From: Evan Cheng <evan.cheng@apple.com>
Date: Fri, 11 Feb 2011 02:28:55 +0000 (+0000)
Subject: Fix buggy fcopysign lowering.
X-Git-Url: http://plrg.eecs.uci.edu/git/?a=commitdiff_plain;h=c143dd4f63889ca6b4f656200f43a1fa7bbf1c34;p=oota-llvm.git

Fix buggy fcopysign lowering.
This
define float @foo(float %x, float %y) nounwind readnone {
entry:
  %0 = tail call float @copysignf(float %x, float %y) nounwind readnone
  ret float %0
}

Was compiled to:
    vmov     s0, r1
    bic      r0, r0, #-2147483648
    vmov     s1, r0
    vcmpe.f32    s0, #0
    vmrs         apsr_nzcv, fpscr
    it           lt
    vneglt.f32   s1, s1
    vmov         r0, s1
    bx           lr

This fails to copy the sign of -0.0f because it's lost during the float to int
conversion. Also, it's sub-optimal when the inputs are in GPR registers.

Now it uses integer and + or operations when it's profitable. And it's correct!
    lsrs    r1, r1, #31
    bfi     r0, r1, #31, #1
    bx      lr
rdar://8984306


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@125357 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 4afc8c7fac7..fe398602717 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -2833,12 +2833,46 @@ SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
   DebugLoc dl = Op.getDebugLoc();
   EVT VT = Op.getValueType();
   EVT SrcVT = Tmp1.getValueType();
-  SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
-  SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
-  SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
-  SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
+  bool F2IisFast = Subtarget->isCortexA9() ||
+    Tmp0.getOpcode() == ISD::BITCAST || Tmp0.getOpcode() == ARMISD::VMOVDRR;
+
+  // Bitcast operand 1 to i32.
+  if (SrcVT == MVT::f64)
+    Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
+                       &Tmp1, 1).getValue(1);
+  Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
+
+  // If float to int conversion isn't going to be super expensive, then simply
+  // or in the signbit.
+  if (F2IisFast) {
+    SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
+    SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
+    Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
+    if (VT == MVT::f32) {
+      Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
+                         DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
+      return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
+                         DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
+    }
+
+    // f64: Or the high part with signbit and then combine two parts.
+    Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
+                       &Tmp0, 1);
+    SDValue Lo = Tmp0.getValue(0);
+    SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
+    Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
+    return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
+  }
+
+  // Remove the signbit of operand 0.
+  Tmp0 = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
+
+  // If operand 1 signbit is one, then negate operand 0.
+  SDValue ARMcc;
+  SDValue Cmp = getARMCmp(Tmp1, DAG.getConstant(0, MVT::i32),
+                          ISD::SETLT, ARMcc, DAG, dl);
   SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
-  return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
+  return DAG.getNode(ARMISD::CNEG, dl, VT, Tmp0, Tmp0, ARMcc, CCR, Cmp);
 }
 
 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
diff --git a/test/CodeGen/ARM/fcopysign.ll b/test/CodeGen/ARM/fcopysign.ll
index a6d741087a8..1050cd26599 100644
--- a/test/CodeGen/ARM/fcopysign.ll
+++ b/test/CodeGen/ARM/fcopysign.ll
@@ -1,18 +1,45 @@
-; RUN: llc < %s -march=arm | grep bic | count 2
-; RUN: llc < %s -march=arm -mattr=+v6,+vfp2 | \
-; RUN:   grep vneg | count 2
+; RUN: llc < %s -mtriple=armv7-apple-darwin -mcpu=cortex-a8 | FileCheck %s -check-prefix=SOFT
+; RUN: llc < %s -mtriple=armv7-gnueabi -float-abi=hard -mcpu=cortex-a8 | FileCheck %s -check-prefix=HARD
 
-define float @test1(float %x, double %y) {
-	%tmp = fpext float %x to double
-	%tmp2 = tail call double @copysign( double %tmp, double %y )
-	%tmp3 = fptrunc double %tmp2 to float
-	ret float %tmp3
+; rdar://8984306
+define float @test1(float %x, float %y) nounwind {
+entry:
+; SOFT: test1:
+; SOFT: lsr r1, r1, #31
+; SOFT: bfi r0, r1, #31, #1
+
+; HARD: test1:
+; HARD: vabs.f32 d0, d0
+; HARD: cmp r0, #0
+; HARD: vneglt.f32 s0, s0
+  %0 = tail call float @copysignf(float %x, float %y) nounwind
+  ret float %0
+}
+
+define double @test2(double %x, double %y) nounwind {
+entry:
+; SOFT: test2:
+; SOFT: lsr r2, r3, #31
+; SOFT: bfi r1, r2, #31, #1
+
+; HARD: test2:
+; HARD: vabs.f64 d0, d0
+; HARD: cmp r1, #0
+; HARD: vneglt.f64 d0, d0
+  %0 = tail call double @copysign(double %x, double %y) nounwind
+  ret double %0
 }
 
-define double @test2(double %x, float %y) {
-	%tmp = fpext float %y to double
-	%tmp2 = tail call double @copysign( double %x, double %tmp )
-	ret double %tmp2
+define double @test3(double %x, double %y, double %z) nounwind {
+entry:
+; SOFT: test3:
+; SOFT: vabs.f64
+; SOFT: cmp {{.*}}, #0
+; SOFT: vneglt.f64
+  %0 = fmul double %x, %y
+  %1 = tail call double @copysign(double %0, double %z) nounwind
+  ret double %1
 }
 
-declare double @copysign(double, double)
+declare double @copysign(double, double) nounwind
+declare float @copysignf(float, float) nounwind