From 5689f67ef75ea0b927e0bb3c3cda9078e9bf100f Mon Sep 17 00:00:00 2001
From: Ahmed Bougacha <ahmed.bougacha@gmail.com>
Date: Thu, 13 Aug 2015 01:32:30 +0000
Subject: [PATCH] [CodeGen] Mark the promoted FCOPYSIGN result FP_ROUND as
 TRUNCating.

Now that we can properly promote mismatched FCOPYSIGNs (r244858), we
can mark the FP_ROUND on the result as truncating, to expose folding.

FCOPYSIGN doesn't change anything but the sign bit, so
  (fp_round (fcopysign (fpext a), b))
is equivalent to (modulo the sign bit):
  (fp_round (fpext a))
which is a no-op.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@244862 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/SelectionDAG/LegalizeDAG.cpp |  9 ++++++++-
 test/CodeGen/AArch64/f16-instructions.ll | 15 +++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 3b15752e2e4..552f92bbd46 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -4301,8 +4301,15 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
     Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
     Tmp2 = Node->getOperand(1);
     Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
+
+    // fcopysign doesn't change anything but the sign bit, so
+    //   (fp_round (fcopysign (fpext a), b))
+    // is as precise as
+    //   (fp_round (fpext a))
+    // which is a no-op. Mark it as a TRUNCating FP_ROUND.
+    const bool isTrunc = (Node->getOpcode() == ISD::FCOPYSIGN);
     Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
-                                  Tmp3, DAG.getIntPtrConstant(0, dl)));
+                                  Tmp3, DAG.getIntPtrConstant(isTrunc, dl)));
     break;
   }
   case ISD::FFLOOR:
diff --git a/test/CodeGen/AArch64/f16-instructions.ll b/test/CodeGen/AArch64/f16-instructions.ll
index 99dc8e33014..3dd54e64623 100644
--- a/test/CodeGen/AArch64/f16-instructions.ll
+++ b/test/CodeGen/AArch64/f16-instructions.ll
@@ -702,6 +702,21 @@ define half @test_copysign_f64(half %a, double %b) #0 {
   ret half %r
 }
 
+; Check that the FP promotion will use a truncating FP_ROUND, so we can fold
+; away the (fpext (fp_round <result>)) here.
+
+; CHECK-LABEL: test_copysign_extended:
+; CHECK-NEXT: fcvt s1, h1
+; CHECK-NEXT: fcvt s0, h0
+; CHECK-NEXT: movi.4s v2, #0x80, lsl #24
+; CHECK-NEXT: bit.16b v0, v1, v2
+; CHECK-NEXT: ret
+define float @test_copysign_extended(half %a, half %b) #0 {
+  %r = call half @llvm.copysign.f16(half %a, half %b)
+  %xr = fpext half %r to float
+  ret float %xr
+}
+
 ; CHECK-LABEL: test_floor:
 ; CHECK-NEXT: fcvt [[FLOAT32:s[0-9]+]], h0
 ; CHECK-NEXT: frintm [[INT32:s[0-9]+]], [[FLOAT32]]
-- 
2.34.1