[GlobalISel][AArch64] Add instruction selection support for @llvm.log2

[paquette]

Summary

This teaches GlobalISel to emit a RTLib call for @llvm.log2 when it encounters it.

It updates the existing floating point tests to show that we don't fall back on the intrinsic, and select the correct instructions. It also adds a legalizer test for G_FLOG2.

[aemerson]

LGTM

[paquette]

Committed in r352673

lib/CodeGen/GlobalISel/LegalizerHelper.cpp

static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
   case TargetOpcode::G_FLOG10:
     assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
     return Size == 128 ? RTLIB::LOG10_F128
                        : Size == 64 ? RTLIB::LOG10_F64 : RTLIB::LOG10_F32;
   case TargetOpcode::G_FLOG:
     assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
     return Size == 128 ? RTLIB::LOG_F128
                        : Size == 64 ? RTLIB::LOG_F64 : RTLIB::LOG_F32;
+  case TargetOpcode::G_FLOG2:
+    assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
+    return Size == 128 ? RTLIB::LOG2_F128
+                       : Size == 64 ? RTLIB::LOG2_F64 : RTLIB::LOG2_F32;
   }
   llvm_unreachable("Unknown libcall function");
 }
 
 LegalizerHelper::LegalizeResult
 llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall,
                     const CallLowering::ArgInfo &Result,
                     ArrayRef<CallLowering::ArgInfo> Args) {
LegalizerHelper::libcall(MachineInstr &MI) {
   case TargetOpcode::G_FMUL:
   case TargetOpcode::G_FDIV:
   case TargetOpcode::G_FMA:
   case TargetOpcode::G_FPOW:
   case TargetOpcode::G_FREM:
   case TargetOpcode::G_FCOS:
   case TargetOpcode::G_FSIN:
   case TargetOpcode::G_FLOG10:
-  case TargetOpcode::G_FLOG: {
+  case TargetOpcode::G_FLOG:
+  case TargetOpcode::G_FLOG2: {
     if (Size > 64) {
       LLVM_DEBUG(dbgs() << "Size " << Size << " too large to legalize.\n");
       return UnableToLegalize;
     }
     Type *HLTy = Size == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx);
     auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
     if (Status != Legalized)
       return Status;
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   case TargetOpcode::G_FABS:
   case TargetOpcode::G_FDIV:
   case TargetOpcode::G_FREM:
   case TargetOpcode::G_FCEIL:
   case TargetOpcode::G_FCOS:
   case TargetOpcode::G_FSIN:
   case TargetOpcode::G_FLOG10:
   case TargetOpcode::G_FLOG:
+  case TargetOpcode::G_FLOG2:
     assert(TypeIdx == 0);
     Observer.changingInstr(MI);
 
     for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I)
       widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT);
 
     widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
     Observer.changedInstr(MI);

lib/Target/AArch64/AArch64LegalizerInfo.cpp

getActionDefinitionsBuilder(G_FCEIL)
       // encounter it.
       .widenScalarIf(
           [=, &ST](const LegalityQuery &Query) {
             return Query.Types[0] == s16 && !ST.hasFullFP16();
           },
           [=](const LegalityQuery &Query) { return std::make_pair(0, s32); })
       .legalFor({s16, s32, s64, v2s32, v4s32, v2s64, v2s16, v4s16, v8s16});
 
-  getActionDefinitionsBuilder({G_FCOS, G_FSIN, G_FLOG10, G_FLOG})
+  getActionDefinitionsBuilder({G_FCOS, G_FSIN, G_FLOG10, G_FLOG, G_FLOG2})
       // We need a call for these, so we always need to scalarize.
       .scalarize(0)
       // Regardless of FP16 support, widen 16-bit elements to 32-bits.
       .minScalar(0, s32)
       .libcallFor({s32, s64, v2s32, v4s32, v2s64});
 
   getActionDefinitionsBuilder(G_INSERT)
       .unsupportedIf([=](const LegalityQuery &Query) {

lib/Target/AArch64/AArch64RegisterBankInfo.cpp

static bool isPreISelGenericFloatingPointOpcode(unsigned Opc) {
   case TargetOpcode::G_FPEXT:
   case TargetOpcode::G_FPTRUNC:
   case TargetOpcode::G_FCEIL:
   case TargetOpcode::G_FNEG:
   case TargetOpcode::G_FCOS:
   case TargetOpcode::G_FSIN:
   case TargetOpcode::G_FLOG10:
   case TargetOpcode::G_FLOG:
+  case TargetOpcode::G_FLOG2:
     return true;
   }
   return false;
 }
 
 const RegisterBankInfo::InstructionMapping &
 AArch64RegisterBankInfo::getSameKindOfOperandsMapping(
     const MachineInstr &MI) const {

