[DAGCombiner] try to convert pow(x, 1/3) to cbrt(x)

rotateright · rotateright · commit 3eaf500a6d66 · 2018-09-16T16:50:26.000Z
This is a follow-up suggested in D51630 and originally proposed as an IR transform in D49040. Copying the motivational statement by @evandro from that patch: "This transformation helps some benchmarks in SPEC CPU2000 and CPU2006, such as 188.ammp, 447.dealII, 453.povray, and especially 300.twolf, as well as some proprietary benchmarks. Otherwise, no regressions on x86-64 or A64." I'm proposing to add only the minimum support for a DAG node here. Since we don't have an LLVM IR intrinsic for cbrt, and there are no other DAG ways to create a FCBRT node yet, I don't think we need to worry about DAG builder, legalization, a strict variant, etc. We should be able to expand as needed when adding more functionality/transforms. For reference, these are transform suggestions currently listed in SimplifyLibCalls.cpp: // * cbrt(expN(X)) -> expN(x/3) // * cbrt(sqrt(x)) -> pow(x,1/6) // * cbrt(cbrt(x)) -> pow(x,1/9) Also, given that we bail out on long double for now, there should not be any logical differences between platforms (unless there's some platform out there that has pow() but not cbrt()). Differential Revision: https://reviews.llvm.org/D51753 llvm-svn: 342348
diff --git a/llvm/include/llvm/CodeGen/ISDOpcodes.h b/llvm/include/llvm/CodeGen/ISDOpcodes.h
@@ -550,11 +550,8 @@ namespace ISD {
     /// is often a storage-only type but has native conversions.
     FP16_TO_FP, FP_TO_FP16,
 
-    /// FNEG, FABS, FSQRT, FSIN, FCOS, FPOWI, FPOW,
-    /// FLOG, FLOG2, FLOG10, FEXP, FEXP2,
-    /// FCEIL, FTRUNC, FRINT, FNEARBYINT, FROUND, FFLOOR - Perform various unary
-    /// floating point operations. These are inspired by libm.
-    FNEG, FABS, FSQRT, FSIN, FCOS, FPOWI, FPOW,
+    /// Perform various unary floating-point operations inspired by libm.
+    FNEG, FABS, FSQRT, FCBRT, FSIN, FCOS, FPOWI, FPOW,
     FLOG, FLOG2, FLOG10, FEXP, FEXP2,
     FCEIL, FTRUNC, FRINT, FNEARBYINT, FROUND, FFLOOR,
     /// FMINNUM/FMAXNUM - Perform floating-point minimum or maximum on two
diff --git a/llvm/include/llvm/IR/RuntimeLibcalls.def b/llvm/include/llvm/IR/RuntimeLibcalls.def
@@ -128,6 +128,11 @@ HANDLE_LIBCALL(SQRT_F64, "sqrt")
 HANDLE_LIBCALL(SQRT_F80, "sqrtl")
 HANDLE_LIBCALL(SQRT_F128, "sqrtl")
 HANDLE_LIBCALL(SQRT_PPCF128, "sqrtl")
+HANDLE_LIBCALL(CBRT_F32, "cbrtf")
+HANDLE_LIBCALL(CBRT_F64, "cbrt")
+HANDLE_LIBCALL(CBRT_F80, "cbrtl")
+HANDLE_LIBCALL(CBRT_F128, "cbrtl")
+HANDLE_LIBCALL(CBRT_PPCF128, "cbrtl")
 HANDLE_LIBCALL(LOG_F32, "logf")
 HANDLE_LIBCALL(LOG_F64, "log")
 HANDLE_LIBCALL(LOG_F80, "logl")
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -11571,6 +11571,34 @@ SDValue DAGCombiner::visitFPOW(SDNode *N) {
   if (!ExponentC)
     return SDValue();
 
+  // Try to convert x ** (1/3) into cube root.
+  // TODO: Handle the various flavors of long double.
+  // TODO: Since we're approximating, we don't need an exact 1/3 exponent.
+  //       Some range near 1/3 should be fine.
+  EVT VT = N->getValueType(0);
+  if ((VT == MVT::f32 && ExponentC->getValueAPF().isExactlyValue(1.0f/3.0f)) ||
+      (VT == MVT::f64 && ExponentC->getValueAPF().isExactlyValue(1.0/3.0))) {
+    // pow(-0.0, 1/3) = +0.0; cbrt(-0.0) = -0.0.
+    // pow(-inf, 1/3) = +inf; cbrt(-inf) = -inf.
+    // pow(-val, 1/3) =  nan; cbrt(-val) = -num.
+    // For regular numbers, rounding may cause the results to differ.
+    // Therefore, we require { nsz ninf nnan afn } for this transform.
+    // TODO: We could select out the special cases if we don't have nsz/ninf.
+    SDNodeFlags Flags = N->getFlags();
+    if (!Flags.hasNoSignedZeros() || !Flags.hasNoInfs() || !Flags.hasNoNaNs() ||
+        !Flags.hasApproximateFuncs())
+      return SDValue();
+
+    // Do not create a cbrt() libcall if the target does not have it, and do not
+    // turn a pow that has lowering support into a cbrt() libcall.
+    if (!DAG.getLibInfo().has(LibFunc_cbrt) ||
+        (!DAG.getTargetLoweringInfo().isOperationExpand(ISD::FPOW, VT) &&
+         DAG.getTargetLoweringInfo().isOperationExpand(ISD::FCBRT, VT)))
+      return SDValue();
+
+    return DAG.getNode(ISD::FCBRT, SDLoc(N), VT, N->getOperand(0), Flags);
+  }
+
   // Try to convert x ** (1/4) into square roots.
   // x ** (1/2) is canonicalized to sqrt, so we do not bother with that case.
   // TODO: This could be extended (using a target hook) to handle smaller
@@ -11587,7 +11615,6 @@ SDValue DAGCombiner::visitFPOW(SDNode *N) {
       return SDValue();
 
