ARM updating VST2 pseudo-lowering fixed vs. register update.

rdar://10663487

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147876 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMExpandPseudoInsts.cpp b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
index 35f3c25500a6d6ea493dc5da679c8a4934fd6147..2de0fa60f324b048421ae18aa95625a046224f92 100644 (file)
--- a/lib/Target/ARM/ARMExpandPseudoInsts.cpp
+++ b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
@@ -314,7 +314,7 @@ static const NEONLdStTableEntry NEONLdStTable[] = {
 { ARM::VST2d16PseudoWB_fixed,   ARM::VST2d16wb_fixed, false, true, false,  SingleSpc,  2, 4 ,false},
 { ARM::VST2d16PseudoWB_register,   ARM::VST2d16wb_register, false, true, true,  SingleSpc,  2, 4 ,false},
 { ARM::VST2d32Pseudo,       ARM::VST2d32,      false, false, false, SingleSpc,  2, 2 ,false},
-{ ARM::VST2d32PseudoWB_fixed,   ARM::VST2d32wb_fixed, false, true, true,  SingleSpc,  2, 2 ,false},
+{ ARM::VST2d32PseudoWB_fixed,   ARM::VST2d32wb_fixed, false, true, false,  SingleSpc,  2, 2 ,false},
 { ARM::VST2d32PseudoWB_register,   ARM::VST2d32wb_register, false, true, true,  SingleSpc,  2, 2 ,false},
 { ARM::VST2d8Pseudo,        ARM::VST2d8,       false, false, false, SingleSpc,  2, 8 ,false},
 { ARM::VST2d8PseudoWB_fixed,    ARM::VST2d8wb_fixed, false, true, false,  SingleSpc,  2, 8 ,false},

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 7bd0bace9090e346c12f7fb617d1802148c7036d..d66d05ac730a6c09e4d89bf362667dbf69640fc5 100644 (file)
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1822,7 +1822,7 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
         Opc = getVLDSTRegisterUpdateOpcode(Opc);
       // We use a VST1 for v1i64 even if the pseudo says vld2/3/4, so
       // check for that explicitly too. Horribly hacky, but temporary.
-      if ((NumVecs != 1 && Opc != ARM::VST1q64PseudoWB_fixed) ||
+      if ((NumVecs > 2 && Opc != ARM::VST1q64PseudoWB_fixed) ||
           !isa<ConstantSDNode>(Inc.getNode()))
         Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
     }

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 2c564a7e8e6490e02e7ccd0bcd64a742029cf266..1e3410221e1f83081da3574f23b2756551fc15a9 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -1673,12 +1673,12 @@ defm VST2q8wb    : VST2QWB<{0,0,?,?}, "8">;
 defm VST2q16wb   : VST2QWB<{0,1,?,?}, "16">;
 defm VST2q32wb   : VST2QWB<{1,0,?,?}, "32">;
 
-def VST2d8PseudoWB_fixed     : VSTQWBPseudo<IIC_VST2u>;
-def VST2d16PseudoWB_fixed    : VSTQWBPseudo<IIC_VST2u>;
-def VST2d32PseudoWB_fixed    : VSTQWBPseudo<IIC_VST2u>;
-def VST2d8PseudoWB_register  : VSTQWBPseudo<IIC_VST2u>;
-def VST2d16PseudoWB_register : VSTQWBPseudo<IIC_VST2u>;
-def VST2d32PseudoWB_register : VSTQWBPseudo<IIC_VST2u>;
+def VST2d8PseudoWB_fixed     : VSTQWBfixedPseudo<IIC_VST2u>;
+def VST2d16PseudoWB_fixed    : VSTQWBfixedPseudo<IIC_VST2u>;
+def VST2d32PseudoWB_fixed    : VSTQWBfixedPseudo<IIC_VST2u>;
+def VST2d8PseudoWB_register  : VSTQWBregisterPseudo<IIC_VST2u>;
+def VST2d16PseudoWB_register : VSTQWBregisterPseudo<IIC_VST2u>;
+def VST2d32PseudoWB_register : VSTQWBregisterPseudo<IIC_VST2u>;
 
 def VST2q8PseudoWB_fixed     : VSTQQWBPseudo<IIC_VST2x2u>;
 def VST2q16PseudoWB_fixed    : VSTQQWBPseudo<IIC_VST2x2u>;

diff --git a/test/CodeGen/ARM/vst2.ll b/test/CodeGen/ARM/vst2.ll
index 915a84b677672285d663d14b7d14b56991286a65..497214316a9efb1875f8efda81965914f45e0a5e 100644 (file)
--- a/test/CodeGen/ARM/vst2.ll
+++ b/test/CodeGen/ARM/vst2.ll
@@ -110,6 +110,15 @@ define void @vst2Qf(float* %A, <4 x float>* %B) nounwind {
        ret void
 }
 
+define i8* @vst2update(i8* %out, <4 x i16>* %B) nounwind {
+;CHECK: vst2update
+;CHECK: vst2.16 {d16, d17}, [r0]!
+       %tmp1 = load <4 x i16>* %B
+       tail call void @llvm.arm.neon.vst2.v4i16(i8* %out, <4 x i16> %tmp1, <4 x i16> %tmp1, i32 2)
+       %t5 = getelementptr inbounds i8* %out, i32 16
+       ret i8* %t5
+}
+
 declare void @llvm.arm.neon.vst2.v8i8(i8*, <8 x i8>, <8 x i8>, i32) nounwind
 declare void @llvm.arm.neon.vst2.v4i16(i8*, <4 x i16>, <4 x i16>, i32) nounwind
 declare void @llvm.arm.neon.vst2.v2i32(i8*, <2 x i32>, <2 x i32>, i32) nounwind