test/CodeGen/AArch64/GlobalISel/legalize-log2.mir

+# RUN: llc -verify-machineinstrs -mtriple aarch64--- \
+# RUN: -run-pass=legalizer -mattr=+fullfp16 -global-isel %s -o - \
+# RUN: | FileCheck %s
+...
+---
+name:            test_v4f16.log2
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $d0
+    ; CHECK-LABEL: name:            test_v4f16.log2
+    ; CHECK: [[V1:%[0-9]+]]:_(s16), [[V2:%[0-9]+]]:_(s16), [[V3:%[0-9]+]]:_(s16), [[V4:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES %{{[0-9]+}}(<4 x s16>)
+
+    ; CHECK-DAG: [[V1_S32:%[0-9]+]]:_(s32) = G_FPEXT [[V1]](s16)
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-NEXT: $s0 = COPY [[V1_S32]](s32)
+    ; CHECK-NEXT: BL &log2
+    ; CHECK-NEXT: [[ELT1_S32:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-NEXT: ADJCALLSTACKUP
+    ; CHECK-NEXT: [[ELT1:%[0-9]+]]:_(s16) = G_FPTRUNC [[ELT1_S32]](s32)
+
+    ; CHECK-DAG: [[V2_S32:%[0-9]+]]:_(s32) = G_FPEXT [[V2]](s16)
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-NEXT: $s0 = COPY [[V2_S32]](s32)
+    ; CHECK-NEXT: BL &log2
+    ; CHECK-NEXT: [[ELT2_S32:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-NEXT: ADJCALLSTACKUP
+    ; CHECK-NEXT: [[ELT2:%[0-9]+]]:_(s16) = G_FPTRUNC [[ELT2_S32]](s32)
+
+    ; CHECK-DAG: [[V3_S32:%[0-9]+]]:_(s32) = G_FPEXT [[V3]](s16)
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-NEXT: $s0 = COPY [[V3_S32]](s32)
+    ; CHECK-NEXT: BL &log2
+    ; CHECK-NEXT: [[ELT3_S32:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-NEXT: ADJCALLSTACKUP
+    ; CHECK-NEXT: [[ELT3:%[0-9]+]]:_(s16) = G_FPTRUNC [[ELT3_S32]](s32)
+
+    ; CHECK-DAG: [[V4_S32:%[0-9]+]]:_(s32) = G_FPEXT [[V4]](s16)
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-NEXT: $s0 = COPY [[V4_S32]](s32)
+    ; CHECK-NEXT: BL &log2
+    ; CHECK-NEXT: [[ELT4_S32:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-NEXT: ADJCALLSTACKUP
+    ; CHECK-NEXT: [[ELT4:%[0-9]+]]:_(s16) = G_FPTRUNC [[ELT4_S32]](s32)
+
+    ; CHECK-DAG: %{{[0-9]+}}:_(<4 x s16>) = G_BUILD_VECTOR [[ELT1]](s16), [[ELT2]](s16), [[ELT3]](s16), [[ELT4]](s16)
+
+    %0:_(<4 x s16>) = COPY $d0
+    %1:_(<4 x s16>) = G_FLOG2 %0
+    $d0 = COPY %1(<4 x s16>)
+    RET_ReallyLR implicit $d0
+
+...
+---
+name:            test_v8f16.log2
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; CHECK-LABEL: name:            test_v8f16.log2
+
+    ; This is big, so let's just check for the 8 calls to log2, the the
+    ; G_UNMERGE_VALUES, and the G_BUILD_VECTOR. The other instructions ought
+    ; to be covered by the other tests.
+
+    ; CHECK: G_UNMERGE_VALUES
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK: BL &log2
+    ; CHECK-DAG: G_BUILD_VECTOR
+
+    %0:_(<8 x s16>) = COPY $q0
+    %1:_(<8 x s16>) = G_FLOG2 %0
+    $q0 = COPY %1(<8 x s16>)
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:            test_v2f32.log2
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $d0
+
+    ; CHECK-LABEL: name:            test_v2f32.log2
+    ; CHECK: [[V1:%[0-9]+]]:_(s32), [[V2:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES %{{[0-9]+}}(<2 x s32>)
