[llvm] 6245690 - [PowerPC] improve performance on the isNan and !isNan function in case of -ffp-model=strict (#204170)

via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 23 06:32:25 PDT 2026


Author: zhijian lin
Date: 2026-07-23T09:32:21-04:00
New Revision: 624569002f06a27543ac964b2b338d677deb4e37

URL: https://github.com/llvm/llvm-project/commit/624569002f06a27543ac964b2b338d677deb4e37
DIFF: https://github.com/llvm/llvm-project/commit/624569002f06a27543ac964b2b338d677deb4e37.diff

LOG: [PowerPC] improve performance on the isNan and !isNan function  in case of -ffp-model=strict (#204170)

For the IR representation:

isnan(x) → %0 = tail call noundef i1 @llvm.is.fpclass.f64(double %x, i32
3)
!isnan(x) → %0 = tail call noundef i1 @llvm.is.fpclass.f64(double %x,
i32 1020)

Under `-ffp-model=strict`, the generic
TargetLowering::expandIS_FPCLASS() is used to lower these when
Subtarget.hasP9Vector() && Subtarget.useCRBits() is false. However,
PowerPC has more optimal assembly sequences for isnan(x) and !isnan(x)
on POWER7/8 and generic PPC targets.
We implement a custom lowering for isnan(x) and !isnan(x) under
`-ffp-model=strict`, using `fcmpu `for POWER7/8 and generic PPC targets,
and` xscmpudp` for targets where VSX is available.

Added: 
    llvm/test/CodeGen/PowerPC/fp-classify-nan.ll

Modified: 
    llvm/lib/Target/PowerPC/PPCISelLowering.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 91125abb89504..8a03946a2ca10 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -817,6 +817,9 @@ PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
     setOperationAction(ISD::FCANONICALIZE, MVT::f32, Legal);
   }
 
+  setOperationAction(ISD::IS_FPCLASS, MVT::f32, Custom);
+  setOperationAction(ISD::IS_FPCLASS, MVT::f64, Custom);
+
   if (Subtarget.hasAltivec()) {
     for (MVT VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
       setOperationAction(ISD::AVGCEILS, VT, Legal);
@@ -1258,11 +1261,8 @@ PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
     if (Subtarget.hasP9Vector()) {
       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
-
       // Test data class instructions store results in CR bits.
       if (Subtarget.useCRBits()) {
-        setOperationAction(ISD::IS_FPCLASS, MVT::f32, Custom);
-        setOperationAction(ISD::IS_FPCLASS, MVT::f64, Custom);
         setOperationAction(ISD::IS_FPCLASS, MVT::f128, Custom);
         setOperationAction(ISD::IS_FPCLASS, MVT::ppcf128, Custom);
       }
@@ -11962,18 +11962,78 @@ static SDValue getDataClassTest(SDValue Op, FPClassTest Mask, const SDLoc &Dl,
 
 SDValue PPCTargetLowering::LowerIS_FPCLASS(SDValue Op,
                                            SelectionDAG &DAG) const {
-  assert(Subtarget.hasP9Vector() && "Test data class requires Power9");
   SDValue LHS = Op.getOperand(0);
   uint64_t RHSC = Op.getConstantOperandVal(1);
   SDLoc Dl(Op);
   FPClassTest Category = static_cast<FPClassTest>(RHSC);
-  if (LHS.getValueType() == MVT::ppcf128) {
+  EVT VT = LHS.getValueType();
+
+  assert((VT == MVT::f32 || VT == MVT::f64 ||
+          ((VT == MVT::f128 || VT == MVT::ppcf128) && Subtarget.hasVSX() &&
+           Subtarget.useCRBits())) &&
+         "invalid customize type for IS_FPCLASS.");
+  // Handle ppcf128 by extracting the higher part
+  if (VT == MVT::ppcf128) {
     // The higher part determines the value class.
     LHS = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::f64, LHS,
                       DAG.getConstant(1, Dl, MVT::i32));
+    VT = MVT::f64;
+  }
+
+  // If we have P9Vector and useCRBits, use the data class test instructions
+  if (Subtarget.hasP9Vector() && Subtarget.useCRBits()) {
+    return getDataClassTest(LHS, Category, Dl, DAG, Subtarget);
+  }
+
+  // For non-P9Vector targets, we can only check for NaN using fcmpu/xscmpudp
+  // These instructions set CR bits based on comparison with itself:
+  // - If value is NaN, the comparison is unordered (FU bit set)
+  // - If value is not NaN, the comparison is equal (EQ bit set)
+
+  if ((Category & ~fcNan) && (Category != ~fcNan)) {
+    // If not checking for NaN or non-NaN, we can't handle this without P9Vector
+    return SDValue();
   }
 
