[X86][AVX10.2] Map vector saturated converts to public intrinsics (llvm#121483)

JaydeepChauhan14 · web-flow · commit 36dd42135579 · 2025-01-05T10:23:37.000+08:00
We already have support for saturated convert ISA in llvm. With this
patch we mapped public llvm intrinsic onto saturated convert ISA. It
includes support for float, double into sign and unsigned int.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -341,8 +341,17 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     }
   }
   if (Subtarget.hasAVX10_2()) {
-    setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i32, Legal);
-    setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i32, Legal);
+    setOperationAction(ISD::FP_TO_UINT_SAT, MVT::v2i32, Custom);
+    setOperationAction(ISD::FP_TO_SINT_SAT, MVT::v2i32, Custom);
+    for (MVT VT : {MVT::i32, MVT::v4i32, MVT::v8i32, MVT::v16i32, MVT::v2i64,
+                   MVT::v4i64}) {
+      setOperationAction(ISD::FP_TO_UINT_SAT, VT, Legal);
+      setOperationAction(ISD::FP_TO_SINT_SAT, VT, Legal);
+    }
+    if (Subtarget.hasAVX10_2_512()) {
+      setOperationAction(ISD::FP_TO_UINT_SAT, MVT::v8i64, Legal);
+      setOperationAction(ISD::FP_TO_SINT_SAT, MVT::v8i64, Legal);
+    }
     if (Subtarget.is64Bit()) {
       setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Legal);
       setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Legal);
@@ -2656,6 +2665,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
                        ISD::UINT_TO_FP,
                        ISD::STRICT_SINT_TO_FP,
                        ISD::STRICT_UINT_TO_FP,
+                       ISD::FP_TO_SINT_SAT,
+                       ISD::FP_TO_UINT_SAT,
                        ISD::SETCC,
                        ISD::MUL,
                        ISD::XOR,
@@ -33665,6 +33676,26 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
     }
     return;
   }
+  case ISD::FP_TO_SINT_SAT:
+  case ISD::FP_TO_UINT_SAT: {
+    if (!Subtarget.hasAVX10_2())
+      return;
+
+    bool IsSigned = Opc == ISD::FP_TO_SINT_SAT;
+    EVT VT = N->getValueType(0);
+    SDValue Op = N->getOperand(0);
+    EVT OpVT = Op.getValueType();
+    SDValue Res;
+
+    if (VT == MVT::v2i32 && OpVT == MVT::v2f64) {
+      if (IsSigned)
+        Res = DAG.getNode(X86ISD::FP_TO_SINT_SAT, dl, MVT::v4i32, Op);
+      else
+        Res = DAG.getNode(X86ISD::FP_TO_UINT_SAT, dl, MVT::v4i32, Op);
+      Results.push_back(Res);
+    }
+    return;
+  }
   case ISD::FP_TO_SINT:
   case ISD::STRICT_FP_TO_SINT:
   case ISD::FP_TO_UINT:
@@ -34645,6 +34676,8 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(VPERMV3)
   NODE_NAME_CASE(VPERMI)
   NODE_NAME_CASE(VPTERNLOG)
+  NODE_NAME_CASE(FP_TO_SINT_SAT)
+  NODE_NAME_CASE(FP_TO_UINT_SAT)
   NODE_NAME_CASE(VFIXUPIMM)
   NODE_NAME_CASE(VFIXUPIMM_SAE)
   NODE_NAME_CASE(VFIXUPIMMS)
@@ -56202,6 +56235,33 @@ static SDValue combineSIntToFP(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Custom handling for VCVTTPS2QQS/VCVTTPS2UQQS
