Christopher,
I'll answer my own question .. You're compressing the alignment to a
smaller value by extracting its log. I didn't get that until I read the
Reader.cpp code :)
Reid.
On Sat, 2007-04-21 at 09:46 -0700, Reid Spencer wrote:
> Christopher,
>
> On Sat, 2007-04-21 at 03:17 -0500, Christopher Lamb wrote:
> >
> > Changes in directory llvm/lib/Bytecode/Writer:
> >
> > Writer.cpp updated: 1.173 -> 1.174
> > ---
> > Log message:
> >
> >
> > add support for alignment attributes on load/store instructions
> >
> >
> > ---
> > Diffs of the changes: (+27 -5)
> >
> > Writer.cpp | 32 +++++++++++++++++++++++++++-----
> > 1 files changed, 27 insertions(+), 5 deletions(-)
> >
> >
> > Index: llvm/lib/Bytecode/Writer/Writer.cpp
> > diff -u llvm/lib/Bytecode/Writer/Writer.cpp:1.173
> > llvm/lib/Bytecode/Writer/Writer.cpp:1.174
> > --- llvm/lib/Bytecode/Writer/Writer.cpp:1.173 Mon Apr 16 18:32:28 2007
> > +++ llvm/lib/Bytecode/Writer/Writer.cpp Sat Apr 21 03:16:25 2007
> > @@ -616,7 +616,7 @@
> > unsigned Opcode = I.getOpcode();
> > unsigned NumOperands = I.getNumOperands();
> >
> > - // Encode 'tail call' as 61, 'volatile load' as 62, and 'volatile store'
> > as
> > + // Encode 'tail call' as 61
> > // 63.
> > if (const CallInst *CI = dyn_cast<CallInst>(&I)) {
> > if (CI->getCallingConv() == CallingConv::C) {
> > @@ -632,10 +632,6 @@
> > } else {
> > Opcode = 58; // Call escape sequence.
> > }
> > - } else if (isa<LoadInst>(I) && cast<LoadInst>(I).isVolatile()) {
> > - Opcode = 62;
> > - } else if (isa<StoreInst>(I) && cast<StoreInst>(I).isVolatile()) {
> > - Opcode = 63;
> > }
> >
> > // Figure out which type to encode with the instruction. Typically we
> > want
> > @@ -744,6 +740,32 @@
> > } else if (isa<InvokeInst>(I)) {
> > // Invoke escape seq has at least 4 operands to encode.
> > ++NumOperands;
> > + } else if (const LoadInst *LI = dyn_cast<LoadInst>(&I)) {
> > + // Encode attributed load as opcode 62
> > + // We need to encode the attributes of the load instruction as the
> > second
> > + // operand. Its not really a slot, but we don't want to break the
> > + // instruction format for these instructions.
> > + if (LI->getAlignment() || LI->isVolatile()) {
> > + NumOperands = 2;
> > + Slots[1] = ((Log2_32(LI->getAlignment())+1)<<1) +
> > + (LI->isVolatile() ? 1 : 0);
>
> Is this Log2_32 really needed? The alignment should already be a power
> of two. If the alignment is something like 7 then you'll get rounding
> to 8 but I contend that an alignment of 7 can't happen. This can be
> assured in the bcreader and the asmparser.
>
> > + if (Slots[1] > MaxOpSlot)
> > + MaxOpSlot = Slots[1];
> > + Opcode = 62;
> > + }
> > + } else if (const StoreInst *SI = dyn_cast<StoreInst>(&I)) {
> > + // Encode attributed store as opcode 63
> > + // We need to encode the attributes of the store instruction as the
> > third
> > + // operand. Its not really a slot, but we don't want to break the
> > + // instruction format for these instructions.
> > + if (SI->getAlignment() || SI->isVolatile()) {
> > + NumOperands = 3;
> > + Slots[2] = ((Log2_32(SI->getAlignment())+1)<<1) +
> > + (SI->isVolatile() ? 1 : 0);
> > + if (Slots[2] > MaxOpSlot)
> > + MaxOpSlot = Slots[2];
> > + Opcode = 63;
> > + }
> > }
> >
> > // Decide which instruction encoding to use. This is determined
> > primarily
> >
> >
> >
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits@cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits@cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
_______________________________________________
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
