[DRAFT] Added handling for AVX512DQ+AVX512VL+AVX512F to lowerFPToIntToFP

[KavinTheG]

Fixes #160111

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Kavin Gnanapandithan (KavinTheG)

Changes

Progress on issue llvm#160111

---

Full diff: https://github.com/llvm/llvm-project/pull/160582.diff

1 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+39-10)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2feb76e0eb7b4..b98f190f6915b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -19897,21 +19897,50 @@ static SDValue lowerFPToIntToFP(SDValue CastToFP, const SDLoc &DL,
   // See if we have 128-bit vector cast instructions for this type of cast.
   // We need cvttps2dq/cvttpd2dq and cvtdq2ps/cvtdq2pd.
   if (!Subtarget.hasSSE2() || (VT != MVT::f32 && VT != MVT::f64) ||
-      IntVT != MVT::i32)
+      (IntVT != MVT::i32 && IntVT != MVT::i64))
     return SDValue();
 
   unsigned SrcSize = SrcVT.getSizeInBits();
   unsigned IntSize = IntVT.getSizeInBits();
   unsigned VTSize = VT.getSizeInBits();
-  MVT VecSrcVT = MVT::getVectorVT(SrcVT, 128 / SrcSize);
-  MVT VecIntVT = MVT::getVectorVT(IntVT, 128 / IntSize);
-  MVT VecVT = MVT::getVectorVT(VT, 128 / VTSize);
-
-  // We need target-specific opcodes if this is v2f64 -> v4i32 -> v2f64.
-  unsigned ToIntOpcode =
-      SrcSize != IntSize ? X86ISD::CVTTP2SI : (unsigned)ISD::FP_TO_SINT;
-  unsigned ToFPOpcode =
-      IntSize != VTSize ? X86ISD::CVTSI2P : (unsigned)ISD::SINT_TO_FP;
+  unsigned ToIntOpcode, ToFPOpcode;
+  unsigned Width = 128;
+  bool IsUnsigned = CastToInt.getOpcode() == ISD::FP_TO_UINT;
+
+  if (Subtarget.hasAVX512()) {
+    if (Subtarget.hasVLX())  {
+      // AVX512VL could handle for FP_TO_UINT/UINT_TO_FP (f64/32 <-> i32) AVX512F as well but Width = 512
+      if (IntVT == MVT::i32) {
+        ToIntOpcode = IsUnsigned ? X86ISD::CVTTP2UI : X86ISD::CVTTP2SI;  
+        ToFPOpcode = IsUnsigned ? X86ISD::CVTUI2P : X86ISD::CVTSI2P;  
+      } else {
+        ToIntOpcode = CastToInt.getOpcode();
+        ToFPOpcode = IsUnsigned ? ISD::UINT_TO_FP : ISD::SINT_TO_FP;
+      } 
+    } else if (Subtarget.hasDQI() && IntVT == MVT::i64) {
+      // AVX512DQ + AVX512VL could handle f64/32 <-> i64 SINT & UINT
+      ToIntOpcode = CastToInt.getOpcode(); 
+      ToFPOpcode = IsUnsigned ? ISD::UINT_TO_FP : ISD::SINT_TO_FP;
+    } else {
+      // AVX512F
+      Width = 512;
+      ToIntOpcode = CastToInt.getOpcode(); 
+      ToFPOpcode = IsUnsigned ? ISD::UINT_TO_FP : ISD::SINT_TO_FP;
+    }
+  } else {
+    if (IntVT != MVT::i32 || IsUnsigned)
+      return SDValue();
+    // We need target-specific opcodes if this is v2f64 -> v4i32 -> v2f64.
+    ToIntOpcode =
+        SrcSize != IntSize ? X86ISD::CVTTP2SI : (unsigned)ISD::FP_TO_SINT;
+    ToFPOpcode =
+        IntSize != VTSize ? X86ISD::CVTSI2P : (unsigned)ISD::SINT_TO_FP;
+  }
+
+  MVT VecSrcVT = MVT::getVectorVT(SrcVT, Width / SrcSize);
+  MVT VecIntVT = MVT::getVectorVT(IntVT, Width / IntSize);
+  MVT VecVT = MVT::getVectorVT(VT, Width / VTSize);
+
 
   // sint_to_fp (fp_to_sint X) --> extelt (sint_to_fp (fp_to_sint (s2v X))), 0
   //

[RKSimon]

You'd be better off moving the hasDQI logic in here to simplify below:

!(IntVT == MVT::i32 || (IntVT == MVT::i64 && Subtarget.hasDQI())))