+static SDValue combineFP_TO_xINT_SAT(SDNode *N, SelectionDAG &DAG,
+                                     const X86Subtarget &Subtarget) {
+  if (!Subtarget.hasAVX10_2())
+    return SDValue();
+
+  bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT_SAT;
+  EVT SrcVT = N->getOperand(0).getValueType();
+  EVT DstVT = N->getValueType(0);
+  SDLoc dl(N);
+
+  if (SrcVT == MVT::v2f32 && DstVT == MVT::v2i64) {
+    SDValue V2F32Value = DAG.getUNDEF(SrcVT);
+
+    // Concatenate the original v2f32 input and V2F32Value to create v4f32
+    SDValue NewSrc = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
+                                 N->getOperand(0), V2F32Value);
+
+    // Select the FP_TO_SINT_SAT/FP_TO_UINT_SAT node
+    if (IsSigned)
+      return DAG.getNode(X86ISD::FP_TO_SINT_SAT, dl, MVT::v2i64, NewSrc);
+
+    return DAG.getNode(X86ISD::FP_TO_UINT_SAT, dl, MVT::v2i64, NewSrc);
+  }
+  return SDValue();
+}
+
 static bool needCarryOrOverflowFlag(SDValue Flags) {
   assert(Flags.getValueType() == MVT::i32 && "Unexpected VT!");
 
@@ -59315,6 +59375,8 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::INTRINSIC_WO_CHAIN:  return combineINTRINSIC_WO_CHAIN(N, DAG, DCI);
   case ISD::INTRINSIC_W_CHAIN:  return combineINTRINSIC_W_CHAIN(N, DAG, DCI);
   case ISD::INTRINSIC_VOID:  return combineINTRINSIC_VOID(N, DAG, DCI);
+  case ISD::FP_TO_SINT_SAT:
+  case ISD::FP_TO_UINT_SAT: return combineFP_TO_xINT_SAT(N, DAG, Subtarget);
     // clang-format on
   }
 
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -908,6 +908,10 @@ namespace llvm {
     // Load x87 FPU environment from memory.
     FLDENVm,
 
+    // Custom handling for FP_TO_xINT_SAT
+    FP_TO_SINT_SAT,
+    FP_TO_UINT_SAT,
+
     /// This instruction implements FP_TO_SINT with the
     /// integer destination in memory and a FP reg source.  This corresponds
     /// to the X86::FIST*m instructions and the rounding mode change stuff. It
diff --git a/llvm/lib/Target/X86/X86InstrAVX10.td b/llvm/lib/Target/X86/X86InstrAVX10.td
@@ -834,6 +834,70 @@ let Predicates = [HasAVX10_2] in {
 // patterns have been disabled with null_frag.
 // Patterns VCVTTPD2DQSZ128
 
+// VCVTTPD2DQS
+def : Pat<(v4i32(X86fp2sisat(v2f64 VR128X:$src))),
+          (VCVTTPD2DQSZ128rr VR128X:$src)>;
+def : Pat<(v4i32(fp_to_sint_sat(v4f64 VR256X:$src), i32)),
+          (VCVTTPD2DQSZ256rr VR256X:$src)>;
+def : Pat<(v8i32(fp_to_sint_sat(v8f64 VR512:$src), i32)),
+          (VCVTTPD2DQSZrr VR512:$src)>;
+
+// VCVTTPD2QQS
+def : Pat<(v2i64(fp_to_sint_sat(v2f64 VR128X:$src), i64)),
+          (VCVTTPD2QQSZ128rr VR128X:$src)>;
+def : Pat<(v4i64(fp_to_sint_sat(v4f64 VR256X:$src), i64)),
+          (VCVTTPD2QQSZ256rr VR256X:$src)>;
+def : Pat<(v8i64(fp_to_sint_sat(v8f64 VR512:$src), i64)),
+          (VCVTTPD2QQSZrr VR512:$src)>;
+
+// VCVTTPD2UDQS
+def : Pat<(v4i32(X86fp2uisat(v2f64 VR128X:$src))),
+          (VCVTTPD2UDQSZ128rr VR128X:$src)>;
