Change the NEON VDUPfdf and VDUPfqf pseudo-instructions to actually be
pseudo instructions.
llvm-svn: 115840
diff --git a/llvm/lib/Target/ARM/ARMExpandPseudoInsts.cpp b/llvm/lib/Target/ARM/ARMExpandPseudoInsts.cpp
index ecf4aee..6755487 100644
--- a/llvm/lib/Target/ARM/ARMExpandPseudoInsts.cpp
+++ b/llvm/lib/Target/ARM/ARMExpandPseudoInsts.cpp
@@ -17,6 +17,7 @@
#define DEBUG_TYPE "arm-pseudo"
#include "ARM.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMRegisterInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/Target/TargetRegisterInfo.h"
@@ -710,6 +711,31 @@
MI.eraseFromParent();
break;
}
+ case ARM::VDUPfqf:
+ case ARM::VDUPfdf:{
+ unsigned NewOpc = Opcode == ARM::VDUPfqf ? ARM::VDUPLNfq : ARM::VDUPLNfd;
+ MachineInstrBuilder MIB =
+ BuildMI(MBB, MBBI, MI.getDebugLoc(), TII->get(NewOpc));
+ unsigned OpIdx = 0;
+ unsigned SrcReg = MI.getOperand(1).getReg();
+ unsigned Lane = getARMRegisterNumbering(SrcReg) & 1;
+ unsigned DReg = TRI->getMatchingSuperReg(SrcReg,
+ Lane & 1 ? ARM::ssub_1 : ARM::ssub_0, &ARM::DPR_VFP2RegClass);
+ // The lane is [0,1] for the containing DReg superregister.
+ // Copy the dst/src register operands.
+ MIB.addOperand(MI.getOperand(OpIdx++));
+ MIB.addReg(DReg);
+ ++OpIdx;
+ // Add the lane select operand.
+ MIB.addImm(Lane);
+ // Add the predicate operands.
+ MIB.addOperand(MI.getOperand(OpIdx++));
+ MIB.addOperand(MI.getOperand(OpIdx++));
+
+ TransferImpOps(MI, MIB, MIB);
+ MI.eraseFromParent();
+ break;
+ }
case ARM::VLD1q8Pseudo:
case ARM::VLD1q16Pseudo:
diff --git a/llvm/lib/Target/ARM/ARMInstrNEON.td b/llvm/lib/Target/ARM/ARMInstrNEON.td
index eaf2f79..e3ff884 100644
--- a/llvm/lib/Target/ARM/ARMInstrNEON.td
+++ b/llvm/lib/Target/ARM/ARMInstrNEON.td
@@ -3609,14 +3609,9 @@
(DSubReg_i32_reg imm:$lane))),
(SubReg_i32_lane imm:$lane)))>;
-def VDUPfdf : N2V<0b11, 0b11, {?,1}, {0,0}, 0b11000, 0, 0,
- (outs DPR:$dst), (ins SPR:$src),
- IIC_VMOVD, "vdup", "32", "$dst, ${src:lane}", "",
+def VDUPfdf : PseudoNeonI<(outs DPR:$dst), (ins SPR:$src), IIC_VMOVD, "",
[(set DPR:$dst, (v2f32 (NEONvdup (f32 SPR:$src))))]>;
-
-def VDUPfqf : N2V<0b11, 0b11, {?,1}, {0,0}, 0b11000, 1, 0,
- (outs QPR:$dst), (ins SPR:$src),
- IIC_VMOVQ, "vdup", "32", "$dst, ${src:lane}", "",
+def VDUPfqf : PseudoNeonI<(outs QPR:$dst), (ins SPR:$src), IIC_VMOVD, "",
[(set QPR:$dst, (v4f32 (NEONvdup (f32 SPR:$src))))]>;
// VMOVN : Vector Narrowing Move