     // Don't double the number of libcalls. We are trying to inline fast code.
-    EVT VT = N->getValueType(0);
     if (!DAG.getTargetLoweringInfo().isOperationLegalOrCustom(ISD::FSQRT, VT))
       return SDValue();
 
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -4047,6 +4047,11 @@ void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) {
                                       RTLIB::SQRT_F80, RTLIB::SQRT_F128,
                                       RTLIB::SQRT_PPCF128));
     break;
+  case ISD::FCBRT:
+    Results.push_back(ExpandFPLibCall(Node, RTLIB::CBRT_F32, RTLIB::CBRT_F64,
+                                      RTLIB::CBRT_F80, RTLIB::CBRT_F128,
+                                      RTLIB::CBRT_PPCF128));
+    break;
   case ISD::FSIN:
   case ISD::STRICT_FSIN:
     Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
@@ -181,6 +181,7 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
   case ISD::FNEG:                       return "fneg";
   case ISD::FSQRT:                      return "fsqrt";
   case ISD::STRICT_FSQRT:               return "strict_fsqrt";
+  case ISD::FCBRT:                      return "fcbrt";
   case ISD::FSIN:                       return "fsin";
   case ISD::STRICT_FSIN:                return "strict_fsin";
   case ISD::FCOS:                       return "fcos";
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -666,6 +666,7 @@ void TargetLoweringBase::initActions() {
 
   // These library functions default to expand.
   for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
+    setOperationAction(ISD::FCBRT,      VT, Expand);
     setOperationAction(ISD::FLOG ,      VT, Expand);
     setOperationAction(ISD::FLOG2,      VT, Expand);
     setOperationAction(ISD::FLOG10,     VT, Expand);
diff --git a/llvm/test/CodeGen/X86/pow.ll b/llvm/test/CodeGen/X86/pow.ll
@@ -7,6 +7,8 @@ declare <4 x float> @llvm.pow.v4f32(<4 x float>, <4 x float>)
 declare double @llvm.pow.f64(double, double)
 declare <2 x double> @llvm.pow.v2f64(<2 x double>, <2 x double>)
 
+declare x86_fp80 @llvm.pow.f80(x86_fp80, x86_fp80)
+
 define float @pow_f32_one_fourth_fmf(float %x) nounwind {
 ; CHECK-LABEL: pow_f32_one_fourth_fmf:
 ; CHECK:       # %bb.0:
@@ -165,8 +167,7 @@ define <2 x double> @pow_v2f64_one_fourth_not_enough_fmf(<2 x double> %x) nounwi
 define float @pow_f32_one_third_fmf(float %x) nounwind {
 ; CHECK-LABEL: pow_f32_one_third_fmf:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; CHECK-NEXT:    jmp powf # TAILCALL
+; CHECK-NEXT:    jmp cbrtf # TAILCALL
   %one = uitofp i32 1 to float
   %three = uitofp i32 3 to float
   %exp = fdiv float %one, %three
@@ -177,12 +178,53 @@ define float @pow_f32_one_third_fmf(float %x) nounwind {
 define double @pow_f64_one_third_fmf(double %x) nounwind {
 ; CHECK-LABEL: pow_f64_one_third_fmf:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
-; CHECK-NEXT:    jmp pow # TAILCALL
+; CHECK-NEXT:    jmp cbrt # TAILCALL
   %one = uitofp i32 1 to double
   %three = uitofp i32 3 to double
   %exp = fdiv double %one, %three
   %r = call nsz nnan ninf afn double @llvm.pow.f64(double %x, double %exp)
   ret double %r
 }
 
+; TODO: We could turn this into cbrtl, but currently we only handle float/double types.
+
+define x86_fp80 @pow_f80_one_third_fmf(x86_fp80 %x) nounwind {
+; CHECK-LABEL: pow_f80_one_third_fmf:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    subq $40, %rsp
+; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
+; CHECK-NEXT:    fldt {{.*}}(%rip)
+; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
+; CHECK-NEXT:    fstpt (%rsp)
+; CHECK-NEXT:    callq powl
+; CHECK-NEXT:    addq $40, %rsp
+; CHECK-NEXT:    retq
+  %one = uitofp i32 1 to x86_fp80
+  %three = uitofp i32 3 to x86_fp80
+  %exp = fdiv x86_fp80 %one, %three
+  %r = call nsz nnan ninf afn x86_fp80 @llvm.pow.f80(x86_fp80 %x, x86_fp80 %exp)
+  ret x86_fp80 %r
+}
+
+; We might want to allow this. The exact hex value for 1/3 as a double is 0x3fd5555555555555.
+
+define double @pow_f64_not_exactly_one_third_fmf(double %x) nounwind {
+; CHECK-LABEL: pow_f64_not_exactly_one_third_fmf:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; CHECK-NEXT:    jmp pow # TAILCALL
+  %r = call nsz nnan ninf afn double @llvm.pow.f64(double %x, double 0x3fd5555555555556)
+  ret double %r
+}
+
+; We require all 4 of nsz, ninf, nnan, afn.
+
+define double @pow_f64_not_enough_fmf(double %x) nounwind {
+; CHECK-LABEL: pow_f64_not_enough_fmf:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
+; CHECK-NEXT:    jmp pow # TAILCALL
+  %r = call nsz ninf afn double @llvm.pow.f64(double %x, double 0x3fd5555555555555)
+  ret double %r
+}
+