+def : Pat<(v4i32(fp_to_uint_sat(v4f64 VR256X:$src), i32)),
+          (VCVTTPD2UDQSZ256rr VR256X:$src)>;
+def : Pat<(v8i32(fp_to_uint_sat(v8f64 VR512:$src), i32)),
+          (VCVTTPD2UDQSZrr VR512:$src)>;
+
+// VCVTTPD2UQQS
+def : Pat<(v2i64(fp_to_uint_sat(v2f64 VR128X:$src), i64)),
+          (VCVTTPD2UQQSZ128rr VR128X:$src)>;
+def : Pat<(v4i64(fp_to_uint_sat(v4f64 VR256X:$src), i64)),
+          (VCVTTPD2UQQSZ256rr VR256X:$src)>;
+def : Pat<(v8i64(fp_to_uint_sat(v8f64 VR512:$src), i64)),
+          (VCVTTPD2UQQSZrr VR512:$src)>;
+
+// VCVTTPS2DQS
+def : Pat<(v4i32(fp_to_sint_sat(v4f32 VR128X:$src), i32)),
+          (VCVTTPS2DQSZ128rr VR128X:$src)>;
+def : Pat<(v8i32(fp_to_sint_sat(v8f32 VR256X:$src), i32)),
+          (VCVTTPS2DQSZ256rr VR256X:$src)>;
+def : Pat<(v16i32(fp_to_sint_sat(v16f32 VR512:$src), i32)),
+          (VCVTTPS2DQSZrr VR512:$src)>;
+
+// VCVTTPS2QQS
+def : Pat<(v2i64(X86fp2sisat(v4f32 VR128X:$src))),
+          (VCVTTPS2QQSZ128rr VR128X:$src)>;
+def : Pat<(v4i64(fp_to_sint_sat(v4f32 VR128X:$src), i64)),
+          (VCVTTPS2QQSZ256rr VR128X:$src)>;
+def : Pat<(v8i64(fp_to_sint_sat(v8f32 VR256X:$src), i64)),
+          (VCVTTPS2QQSZrr VR256X:$src)>;
+
+// VCVTTPS2UDQS
+def : Pat<(v4i32(fp_to_uint_sat(v4f32 VR128X:$src), i32)),
+          (VCVTTPS2UDQSZ128rr VR128X:$src)>;
+def : Pat<(v8i32(fp_to_uint_sat(v8f32 VR256X:$src), i32)),
+          (VCVTTPS2UDQSZ256rr VR256X:$src)>;
+def : Pat<(v16i32(fp_to_uint_sat(v16f32 VR512:$src), i32)),
+          (VCVTTPS2UDQSZrr VR512:$src)>;
+
+// VCVTTPS2UQQS
+def : Pat<(v2i64(X86fp2uisat(v4f32 VR128X:$src))),
+          (VCVTTPS2UQQSZ128rr VR128X:$src)>;
+def : Pat<(v4i64(fp_to_uint_sat(v4f32 VR128X:$src), i64)),
+          (VCVTTPS2UQQSZ256rr VR128X:$src)>;
+def : Pat<(v8i64(fp_to_uint_sat(v8f32 VR256X:$src), i64)),
+          (VCVTTPS2UQQSZrr VR256X:$src)>;
+
 def : Pat<(v4i32 (X86cvttp2sis (v2f64 VR128X:$src))),
           (VCVTTPD2DQSZ128rr VR128X:$src)>;
 def : Pat<(v4i32 (X86cvttp2sis (loadv2f64 addr:$src))),
diff --git a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -390,6 +390,13 @@ def SDTFmaRound : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
                            SDTCisSameAs<1,2>, SDTCisSameAs<1,3>,
                            SDTCisFP<0>, SDTCisVT<4, i32>]>;
 
+def SDTFPToxIntSatOp
+    : SDTypeProfile<1,
+                    1, [SDTCisVec<0>, SDTCisVec<1>, SDTCisInt<0>, SDTCisFP<1>]>;
+
+def X86fp2sisat : SDNode<"X86ISD::FP_TO_SINT_SAT", SDTFPToxIntSatOp>;
+def X86fp2uisat : SDNode<"X86ISD::FP_TO_UINT_SAT", SDTFPToxIntSatOp>;
+
 def X86PAlignr : SDNode<"X86ISD::PALIGNR",
                         SDTypeProfile<1, 3, [SDTCVecEltisVT<0, i8>,
                                              SDTCisSameAs<0,1>,
diff --git a/llvm/test/CodeGen/X86/avx10_2_512fptosi_satcvtds.ll b/llvm/test/CodeGen/X86/avx10_2_512fptosi_satcvtds.ll
@@ -0,0 +1,85 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-linux -mattr=+avx10.2-512 | FileCheck %s --check-prefixes=CHECK,X86
