From d616c642bb02ceef5ae0abf7de74195c0457ab4d Mon Sep 17 00:00:00 2001
From: James Molloy <james.molloy@arm.com>
Date: Tue, 11 Aug 2015 12:06:28 +0000
Subject: [PATCH] [ARM] Match fminnan/fmaxnan for vector vmin/vmax instead of
 an intrinsic

Lower Intrinsic::arm_neon_vmins/vmaxs to fminnan/fmaxnan and match that instead. This is important because SDAG will soon be able to select FMINNAN itself, so we need a unified lowering path for intrinsics and SDAG.

NFCI.

llvm-svn: 244593
---
 llvm/lib/Target/ARM/ARMISelLowering.cpp | 16 ++++++++++++++++
 llvm/lib/Target/ARM/ARMInstrNEON.td     |  8 ++++----
 2 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index ecd439f..d796894 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -957,6 +957,12 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::FMINNAN, MVT::f64, Legal);
     setOperationAction(ISD::FMAXNAN, MVT::f64, Legal);
   }
+  if (Subtarget->hasNEON()) {
+    setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
+    setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
+    setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
+    setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
+  }
 
   // We have target-specific dag combine patterns for the following nodes:
   // ARMISD::VMOVRRD  - No need to call setTargetDAGCombine
@@ -2803,6 +2809,16 @@ ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
     return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
                        Op.getOperand(1), Op.getOperand(2));
   }
+  case Intrinsic::arm_neon_vmins:
+  case Intrinsic::arm_neon_vmaxs: {
+    // v{min,max}s is overloaded between signed integers and floats.
+    if (!Op.getValueType().isFloatingPoint())
+      return SDValue();
+    unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
+      ? ISD::FMINNAN : ISD::FMAXNAN;
+    return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
+                       Op.getOperand(1), Op.getOperand(2));
+  }
   }
 }
 
diff --git a/llvm/lib/Target/ARM/ARMInstrNEON.td b/llvm/lib/Target/ARM/ARMInstrNEON.td
index d7741a9..22b5547 100644
--- a/llvm/lib/Target/ARM/ARMInstrNEON.td
+++ b/llvm/lib/Target/ARM/ARMInstrNEON.td
@@ -5032,10 +5032,10 @@ defm VMAXu    : N3VInt_QHS<1, 0, 0b0110, 0, N3RegFrm,
                            "vmax", "u", int_arm_neon_vmaxu, 1>;
 def  VMAXfd   : N3VDInt<0, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VBIND,
                         "vmax", "f32",
-                        v2f32, v2f32, int_arm_neon_vmaxs, 1>;
+                        v2f32, v2f32, fmaxnan, 1>;
 def  VMAXfq   : N3VQInt<0, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VBINQ,
                         "vmax", "f32",
-                        v4f32, v4f32, int_arm_neon_vmaxs, 1>;
+                        v4f32, v4f32, fmaxnan, 1>;
 
 // VMAXNM
 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
@@ -5058,10 +5058,10 @@ defm VMINu    : N3VInt_QHS<1, 0, 0b0110, 1, N3RegFrm,
                            "vmin", "u", int_arm_neon_vminu, 1>;
 def  VMINfd   : N3VDInt<0, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VBIND,
                         "vmin", "f32",
-                        v2f32, v2f32, int_arm_neon_vmins, 1>;
+                        v2f32, v2f32, fminnan, 1>;
 def  VMINfq   : N3VQInt<0, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VBINQ,
                         "vmin", "f32",
-                        v4f32, v4f32, int_arm_neon_vmins, 1>;
+                        v4f32, v4f32, fminnan, 1>;
 
 // VMINNM
 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
-- 
2.7.4