Revert "AMDGPU: Fix the double rounding issue in v2f64 -> v2f16 conversion (llvm#135659)"

This reverts commit 8b46b98.

Author: jrbyrnes

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -1054,7 +1054,8 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
 
   auto &FPTruncActions = getActionDefinitionsBuilder(G_FPTRUNC);
   if (ST.hasCvtPkF16F32Inst())
-    FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}});
+    FPTruncActions.legalFor(
+        {{S32, S64}, {S16, S32}, {V2S16, V2S32}, {V2S16, V2S64}});
   else
     FPTruncActions.legalFor({{S32, S64}, {S16, S32}});
   FPTruncActions.scalarize(0).lower();

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -915,6 +915,10 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
   }
 
+  if (Subtarget->hasCvtPkF16F32Inst()) {
+    setOperationAction(ISD::FP_ROUND, MVT::v2f16, Legal);
+  }
+
   setTargetDAGCombine({ISD::ADD,
                        ISD::UADDO_CARRY,
                        ISD::SUB,

llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll

@@ -654,9 +654,7 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16(
 ; GFX950-SDAG-NEXT:    s_waitcnt vmcnt(0)
 ; GFX950-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, v[2:3]
 ; GFX950-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
-; GFX950-SDAG-NEXT:    v_cvt_f16_f32_e32 v1, v2
-; GFX950-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; GFX950-SDAG-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX950-SDAG-NEXT:    v_cvt_pk_f16_f32 v0, v0, v2
 ; GFX950-SDAG-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GFX950-SDAG-NEXT:    s_endpgm
 ;
@@ -668,11 +666,11 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f16(
 ; GFX950-GISEL-NEXT:    s_mov_b32 s2, -1
 ; GFX950-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX950-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX950-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
-; GFX950-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
-; GFX950-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; GFX950-GISEL-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; GFX950-GISEL-NEXT:    v_pack_b32_f16 v0, v0, v1
+; GFX950-GISEL-NEXT:    v_mov_b64_e32 v[0:1], s[4:5]
+; GFX950-GISEL-NEXT:    v_mov_b64_e32 v[2:3], s[6:7]
+; GFX950-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, v[2:3]
+; GFX950-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX950-GISEL-NEXT:    v_cvt_pk_f16_f32 v0, v0, v2
 ; GFX950-GISEL-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX950-GISEL-NEXT:    s_endpgm
 ;