-  return getDataClassTest(LHS, Category, Dl, DAG, Subtarget);
+  // Determine which comparison instruction to use based on vector support
+  unsigned CmpOp;
+
+  if (Subtarget.hasVSX()) {
+    // Use xscmpudp for VSX targets (both f32 and f64)
+    // For f32, extend to f64 first
+    if (VT == MVT::f32) {
+      LHS = DAG.getNode(ISD::FP_EXTEND, Dl, MVT::f64, LHS);
+    } else if (VT != MVT::f64) {
+      return SDValue();
+    }
+    CmpOp = PPC::XSCMPUDP;
+  } else {
+    // Use fcmpu for non-VSX targets
+    // FCMPUS and FCMPUD both map to the same fcmpu instruction,
+    // just with 
diff erent register classes (f4rc vs f8rc)
+    if (VT == MVT::f64) {
+      CmpOp = PPC::FCMPUD;
+    } else if (VT == MVT::f32) {
+      CmpOp = PPC::FCMPUS;
+    } else {
+      return SDValue();
+    }
+  }
+
+  // Create the comparison: fcmpu/xscmpudp CR, LHS, LHS
+  // The CR field output will be allocated by the register allocator
+  SDValue Cmp = SDValue(DAG.getMachineNode(CmpOp, Dl, MVT::i32, LHS, LHS), 0);
+
+  // Extract the unordered bit (FU) from the CR field
+  // For NaN detection: FU bit is set if operands are unordered (i.e., NaN)
+  SDValue NanCheck = SDValue(
+      DAG.getMachineNode(
+          TargetOpcode::EXTRACT_SUBREG, Dl, MVT::i1, Cmp,
+          DAG.getTargetConstant(Category == ~fcNan ? PPC::sub_un : PPC::sub_eq,
+                                Dl, MVT::i32)),
+      0);
+
+  return DAG.getNOT(Dl, NanCheck, MVT::i1);
 }
 
