Lines Matching full:andi
44 cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
45 cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
1273 // better to rotate, mask explicitly (using andi/andis), and then or the
1289 // Compute the masks for andi/andis that would be necessary.
1296 // group, using one of andi/andis can break even. If we have three
1297 // groups, we can use both andi and andis and break even (to use both
1298 // andi and andis we also need to or the results together). We need four
1299 // groups if we also need to rotate. To use andi/andis we need to do more
1303 // FIXME: We've biased here against using andi/andis, which is right for
1305 // andi/andis have single-cycle latency whereas the rotate-and-mask
1307 // andi/andis in break-even cases.
1376 // Take care of cases that should use andi/andis first.
1638 // We can use the 32-bit andi/andis technique if the mask does not
1642 // Compute the masks for andi/andis that would be necessary.
1672 // When we'd use andi/andis, we bias toward using the rotates (andi only
1742 // Take care of cases that should use andi/andis first.
1828 // We can use the 32-bit andi/andis technique if the mask does not
1832 // Compute the masks for andi/andis that would be necessary.
2714 // FIXME: Remove this once the ANDI glue bug is fixed:
2725 SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
2736 SDValue(AndI.getNode(), 1) /* glue */);
4011 // For ANDI and ANDIS, the higher-order bits are zero if either that is true
