[PowerPC] Treat llvm.fmuladd intrinsic as using CTR

This fixes bug 55463, similar to D78668. This is a temporary fix since
we will switch to post-isel CTR loop determination in the future.

Reviewed By: dim, shchenz

Differential Revision: https://reviews.llvm.org/D125746

(cherry picked from commit d9d15af7873fe16d7a0dde4def30f40fa9901777)

2 files changed, 137 insertions, 3 deletions

diff --git a/llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp b/llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
index cc5738a5d7b6..48be7e0860df 100644
--- a/llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
+++ b/llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
@@ -491,15 +491,13 @@ bool PPCTTIImpl::mightUseCTR(BasicBlock *BB, TargetLibraryInfo *LibInfo,
           case Intrinsic::experimental_constrained_sin:
           case Intrinsic::experimental_constrained_cos:
             return true;
-          // There is no corresponding FMA instruction for PPC double double.
-          // Thus, we need to disable CTR loop generation for this type.
-          case Intrinsic::fmuladd:
           case Intrinsic::copysign:
             if (CI->getArgOperand(0)->getType()->getScalarType()->
                 isPPC_FP128Ty())
               return true;
             else
               continue; // ISD::FCOPYSIGN is never a library call.
+          case Intrinsic::fmuladd:
           case Intrinsic::fma:                Opcode = ISD::FMA;        break;
           case Intrinsic::sqrt:               Opcode = ISD::FSQRT;      break;
           case Intrinsic::floor:              Opcode = ISD::FFLOOR;     break;
diff --git a/llvm/test/CodeGen/PowerPC/pr55463.ll b/llvm/test/CodeGen/PowerPC/pr55463.ll
new file mode 100644
index 000000000000..63d4170da2d5
--- /dev/null
+++ b/llvm/test/CodeGen/PowerPC/pr55463.ll
@@ -0,0 +1,136 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=powerpcspe -verify-machineinstrs < %s | FileCheck %s
+
+define void @baz() #0 {
+; CHECK-LABEL: baz:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    mflr 0
+; CHECK-NEXT:    stw 0, 4(1)
+; CHECK-NEXT:    stwu 1, -48(1)
+; CHECK-NEXT:    li 3, .LCPI0_0@l
+; CHECK-NEXT:    li 5, .LCPI0_1@l
+; CHECK-NEXT:    lis 4, .LCPI0_0@ha
+; CHECK-NEXT:    lis 6, .LCPI0_1@ha
+; CHECK-NEXT:    evstdd 29, 24(1) # 8-byte Folded Spill
+; CHECK-NEXT:    evstdd 30, 32(1) # 8-byte Folded Spill
+; CHECK-NEXT:    evlddx 30, 4, 3
+; CHECK-NEXT:    # implicit-def: $r3
+; CHECK-NEXT:    evlddx 29, 6, 5
+; CHECK-NEXT:    evstdd 28, 16(1) # 8-byte Folded Spill
+; CHECK-NEXT:    # implicit-def: $r28
+; CHECK-NEXT:    .p2align 4
+; CHECK-NEXT:  .LBB0_1: # %bb1
+; CHECK-NEXT:    #
+; CHECK-NEXT:    efdcfsi 8, 3
+; CHECK-NEXT:    mr 4, 30
+; CHECK-NEXT:    mr 6, 29
+; CHECK-NEXT:    evmergehi 3, 30, 30
+; CHECK-NEXT:    evmergehi 5, 29, 29
+; CHECK-NEXT:    # kill: def $r3 killed $r3 killed $s3
+; CHECK-NEXT:    # kill: def $r5 killed $r5 killed $s5
+; CHECK-NEXT:    evmergehi 7, 8, 8
+; CHECK-NEXT:    # kill: def $r8 killed $r8 killed $s8
+; CHECK-NEXT:    # kill: def $r7 killed $r7 killed $s7
+; CHECK-NEXT:    bl fma
+; CHECK-NEXT:    evmergelo 3, 3, 4
+; CHECK-NEXT:    addi 5, 28, 1
+; CHECK-NEXT:    cmplw 5, 28
