[X86] Teach FCOPYSIGN lowering to recognize constant magnitudes.

[ab]

Summary

For code like:

float foo(float x) { return copysign(1.0, x); }

We used to generate:

andps        <-0.000000e+00,0,0,0>, %xmm0
movss        <1.000000e+00>, %xmm1
andps        <nan>, %xmm1
orps	        %xmm0, %xmm1

Basically doing an abs(1.0f) in the two middle instructions.

We now generate:

andps   <-0.000000e+00,0,0,0>, %xmm0
orps    <1.000000e+00,0,0,0>, %xmm0

[resistor]

Should this be target independent?

—Owen

[ab]

Ping?

• Ahmed

[resistor]

LGTM.

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
   Constant *C = ConstantVector::get(CV);
   SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
   SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
                               MachinePointerInfo::getConstantPool(),
                               false, false, false, 16);
   SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
 
   // Next, clear the sign bit from the first operand (magnitude).
+  // If it's a constant, we can clear it here.
+  if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
+    APFloat APF = Op0CN->getValueAPF();
+    // If the magnitude is a positive zero, the sign bit alone is enough.
+    if (APF.isPosZero())
+      return SignBit;
+    APF.clearSign();
+    CV[0] = ConstantFP::get(*Context, APF);
+  } else {
     CV[0] = ConstantFP::get(
-      *Context, APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
+        *Context,
+        APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
+  }
   C = ConstantVector::get(CV);
   CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
-  SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
-                              MachinePointerInfo::getConstantPool(),
-                              false, false, false, 16);
-  SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
+  SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
+  // If the magnitude operand wasn't a constant, we need to AND out the sign.
+  if (!isa<ConstantFPSDNode>(Op0))
+    Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
 
   // OR the magnitude value with the sign bit.
   return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
 }
 
 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
   SDValue N0 = Op.getOperand(0);
   SDLoc dl(Op);

llvm/trunk/test/CodeGen/X86/copysign-constant-magnitude.ll

+; RUN: llc < %s | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.10.0"
+
+define void @test_copysign_const_magnitude_d(double %X) {
+; CHECK: [[SIGNMASK:L.+]]:
+; CHECK-NEXT:   .quad -9223372036854775808    ## double -0.000000e+00
+; CHECK-NEXT:   .quad 0                       ## double 0.000000e+00
+; CHECK: [[ZERO:L.+]]:
+; CHECK-NEXT:   .space 16
+; CHECK: [[ONE:L.+]]:
+; CHECK-NEXT:   .quad 4607182418800017408     ## double 1.000000e+00
+; CHECK-NEXT:   .quad 0                       ## double 0.000000e+00
+; CHECK-LABEL: test_copysign_const_magnitude_d:
+
+; CHECK: id
+  %iX = call double @id_d(double %X)
+
+; CHECK-NEXT: andpd [[SIGNMASK]](%rip), %xmm0
+  %d0 = call double @copysign(double 0.000000e+00, double %iX)
+
+; CHECK-NEXT: id
+  %id0 = call double @id_d(double %d0)
+
+; CHECK-NEXT: andpd [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orpd [[ZERO]](%rip), %xmm0
+  %dn0 = call double @copysign(double -0.000000e+00, double %id0)
+
+; CHECK-NEXT: id
+  %idn0 = call double @id_d(double %dn0)
+
+; CHECK-NEXT: andpd [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orpd [[ONE]](%rip), %xmm0
+  %d1 = call double @copysign(double 1.000000e+00, double %idn0)
+
+; CHECK-NEXT: id
+  %id1 = call double @id_d(double %d1)
+
+; CHECK-NEXT: andpd [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orpd [[ONE]](%rip), %xmm0
+  %dn1 = call double @copysign(double -1.000000e+00, double %id1)
+
+; CHECK-NEXT: id
+  %idn1 = call double @id_d(double %dn1)
+
+; CHECK: retq
+  ret void
+}
+
+define void @test_copysign_const_magnitude_f(float %X) {
+; CHECK: [[SIGNMASK:L.+]]:
+; CHECK-NEXT:   .long 2147483648              ## float -0.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK: [[ZERO:L.+]]:
+; CHECK-NEXT:   .space 16
+; CHECK: [[ONE:L.+]]:
+; CHECK-NEXT:   .long 1065353216              ## float 1.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK-NEXT:   .long 0                       ## float 0.000000e+00
+; CHECK-LABEL: test_copysign_const_magnitude_f:
+
+; CHECK: id
+  %iX = call float @id_f(float %X)
+
+; CHECK-NEXT: andps [[SIGNMASK]](%rip), %xmm0
+  %d0 = call float @copysignf(float 0.000000e+00, float %iX)
+
+; CHECK-NEXT: id
+  %id0 = call float @id_f(float %d0)
+
+; CHECK-NEXT: andps [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orps [[ZERO]](%rip), %xmm0
+  %dn0 = call float @copysignf(float -0.000000e+00, float %id0)
+
+; CHECK-NEXT: id
+  %idn0 = call float @id_f(float %dn0)
+
+; CHECK-NEXT: andps [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orps [[ONE]](%rip), %xmm0
+  %d1 = call float @copysignf(float 1.000000e+00, float %idn0)
+
+; CHECK-NEXT: id
+  %id1 = call float @id_f(float %d1)
+
+; CHECK-NEXT: andps [[SIGNMASK]](%rip), %xmm0
+; CHECK-NEXT: orps [[ONE]](%rip), %xmm0
+  %dn1 = call float @copysignf(float -1.000000e+00, float %id1)
+
+; CHECK-NEXT: id
+  %idn1 = call float @id_f(float %dn1)
+
+; CHECK: retq
+  ret void
+}
+
+declare double @copysign(double, double) nounwind readnone
+declare float @copysignf(float, float) nounwind readnone
+
+; Dummy identity functions, so we always have xmm0, and prevent optimizations.
+declare double @id_d(double)
+declare float @id_f(float)

llvm/trunk/test/CodeGen/X86/copysign-zero.ll

-; RUN: llc < %s | not grep orpd
-; RUN: llc < %s | grep andpd | count 1
-
-target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
-target triple = "x86_64-apple-darwin8"
-
-define double @test(double %X) nounwind  {
-entry:
-  %tmp2 = tail call double @copysign( double 0.000000e+00, double %X ) nounwind readnone     ; <double> [#uses=1]
-  ret double %tmp2
-}
-
-declare double @copysign(double, double) nounwind readnone 
-