 // Adjust the length value for a load/store with length to account for the

diff  --git a/llvm/test/CodeGen/PowerPC/fp-classify-nan.ll b/llvm/test/CodeGen/PowerPC/fp-classify-nan.ll
new file mode 100644
index 0000000000000..3a866984f2ba7
--- /dev/null
+++ b/llvm/test/CodeGen/PowerPC/fp-classify-nan.ll
@@ -0,0 +1,126 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P8
+; RUN: llc -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P9
+; RUN: llc -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 -mattr=-vsx < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P8-NO-VSX
+
+; RUN: llc -mtriple=powerpc-ibm-aix7.2.0.0 -mcpu=pwr8 < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P8
+; RUN: llc -mtriple=powerpc-ibm-aix7.2.0.0 -mcpu=pwr9 < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P9
+; RUN: llc -mtriple=powerpc-ibm-aix7.2.0.0 -mcpu=pwr8 -mattr=-vsx < %s \
+; RUN:   -verify-machineinstrs -ppc-asm-full-reg-names | FileCheck %s --check-prefix=P8-NO-VSX
+;
+define zeroext i1 @test_is_nan_f64(double %x) #0 {
+; P8-LABEL: test_is_nan_f64:
+; P8:       # %bb.0:
+; P8-NEXT:    xscmpudp cr0, f1, f1
+; P8-NEXT:    li r3, 1
+; P8-NEXT:    iseleq r3, 0, r3
+; P8-NEXT:    blr
+;
+; P9-LABEL: test_is_nan_f64:
+; P9:       # %bb.0:
+; P9-NEXT:    xststdcdp cr0, f1, 64
+; P9-NEXT:    li r3, 0
+; P9-NEXT:    li r4, 1
+; P9-NEXT:    iseleq r3, r4, r3
+; P9-NEXT:    blr
+;
+; P8-NO-VSX-LABEL: test_is_nan_f64:
+; P8-NO-VSX:       # %bb.0:
+; P8-NO-VSX-NEXT:    fcmpu cr0, f1, f1
+; P8-NO-VSX-NEXT:    li r3, 1
+; P8-NO-VSX-NEXT:    iseleq r3, 0, r3
+; P8-NO-VSX-NEXT:    blr
+  %result = call i1 @llvm.is.fpclass.f64(double %x, i32 3)
+  ret i1 %result
+}
+
+define zeroext i1 @test_is_not_nan_f64(double %x) #0 {
+; P8-LABEL: test_is_not_nan_f64:
+; P8:       # %bb.0:
+; P8-NEXT:    xscmpudp cr0, f1, f1
+; P8-NEXT:    li r3, 1
+; P8-NEXT:    isel r3, 0, r3, un
+; P8-NEXT:    blr
+;
+; P9-LABEL: test_is_not_nan_f64:
+; P9:       # %bb.0:
+; P9-NEXT:    xststdcdp cr0, f1, 64
+; P9-NEXT:    li r3, 1
+; P9-NEXT:    iseleq r3, 0, r3
+; P9-NEXT:    blr
+;
+; P8-NO-VSX-LABEL: test_is_not_nan_f64:
+; P8-NO-VSX:       # %bb.0:
+; P8-NO-VSX-NEXT:    fcmpu cr0, f1, f1
+; P8-NO-VSX-NEXT:    li r3, 1
+; P8-NO-VSX-NEXT:    isel r3, 0, r3, un
+; P8-NO-VSX-NEXT:    blr
+  %result = call i1 @llvm.is.fpclass.f64(double %x, i32 1020)
+  ret i1 %result
+}
+
+define zeroext i1 @test_is_nan_f32(float %x) #0 {
+; P8-LABEL: test_is_nan_f32:
+; P8:       # %bb.0:
+; P8-NEXT:    xscmpudp cr0, f1, f1
+; P8-NEXT:    li r3, 1
+; P8-NEXT:    iseleq r3, 0, r3
+; P8-NEXT:    blr
+;
+; P9-LABEL: test_is_nan_f32:
+; P9:       # %bb.0:
+; P9-NEXT:    xststdcsp cr0, f1, 64
+; P9-NEXT:    li r3, 0
+; P9-NEXT:    li r4, 1
+; P9-NEXT:    iseleq r3, r4, r3
+; P9-NEXT:    blr
+;
+; P8-NO-VSX-LABEL: test_is_nan_f32:
+; P8-NO-VSX:       # %bb.0:
+; P8-NO-VSX-NEXT:    fcmpu cr0, f1, f1
+; P8-NO-VSX-NEXT:    li r3, 1
+; P8-NO-VSX-NEXT:    iseleq r3, 0, r3
+; P8-NO-VSX-NEXT:    blr
+  %result = call i1 @llvm.is.fpclass.f32(float %x, i32 3)
+  ret i1 %result
+}
+
+define zeroext i1 @test_is_not_nan_f32(float %x) #0 {
+; P8-LABEL: test_is_not_nan_f32:
+; P8:       # %bb.0:
+; P8-NEXT:    xscmpudp cr0, f1, f1
+; P8-NEXT:    li r3, 1
+; P8-NEXT:    isel r3, 0, r3, un
+; P8-NEXT:    blr
+;
+; P9-LABEL: test_is_not_nan_f32:
+; P9:       # %bb.0:
+; P9-NEXT:    xststdcsp cr0, f1, 64
+; P9-NEXT:    li r3, 1
+; P9-NEXT:    iseleq r3, 0, r3
+; P9-NEXT:    blr
+;
+; P8-NO-VSX-LABEL: test_is_not_nan_f32:
+; P8-NO-VSX:       # %bb.0:
+; P8-NO-VSX-NEXT:    fcmpu cr0, f1, f1
+; P8-NO-VSX-NEXT:    li r3, 1
+; P8-NO-VSX-NEXT:    isel r3, 0, r3, un
+; P8-NO-VSX-NEXT:    blr
+  %result = call i1 @llvm.is.fpclass.f32(float %x, i32 1020)
+  ret i1 %result
+}
+
+declare i1 @llvm.is.fpclass.f64(double, i32)
+declare i1 @llvm.is.fpclass.f32(float, i32)
+declare float @llvm.fabs.f32(float)
+declare double @llvm.fabs.f64(double)
+declare fp128 @llvm.fabs.f128(fp128)
+declare <4 x float> @llvm.fabs.v4f32(<4 x float>)
+declare <2 x double> @llvm.fabs.v2f64(<2 x double>)
+
+attributes #0 = {strictfp}


        


More information about the llvm-commits mailing list