Add NEON VLD1-lane instructions. Partial fix for Radar 8599955.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@117964 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMExpandPseudoInsts.cpp b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
index 53d2e9df126d4b7c2752a86e7a37b0476534bc2f..774324b4528d047b6a2efc929d5644d9522cd376 100644 (file)
--- a/lib/Target/ARM/ARMExpandPseudoInsts.cpp
+++ b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
@@ -110,6 +110,13 @@ namespace {
 }
 
 static const NEONLdStTableEntry NEONLdStTable[] = {
+{ ARM::VLD1LNq16Pseudo,     ARM::VLD1LNd16,     true, false, EvenDblSpc, 1, 4 },
+{ ARM::VLD1LNq16Pseudo_UPD, ARM::VLD1LNd16_UPD, true, false, EvenDblSpc, 1, 4 },
+{ ARM::VLD1LNq32Pseudo,     ARM::VLD1LNd32,     true, false, EvenDblSpc, 1, 2 },
+{ ARM::VLD1LNq32Pseudo_UPD, ARM::VLD1LNd32_UPD, true, false, EvenDblSpc, 1, 2 },
+{ ARM::VLD1LNq8Pseudo,      ARM::VLD1LNd8,      true, false, EvenDblSpc, 1, 8 },
+{ ARM::VLD1LNq8Pseudo_UPD,  ARM::VLD1LNd8_UPD,  true, false, EvenDblSpc, 1, 8 },
+
 { ARM::VLD1d64QPseudo,      ARM::VLD1d64Q,     true,  false, SingleSpc,  4, 1 },
 { ARM::VLD1d64QPseudo_UPD,  ARM::VLD1d64Q_UPD, true,  true,  SingleSpc,  4, 1 },
 { ARM::VLD1d64TPseudo,      ARM::VLD1d64T,     true,  false, SingleSpc,  3, 1 },
@@ -476,8 +483,9 @@ void ARMExpandPseudo::ExpandLaneOp(MachineBasicBlock::iterator &MBBI) {
     DstIsDead = MI.getOperand(OpIdx).isDead();
     DstReg = MI.getOperand(OpIdx++).getReg();
     GetDSubRegs(DstReg, RegSpc, TRI, D0, D1, D2, D3);
-    MIB.addReg(D0, RegState::Define | getDeadRegState(DstIsDead))
-      .addReg(D1, RegState::Define | getDeadRegState(DstIsDead));
+    MIB.addReg(D0, RegState::Define | getDeadRegState(DstIsDead));
+    if (NumRegs > 1)
+      MIB.addReg(D1, RegState::Define | getDeadRegState(DstIsDead));
     if (NumRegs > 2)
       MIB.addReg(D2, RegState::Define | getDeadRegState(DstIsDead));
     if (NumRegs > 3)
@@ -502,7 +510,9 @@ void ARMExpandPseudo::ExpandLaneOp(MachineBasicBlock::iterator &MBBI) {
   // Add the subregs as sources of the new instruction.
   unsigned SrcFlags = (getUndefRegState(MO.isUndef()) |
                        getKillRegState(MO.isKill()));
-  MIB.addReg(D0, SrcFlags).addReg(D1, SrcFlags);
+  MIB.addReg(D0, SrcFlags);
+  if (NumRegs > 1)
+    MIB.addReg(D1, SrcFlags);
   if (NumRegs > 2)
     MIB.addReg(D2, SrcFlags);
   if (NumRegs > 3)
@@ -943,6 +953,12 @@ bool ARMExpandPseudo::ExpandMBB(MachineBasicBlock &MBB) {
       ExpandVST(MBBI);
       break;
 
+    case ARM::VLD1LNq8Pseudo:
+    case ARM::VLD1LNq16Pseudo:
+    case ARM::VLD1LNq32Pseudo:
+    case ARM::VLD1LNq8Pseudo_UPD:
+    case ARM::VLD1LNq16Pseudo_UPD:
+    case ARM::VLD1LNq32Pseudo_UPD:
     case ARM::VLD2LNd8Pseudo:
     case ARM::VLD2LNd16Pseudo:
     case ARM::VLD2LNd32Pseudo:

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index eadc907529d51b6a762bb8e7db39eed554c7ecfc..ed14814dde873ed2611e69d0df9d76083987026b 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -421,6 +421,8 @@ def VLD4q8oddPseudo_UPD  : VLDQQQQWBPseudo<IIC_VLD4>;
 def VLD4q16oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4>;
 def VLD4q32oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4>;
 
+} // mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1
+
 // Classes for VLD*LN pseudo-instructions with multi-register operands.
 // These are expanded to real instructions after register allocation.
 class VLDQLNPseudo<InstrItinClass itin>
@@ -449,7 +451,46 @@ class VLDQQQQLNWBPseudo<InstrItinClass itin>
                  nohash_imm:$lane), itin, "$addr.addr = $wb, $src = $dst">;
 
 //   VLD1LN   : Vector Load (single element to one lane)
-//   FIXME: Not yet implemented.
+class VLD1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
+             PatFrag LoadOp>
+  : NLdSt<1, 0b10, op11_8, op7_4, (outs DPR:$dst),
+          (ins addrmode6:$addr, DPR:$src, nohash_imm:$lane),
+          IIC_VLD1ln, "vld1", Dt, "\\{$dst[$lane]\\}, $addr",
+          "$src = $dst",
+          [(set DPR:$dst, (vector_insert (Ty DPR:$src),
+                                         (i32 (LoadOp addrmode6:$addr)),
+                                         imm:$lane))]>;
+class VLD1QLNPseudo<ValueType Ty, PatFrag LoadOp> : VLDQLNPseudo<IIC_VLD1ln> {
+  let Pattern = [(set QPR:$dst, (vector_insert (Ty QPR:$src),
+                                               (i32 (LoadOp addrmode6:$addr)),
+                                               imm:$lane))];
+}
+
+def VLD1LNd8  : VLD1LN<0b0000, {?,?,?,0}, "8", v8i8, extloadi8>;
+def VLD1LNd16 : VLD1LN<0b0100, {?,?,0,?}, "16", v4i16, extloadi16>;
+def VLD1LNd32 : VLD1LN<0b1000, {?,0,?,?}, "32", v2i32, load>;
+
+def VLD1LNq8Pseudo  : VLD1QLNPseudo<v16i8, extloadi8>;
+def VLD1LNq16Pseudo : VLD1QLNPseudo<v8i16, extloadi16>;
+def VLD1LNq32Pseudo : VLD1QLNPseudo<v4i32, load>;
+
+let mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1 in {
+
+// ...with address register writeback:
+class VLD1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
+  : NLdSt<1, 0b10, op11_8, op7_4, (outs DPR:$dst, GPR:$wb),
+          (ins addrmode6:$addr, am6offset:$offset,
+           DPR:$src, nohash_imm:$lane), IIC_VLD1lnu, "vld1", Dt,
+          "\\{$dst[$lane]\\}, $addr$offset",
+          "$src = $dst, $addr.addr = $wb", []>;
+
+def VLD1LNd8_UPD  : VLD1LNWB<0b0000, {?,?,?,0}, "8">;
+def VLD1LNd16_UPD : VLD1LNWB<0b0100, {?,?,0,?}, "16">;
+def VLD1LNd32_UPD : VLD1LNWB<0b1000, {?,0,?,?}, "32">;
+
+def VLD1LNq8Pseudo_UPD  : VLDQLNWBPseudo<IIC_VLD1lnu>;
+def VLD1LNq16Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD1lnu>;
+def VLD1LNq32Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD1lnu>;
 
 //   VLD2LN   : Vector Load (single 2-element structure to one lane)
 class VLD2LN<bits<4> op11_8, bits<4> op7_4, string Dt>

diff --git a/lib/Target/ARM/ARMSchedule.td b/lib/Target/ARM/ARMSchedule.td
index 61489d64f376f25f1793e64edf84aee51d8262dc..7f4d2bbd44a76a8dd12e2785483de96b13144b89 100644 (file)
--- a/lib/Target/ARM/ARMSchedule.td
+++ b/lib/Target/ARM/ARMSchedule.td
@@ -134,6 +134,8 @@ def IIC_VLD1u      : InstrItinClass;
 def IIC_VLD1x2u    : InstrItinClass;
 def IIC_VLD1x3u    : InstrItinClass;
 def IIC_VLD1x4u    : InstrItinClass;
+def IIC_VLD1ln     : InstrItinClass;
+def IIC_VLD1lnu    : InstrItinClass;
 def IIC_VLD2       : InstrItinClass;
 def IIC_VLD2x2     : InstrItinClass;
 def IIC_VLD2u      : InstrItinClass;