+
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V1]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT1:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V2]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT2:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: %1:_(<2 x s32>) = G_BUILD_VECTOR [[ELT1]](s32), [[ELT2]](s32)
+
+    %0:_(<2 x s32>) = COPY $d0
+    %1:_(<2 x s32>) = G_FLOG2 %0
+    $d0 = COPY %1(<2 x s32>)
+    RET_ReallyLR implicit $d0
+
+...
+---
+name:            test_v4f32.log2
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $q0
+    ; CHECK-LABEL: name:            test_v4f32.log2
+    ; CHECK: [[V1:%[0-9]+]]:_(s32), [[V2:%[0-9]+]]:_(s32), [[V3:%[0-9]+]]:_(s32), [[V4:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES %{{[0-9]+}}(<4 x s32>)
+
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V1]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT1:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V2]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT2:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V3]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT3:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $s0 = COPY [[V4]](s32)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT4:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: %1:_(<4 x s32>) = G_BUILD_VECTOR [[ELT1]](s32), [[ELT2]](s32), [[ELT3]](s32), [[ELT4]](s32)
+
+    %0:_(<4 x s32>) = COPY $q0
+    %1:_(<4 x s32>) = G_FLOG2 %0
+    $q0 = COPY %1(<4 x s32>)
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:            test_v2f64.log2
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; CHECK-LABEL: name:            test_v2f64.log2
+    ; CHECK: [[V1:%[0-9]+]]:_(s64), [[V2:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES %{{[0-9]+}}(<2 x s64>)
+
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $d0 = COPY [[V1]](s64)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT1:%[0-9]+]]:_(s64) = COPY $d0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: ADJCALLSTACKDOWN
+    ; CHECK-DAG: $d0 = COPY [[V2]](s64)
+    ; CHECK-DAG: BL &log2
+    ; CHECK-DAG: [[ELT2:%[0-9]+]]:_(s64) = COPY $d0
+    ; CHECK-DAG: ADJCALLSTACKUP
+
+    ; CHECK-DAG: %1:_(<2 x s64>) = G_BUILD_VECTOR [[ELT1]](s64), [[ELT2]](s64)
+
+    %0:_(<2 x s64>) = COPY $q0
+    %1:_(<2 x s64>) = G_FLOG2 %0
+    $q0 = COPY %1(<2 x s64>)
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:            test_log2_half
+alignment:       2
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+body:             |
+  bb.0:
+    liveins: $h0
+    ; CHECK-LABEL: name:            test_log2_half
+    ; CHECK: [[REG1:%[0-9]+]]:_(s32) = G_FPEXT %0(s16)
+    ; CHECK-NEXT: ADJCALLSTACKDOWN
+    ; CHECK-NEXT: $s0 = COPY [[REG1]](s32)
+    ; CHECK-NEXT: BL &log2
+    ; CHECK-NEXT: [[REG2:%[0-9]+]]:_(s32) = COPY $s0
+    ; CHECK-NEXT: ADJCALLSTACKUP
+    ; CHECK-NEXT: [[RES:%[0-9]+]]:_(s16) = G_FPTRUNC [[REG2]](s32)
+
+    %0:_(s16) = COPY $h0
+    %1:_(s16) = G_FLOG2 %0
+    $h0 = COPY %1(s16)
+    RET_ReallyLR implicit $h0

