just uses <<, nothing exotic
like log :).
+ LoadInst(Value *Ptr, const std::string &Name, bool isVolatile =
false,
+ Instruction *InsertBefore = 0, unsigned Align = 0);
The instruction* to insert before should be the last member. In
the case of alignment, I don't think we need a helper constructor
for this (in fact, we probably don't need one for volatile
either). Any clients that want to create a load instruction with
a specific alignment can create it, then use setAlignment() :)
These helper constructors were used quite prolifically in the
lowering passes, like the hunk below. Do you prefer them to be
changed to set the attribute directly?
All of the codegen stuff looks good and should stay as you wrote it,
I'm just talking about the LLVM level stuff. If you think that
having an alignment-taking-ctor is useful, feel free to keep it, but
move the instruction-to-insert-before to the end of the operand list.
-Chris
_______________________________________________
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