diff --git a/lib/Target/ARM/ARMScheduleA8.td b/lib/Target/ARM/ARMScheduleA8.td
index 1f19b21e1c33cc0a423b808fd110aba1d3f2b466..b5c8def42f9f02fa07ba7947ce7899742e84ad8f 100644 (file)
--- a/lib/Target/ARM/ARMScheduleA8.td
+++ b/lib/Target/ARM/ARMScheduleA8.td
@@ -457,6 +457,18 @@ def CortexA8Itineraries : ProcessorItineraries<
                                InstrStage<3, [A8_LSPipe]>],
                               [2, 2, 3, 3, 2, 1]>,
   //
+  // VLD1ln
+  InstrItinData<IIC_VLD1ln,   [InstrStage<1, [A8_Pipe0, A8_Pipe1]>,
+                               InstrStage<3, [A8_NLSPipe], 1>,
+                               InstrStage<3, [A8_LSPipe]>],
+                              [3, 1, 1, 1]>,
+  //
+  // VLD1lnu
+  InstrItinData<IIC_VLD1lnu,  [InstrStage<1, [A8_Pipe0, A8_Pipe1]>,
+                               InstrStage<3, [A8_NLSPipe], 1>,
+                               InstrStage<3, [A8_LSPipe]>],
+                              [3, 2, 1, 1, 1, 1]>,
+  //
   // VLD2
   InstrItinData<IIC_VLD2,     [InstrStage<1, [A8_Pipe0, A8_Pipe1]>,
                                InstrStage<2, [A8_NLSPipe], 1>,

diff --git a/lib/Target/ARM/ARMScheduleA9.td b/lib/Target/ARM/ARMScheduleA9.td
index 3096b0ad99026e3ac996ee4cdea49ecc33ca8b68..c78f59383f2161ecbdbefbd86b921f972d959a7e 100644 (file)
--- a/lib/Target/ARM/ARMScheduleA9.td
+++ b/lib/Target/ARM/ARMScheduleA9.td
@@ -787,6 +787,24 @@ def CortexA9Itineraries : ProcessorItineraries<
                                InstrStage<3, [A9_LSUnit]>],
                               [2, 2, 3, 3, 2, 1]>,
   //
+  // VLD1ln
+  InstrItinData<IIC_VLD1ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
+                               InstrStage<1, [A9_MUX0], 0>,
+                               InstrStage<1, [A9_DRegsN],   0, Required>,
+                               InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
+                               InstrStage<3, [A9_NPipe], 1>,
+                               InstrStage<3, [A9_LSUnit]>],
+                              [4, 1, 1, 1]>,
+  //
+  // VLD1lnu
+  InstrItinData<IIC_VLD1lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
+                               InstrStage<1, [A9_MUX0], 0>,
+                               InstrStage<1, [A9_DRegsN],   0, Required>,
+                               InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
+                               InstrStage<3, [A9_NPipe], 1>,
+                               InstrStage<3, [A9_LSUnit]>],
+                              [4, 2, 1, 1, 1, 1]>,
+  //
   // VLD2
   InstrItinData<IIC_VLD2,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
                                InstrStage<1, [A9_MUX0], 0>,

diff --git a/test/CodeGen/ARM/vldlane.ll b/test/CodeGen/ARM/vldlane.ll
index 3a7bafe4d658c0941c4ac964fce3591abffe2f54..57e74dfe2498bc7e4244caaaf87e177c691e07dc 100644 (file)
--- a/test/CodeGen/ARM/vldlane.ll
+++ b/test/CodeGen/ARM/vldlane.ll
@@ -1,5 +1,32 @@
 ; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s
 
+define <8 x i8> @vld1lanei8(i8* %A, <8 x i8>* %B) nounwind {
+;CHECK: vld1lanei8:
+;CHECK: vld1.8 {d16[3]}, [r0]
+       %tmp1 = load <8 x i8>* %B
+       %tmp2 = load i8* %A, align 1
+       %tmp3 = insertelement <8 x i8> %tmp1, i8 %tmp2, i32 3
+        ret <8 x i8> %tmp3
+}
+
+define <4 x i16> @vld1lanei16(i16* %A, <4 x i16>* %B) nounwind {
+;CHECK: vld1lanei16:
+;CHECK: vld1.16 {d16[2]}, [r0]
+       %tmp1 = load <4 x i16>* %B
+       %tmp2 = load i16* %A, align 2
+       %tmp3 = insertelement <4 x i16> %tmp1, i16 %tmp2, i32 2
+        ret <4 x i16> %tmp3
+}
+
+define <2 x i32> @vld1lanei32(i32* %A, <2 x i32>* %B) nounwind {
+;CHECK: vld1lanei32:
+;CHECK: vld1.32 {d16[1]}, [r0]
+       %tmp1 = load <2 x i32>* %B
+       %tmp2 = load i32* %A, align 4
+       %tmp3 = insertelement <2 x i32> %tmp1, i32 %tmp2, i32 1
+        ret <2 x i32> %tmp3
+}
+
 %struct.__neon_int8x8x2_t = type { <8 x i8>,  <8 x i8> }
 %struct.__neon_int16x4x2_t = type { <4 x i16>, <4 x i16> }
 %struct.__neon_int32x2x2_t = type { <2 x i32>, <2 x i32> }