+; RUN: llc < %s -mtriple=x86_64-linux -mattr=+avx10.2-512 | FileCheck %s --check-prefixes=CHECK,X64
+
+; VCVTTPD2DQS
+define <8 x i32> @test_signed_v8i32_v8f64(<8 x double> %f) nounwind {
+; CHECK-LABEL: test_signed_v8i32_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttpd2dqs %zmm0, %ymm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <8 x i32> @llvm.fptosi.sat.v8i32.v8f64(<8 x double> %f)
+  ret <8 x i32> %x
+}
+
+; VCVTTPD2QQS
+define <8 x i64> @test_signed_v8i64_v8f64(<8 x double> %f) nounwind {
+; CHECK-LABEL: test_signed_v8i64_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttpd2qqs %zmm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call <8 x i64> @llvm.fptosi.sat.v8i64.v8f64(<8 x double> %f)
+  ret <8 x i64> %x
+}
+
+; VCVTTPD2UDQS
+define <8 x i32> @test_unsigned_v8i32_v8f64(<8 x double> %f) nounwind {
+; CHECK-LABEL: test_unsigned_v8i32_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttpd2udqs %zmm0, %ymm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <8 x i32> @llvm.fptoui.sat.v8i32.v8f64(<8 x double> %f)
+ ret <8 x i32> %x
+}
+
+; VCVTTPD2UQQS
+define <8 x i64> @test_unsigned_v8i64_v8f64(<8 x double> %f) nounwind {
+; CHECK-LABEL: test_unsigned_v8i64_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttpd2uqqs %zmm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <8 x i64> @llvm.fptoui.sat.v8i64.v8f64(<8 x double> %f)
+  ret <8 x i64> %x
+}
+
+; VCVTTPS2DQS
+define <16 x i32> @test_signed_v16i32_v16f32(<16 x float> %f) nounwind {
+; CHECK-LABEL: test_signed_v16i32_v16f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttps2dqs %zmm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <16 x i32> @llvm.fptosi.sat.v16i32.v16f32(<16 x float> %f)
+  ret <16 x i32> %x
+}
+
+; VCVTTPS2UDQS
+define <16 x i32> @test_unsigned_v16i32_v16f32(<16 x float> %f) nounwind {
+; CHECK-LABEL: test_unsigned_v16i32_v16f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttps2udqs %zmm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <16 x i32> @llvm.fptoui.sat.v16i32.v16f32(<16 x float> %f)
+  ret <16 x i32> %x
+}
+; VCVTTPS2QQS
+define <8 x i64> @test_signed_v8i64_v8f32(<8 x float> %f) nounwind {
+; CHECK-LABEL: test_signed_v8i64_v8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttps2qqs %ymm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <8 x i64> @llvm.fptosi.sat.v8i64.v8f32(<8 x float> %f)
+  ret <8 x i64> %x
+}
+
+; VCVTTPS2UQQS
+define <8 x i64> @test_unsigned_v8i64_v8f32(<8 x float> %f) nounwind {
+; CHECK-LABEL: test_unsigned_v8i64_v8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvttps2uqqs %ymm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+  %x = call  <8 x i64> @llvm.fptoui.sat.v8i64.v8f32(<8 x float> %f)
+  ret <8 x i64> %x
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; X64: {{.*}}
+; X86: {{.*}}
diff --git a/llvm/test/CodeGen/X86/avx10_2fptosi_satcvtds.ll b/llvm/test/CodeGen/X86/avx10_2fptosi_satcvtds.ll
