[X86][FP16] Only generate approximate rsqrt when Reciprocal is true for half type

We have reasonable fast sqrt and accurate rsqrt for half type due to the
limited fractions. So neither do we need multi steps refinement for
rsqrt nor replace sqrt by rsqrt.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D114844

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b82a74b..62b2387 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -23190,6 +23190,10 @@ static SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
 bool X86TargetLowering::isFsqrtCheap(SDValue Op, SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
 
+  // We don't need to replace SQRT with RSQRT for half type.
+  if (VT.getScalarType() == MVT::f16)
+    return true;
+
   // We never want to use both SQRT and RSQRT instructions for the same input.
   if (DAG.getNodeIfExists(X86ISD::FRSQRT, DAG.getVTList(VT), Op))
     return false;
@@ -23236,6 +23240,7 @@ SDValue X86TargetLowering::getSqrtEstimate(SDValue Op,
 
   if (VT.getScalarType() == MVT::f16 && isTypeLegal(VT) &&
       Subtarget.hasFP16()) {
+    assert(Reciprocal && "Don't replace SQRT with RSQRT for half type");
     if (RefinementSteps == ReciprocalEstimate::Unspecified)
       RefinementSteps = 0;
 

diff --git a/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll b/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
index 34f1391..6a5c3e2 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
@@ -123,12 +123,7 @@ define half @test_sqrt_sh2(half %a0, half %a1) {
 define half @test_sqrt_sh3(half %a0, half %a1) {
 ; CHECK-LABEL: test_sqrt_sh3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
-; CHECK-NEXT:    vpand %xmm1, %xmm0, %xmm1
-; CHECK-NEXT:    vcmpltsh {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %k1
-; CHECK-NEXT:    vrsqrtsh %xmm0, %xmm0, %xmm0
-; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vmovsh %xmm1, %xmm0, %xmm0 {%k1}
+; CHECK-NEXT:    vsqrtsh %xmm0, %xmm0, %xmm0
 ; CHECK-NEXT:    retq
   %1 = call fast half @llvm.sqrt.f16(half %a0)
   ret half %1

diff --git a/llvm/test/CodeGen/X86/avx512fp16vl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512fp16vl-intrinsics.ll
index d3af80c..b103b23 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512fp16vl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512fp16vl-intrinsics.ll
@@ -972,10 +972,7 @@ define <8 x half> @test_sqrt_ph_128_fast(<8 x half> %a0, <8 x half> %a1) {
 define <8 x half> @test_sqrt_ph_128_fast2(<8 x half> %a0, <8 x half> %a1) {
 ; CHECK-LABEL: test_sqrt_ph_128_fast2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
-; CHECK-NEXT:    vpand %xmm1, %xmm0, %xmm1
-; CHECK-NEXT:    vcmpgeph {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %xmm1, %k1
-; CHECK-NEXT:    vrsqrtph %xmm0, %xmm0 {%k1} {z}
+; CHECK-NEXT:    vsqrtph %xmm0, %xmm0
 ; CHECK-NEXT:    retq
   %1 = call fast <8 x half> @llvm.sqrt.v8f16(<8 x half> %a0)
   ret <8 x half> %1