[NVPTX] Fix lowering of fp16 ISD::FNEG.

[tra]

Summary

There's no neg.f16 instruction, so negation has to be done via subtraction from zero.

[tra]

use CHECK-DAG for the instructions that could be generated in different order.

lib/Target/NVPTX/NVPTXISelLowering.cpp

if (!STI.allowFP16Math()) {
     // hardware, only sm_53 and sm_60 have full implementation. Others
     // only have token amount of hardware and are likely to run faster
     // by using fp32 units instead.
     setOperationAction(ISD::FADD, MVT::f16, Promote);
     setOperationAction(ISD::FMUL, MVT::f16, Promote);
     setOperationAction(ISD::FSUB, MVT::f16, Promote);
     setOperationAction(ISD::FMA, MVT::f16, Promote);
   }
+  // There's no neg.f16 instruction.
+  setOperationAction(ISD::FNEG, MVT::f16, Expand);
 
   // Library functions.  These default to Expand, but we have instructions
   // for them.
   setOperationAction(ISD::FCEIL,  MVT::f16, Legal);
   setOperationAction(ISD::FCEIL,  MVT::f32, Legal);
   setOperationAction(ISD::FCEIL,  MVT::f64, Legal);
   setOperationAction(ISD::FFLOOR, MVT::f16, Legal);
   setOperationAction(ISD::FFLOOR, MVT::f32, Legal);

test/CodeGen/NVPTX/f16-instructions.ll

 ; CHECK-NOF16-NEXT: cvt.rn.f16.f32 [[R:%h[0-9]+]], [[R32]]
 ; CHECK-NEXT: st.param.b16    [func_retval0+0], [[R]];
 ; CHECK-NEXT: ret;
 define half @test_fsub(half %a, half %b) #0 {
   %r = fsub half %a, %b
   ret half %r
 }
 
+; CHECK-LABEL: test_fneg(
+; CHECK-DAG:  ld.param.b16    [[A:%h[0-9]+]], [test_fneg_param_0];
+; CHECK-F16-NEXT:   mov.b16        [[Z:%h[0-9]+]], 0x0000
+; CHECK-F16-NEXT:   sub.rn.f16     [[R:%h[0-9]+]], [[Z]], [[A]];
+; CHECK-NOF16-DAG:  cvt.f32.f16    [[A32:%f[0-9]+]], [[A]]
+; CHECK-NOF16-NEXT: mov.f32        [[Z:%f[0-9]+]], 0f00000000;
+; CHECK-NOF16-NEXT: sub.rn.f32     [[R32:%f[0-9]+]], [[Z]], [[A32]];
+; CHECK-NOF16-NEXT: cvt.rn.f16.f32 [[R:%h[0-9]+]], [[R32]]
+; CHECK-NEXT: st.param.b16    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define half @test_fneg(half %a) #0 {
+  %r = fsub half 0.0, %a
+  ret half %r
+}
+
 ; CHECK-LABEL: test_fmul(
 ; CHECK-DAG:  ld.param.b16    [[A:%h[0-9]+]], [test_fmul_param_0];
 ; CHECK-DAG:  ld.param.b16    [[B:%h[0-9]+]], [test_fmul_param_1];
 ; CHECK-F16-NEXT: mul.rn.f16      [[R:%h[0-9]+]], [[A]], [[B]];
 ; CHECK-NOF16-DAG:  cvt.f32.f16    [[A32:%f[0-9]+]], [[A]]
 ; CHECK-NOF16-DAG:  cvt.f32.f16    [[B32:%f[0-9]+]], [[B]]
 ; CHECK-NOF16-NEXT: mul.rn.f32     [[R32:%f[0-9]+]], [[A32]], [[B32]];
 ; CHECK-NOF16-NEXT: cvt.rn.f16.f32 [[R:%h[0-9]+]], [[R32]]