+; CHECK-NEXT:    mr 28, 5
+; CHECK-NEXT:    efdctsiz 3, 3
+; CHECK-NEXT:    bge 0, .LBB0_1
+; CHECK-NEXT:  # %bb.2: # %bb8
+; CHECK-NEXT:    bl wibble
+; CHECK-NEXT:    evldd 30, 32(1) # 8-byte Folded Reload
+; CHECK-NEXT:    evldd 29, 24(1) # 8-byte Folded Reload
+; CHECK-NEXT:    evldd 28, 16(1) # 8-byte Folded Reload
+; CHECK-NEXT:    lwz 0, 52(1)
+; CHECK-NEXT:    addi 1, 1, 48
+; CHECK-NEXT:    mtlr 0
+; CHECK-NEXT:    blr
+bb:
+  br label %bb1
+
+bb1:
+  %tmp = phi i32 [ %tmp6, %bb1 ], [ undef, %bb ]
+  %tmp2 = phi i32 [ %tmp3, %bb1 ], [ undef, %bb ]
+  %tmp3 = add nsw i32 %tmp2, 1
+  %tmp4 = sitofp i32 %tmp to double
+  %tmp5 = tail call double @llvm.fmuladd.f64(double 0.000000e+00, double -0.000000e+00, double %tmp4)
+  %tmp6 = fptosi double %tmp5 to i32
+  %tmp7 = icmp eq i32 %tmp2, 0
+  br i1 %tmp7, label %bb8, label %bb1
+
+bb8:
+  call void @wibble(i32 %tmp6)
+  ret void
+}
+
+define void @wombat() #0 {
+; CHECK-LABEL: wombat:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    mflr 0
+; CHECK-NEXT:    stw 0, 4(1)
+; CHECK-NEXT:    stwu 1, -48(1)
+; CHECK-NEXT:    li 3, .LCPI1_0@l
+; CHECK-NEXT:    li 5, .LCPI1_1@l
+; CHECK-NEXT:    lis 4, .LCPI1_0@ha
+; CHECK-NEXT:    lis 6, .LCPI1_1@ha
+; CHECK-NEXT:    evstdd 29, 24(1) # 8-byte Folded Spill
+; CHECK-NEXT:    evstdd 30, 32(1) # 8-byte Folded Spill
+; CHECK-NEXT:    evlddx 30, 4, 3
+; CHECK-NEXT:    # implicit-def: $r3
+; CHECK-NEXT:    evlddx 29, 6, 5
+; CHECK-NEXT:    evstdd 28, 16(1) # 8-byte Folded Spill
+; CHECK-NEXT:    # implicit-def: $r28
+; CHECK-NEXT:    .p2align 4
+; CHECK-NEXT:  .LBB1_1: # %bb1
+; CHECK-NEXT:    #
+; CHECK-NEXT:    efdcfsi 8, 3
+; CHECK-NEXT:    mr 4, 30
+; CHECK-NEXT:    mr 6, 29
+; CHECK-NEXT:    evmergehi 3, 30, 30
+; CHECK-NEXT:    evmergehi 5, 29, 29
+; CHECK-NEXT:    # kill: def $r3 killed $r3 killed $s3
+; CHECK-NEXT:    # kill: def $r5 killed $r5 killed $s5
+; CHECK-NEXT:    evmergehi 7, 8, 8
+; CHECK-NEXT:    # kill: def $r8 killed $r8 killed $s8
+; CHECK-NEXT:    # kill: def $r7 killed $r7 killed $s7
+; CHECK-NEXT:    bl fma
+; CHECK-NEXT:    evmergelo 3, 3, 4
+; CHECK-NEXT:    addi 5, 28, 1
+; CHECK-NEXT:    cmplw 5, 28
+; CHECK-NEXT:    mr 28, 5
+; CHECK-NEXT:    efdctsiz 3, 3
+; CHECK-NEXT:    bge 0, .LBB1_1
+; CHECK-NEXT:  # %bb.2: # %bb8
+; CHECK-NEXT:    bl wibble
+; CHECK-NEXT:    evldd 30, 32(1) # 8-byte Folded Reload
+; CHECK-NEXT:    evldd 29, 24(1) # 8-byte Folded Reload
+; CHECK-NEXT:    evldd 28, 16(1) # 8-byte Folded Reload
+; CHECK-NEXT:    lwz 0, 52(1)
+; CHECK-NEXT:    addi 1, 1, 48
+; CHECK-NEXT:    mtlr 0
+; CHECK-NEXT:    blr
+bb:
+  br label %bb1
+
+bb1:
+  %tmp = phi i32 [ %tmp6, %bb1 ], [ undef, %bb ]
+  %tmp2 = phi i32 [ %tmp3, %bb1 ], [ undef, %bb ]
+  %tmp3 = add nsw i32 %tmp2, 1
+  %tmp4 = sitofp i32 %tmp to double
+  %tmp5 = tail call double @llvm.fma.f64(double 0.000000e+00, double -0.000000e+00, double %tmp4)
+  %tmp6 = fptosi double %tmp5 to i32
+  %tmp7 = icmp eq i32 %tmp2, 0
+  br i1 %tmp7, label %bb8, label %bb1
+
+bb8:
+  call void @wibble(i32 %tmp6)
+  ret void
+}
+
+declare void @wibble(i32)
+declare double @llvm.fmuladd.f64(double, double, double)
+declare double @llvm.fma.f64(double, double, double)
+
+attributes #0 = { nounwind }