[llvm-commits] CVS: llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp
Nate Begeman
natebegeman at mac.com
Fri Oct 22 17:50:34 PDT 2004
Changes in directory llvm/lib/Target/PowerPC:
PPC32ISelSimple.cpp updated: 1.96 -> 1.97
---
Log message:
Kill casts from integer types to unsigned byte, when the cast was only used
as the shift amount operand to a shift instruction. This was causing us to
emit unnecessary clear operations for code such as:
int foo(int x) { return 1 << x; }
---
Diffs of the changes: (+13 -0)
Index: llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp
diff -u llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.96 llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.97
--- llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.96 Fri Oct 22 16:02:08 2004
+++ llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp Fri Oct 22 19:50:23 2004
@@ -3014,6 +3014,19 @@
unsigned DestReg = getReg(CI);
MachineBasicBlock::iterator MI = BB->end();
+ // If this is a cast from an integer type to a ubyte, with one use where the
+ // use is the shift amount argument of a shift instruction, just emit a move
+ // instead (since the shift instruction will only look at the low 5 bits
+ // regardless of how it is sign extended)
+ if (CI.getType() == Type::UByteTy && SrcClass <= cInt && CI.hasOneUse()) {
+ ShiftInst *SI = dyn_cast<ShiftInst>(*(CI.use_begin()));
+ if (SI && (SI->getOperand(1) == &CI)) {
+ unsigned SrcReg = getReg(Op, BB, MI);
+ BuildMI(*BB, MI, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg);
+ return;
+ }
+ }
+
// If this is a cast from an byte, short, or int to an integer type of equal
// or lesser width, and all uses of the cast are store instructions then dont
// emit them, as the store instruction will implicitly not store the zero or
More information about the llvm-commits
mailing list