test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir

 #
 # DEBUG-NEXT: G_FEXP2 (opcode {{[0-9]+}}): 1 type index
 # DEBUG:      .. type index coverage check SKIPPED: no rules defined
 #
 # DEBUG-NEXT: G_FLOG (opcode {{[0-9]+}}): 1 type index
 # DEBUG:      .. the first uncovered type index: 1, OK
 #
 # DEBUG-NEXT: G_FLOG2 (opcode {{[0-9]+}}): 1 type index
-# DEBUG:      .. type index coverage check SKIPPED: no rules defined
+# DEBUG:      .. the first uncovered type index: 1, OK
 #
 # DEBUG-NEXT: G_FLOG10 (opcode {{[0-9]+}}): 1 type index
 # DEBUG:      .. the first uncovered type index: 1, OK
 #
 # DEBUG-NEXT: G_FNEG (opcode {{[0-9]+}}): 1 type index
 # DEBUG:      .. the first uncovered type index: 1, OK
 #
 # DEBUG-NEXT: G_FPEXT (opcode {{[0-9]+}}): 2 type indices

test/CodeGen/AArch64/arm64-vfloatintrinsics.ll

define %v4f16 @test_v4f16.log10(%v4f16 %a) {
   ; This operation is expanded, whether with or without +fullfp16.
   ; CHECK-LABEL:   test_v4f16.log10:
   ; CHECK-COUNT-4: bl log10
   ; GISEL-LABEL:   test_v4f16.log10:
   ; GISEL-COUNT-4: bl log10
   %1 = call %v4f16 @llvm.log10.v4f16(%v4f16 %a)
   ret %v4f16 %1
 }
+
+; FALLBACK-NOT: remark{{.*}}test_v4f16.log2
 define %v4f16 @test_v4f16.log2(%v4f16 %a) {
   ; This operation is expanded, whether with or without +fullfp16.
   ; CHECK-LABEL:   test_v4f16.log2:
   ; CHECK-COUNT-4: bl log2
+  ; GISEL-LABEL:   test_v4f16.log2:
+  ; GISEL-COUNT-4: bl log2
   %1 = call %v4f16 @llvm.log2.v4f16(%v4f16 %a)
   ret %v4f16 %1
 }
 define %v4f16 @test_v4f16.fma(%v4f16 %a, %v4f16 %b, %v4f16 %c) {
   ; CHECK-LABEL:          test_v4f16.fma:
   ; CHECK-NOFP16-COUNT-4: fmadd s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}
   ; CHECK-FP16-NOT:       fcvt
   ; CHECK-FP16:           fmla.4h
define %v8f16 @test_v8f16.log10(%v8f16 %a) {
   ; This operation is expanded, whether with or without +fullfp16.
   ; CHECK-LABEL:   test_v8f16.log10:
   ; CHECK-COUNT-8: bl log10
   ; GISEL-LABEL:   test_v8f16.log10:
   ; GISEL-COUNT-8: bl log10
   %1 = call %v8f16 @llvm.log10.v8f16(%v8f16 %a)
   ret %v8f16 %1
 }
+
+; FALLBACK-NOT: remark{{.*}}test_v8f16.log2
 define %v8f16 @test_v8f16.log2(%v8f16 %a) {
   ; This operation is expanded, whether with or without +fullfp16.
   ; CHECK-LABEL:   test_v8f16.log2:
   ; CHECK-COUNT-8: bl log2
+  ; GISEL-LABEL:   test_v8f16.log2:
+  ; GISEL-COUNT-8: bl log2
   %1 = call %v8f16 @llvm.log2.v8f16(%v8f16 %a)
   ret %v8f16 %1
 }
 define %v8f16 @test_v8f16.fma(%v8f16 %a, %v8f16 %b, %v8f16 %c) {
   ; CHECK-LABEL:          test_v8f16.fma:
   ; CHECK-NOFP16-COUNT-8: fmadd s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}
   ; CHECK-FP16-NOT:       fcvt
   ; CHECK-FP16:           fmla.8h

 ; CHECK: test_v2f32.log10:
 ; GISEL: test_v2f32.log10:
 define %v2f32 @test_v2f32.log10(%v2f32 %a) {
   ; CHECK: log
   ; GISEL: log
   %1 = call %v2f32 @llvm.log10.v2f32(%v2f32 %a)
   ret %v2f32 %1
 }
+
+; FALLBACK-NOT: remark{{.*}}test_v2f32.log2
 ; CHECK: test_v2f32.log2:
+; GISEL: test_v2f32.log2:
 define %v2f32 @test_v2f32.log2(%v2f32 %a) {
   ; CHECK: log
+  ; GISEL: log
   %1 = call %v2f32 @llvm.log2.v2f32(%v2f32 %a)
   ret %v2f32 %1
 }
 ; CHECK-LABEL: test_v2f32.fma:
 define %v2f32 @test_v2f32.fma(%v2f32 %a, %v2f32 %b, %v2f32 %c) {
   ; CHECK: fmla.2s
   %1 = call %v2f32 @llvm.fma.v2f32(%v2f32 %a, %v2f32 %b, %v2f32 %c)
   ret %v2f32 %1

 ; FALLBACK-NOT: remark{{.*}}test_v4f32.log10
 ; CHECK: test_v4f32.log10:
 define %v4f32 @test_v4f32.log10(%v4f32 %a) {
   ; CHECK: log
   ; GISEL: log
   %1 = call %v4f32 @llvm.log10.v4f32(%v4f32 %a)
   ret %v4f32 %1
 }
+
+; FALLBACK-NOT: remark{{.*}}test_v4f32.log2
 ; CHECK: test_v4f32.log2:
+; GISEL: test_v4f32.log2:
 define %v4f32 @test_v4f32.log2(%v4f32 %a) {
   ; CHECK: log
+  ; GISEL: log
   %1 = call %v4f32 @llvm.log2.v4f32(%v4f32 %a)
   ret %v4f32 %1
 }
 ; CHECK: test_v4f32.fma:
 define %v4f32 @test_v4f32.fma(%v4f32 %a, %v4f32 %b, %v4f32 %c) {
   ; CHECK: fma
   %1 = call %v4f32 @llvm.fma.v4f32(%v4f32 %a, %v4f32 %b, %v4f32 %c)
   ret %v4f32 %1

 ; CHECK: test_v2f64.log10:
 ; GISEL: test_v2f64.log10:
 define %v2f64 @test_v2f64.log10(%v2f64 %a) {
   ; CHECK: log
   ; GISEL: log
   %1 = call %v2f64 @llvm.log10.v2f64(%v2f64 %a)
   ret %v2f64 %1
 }
+
+; FALLBACK-NOT: remark{{.*}}test_v2f64.log2
 ; CHECK: test_v2f64.log2:
+; GISEL: test_v2f64.log2:
 define %v2f64 @test_v2f64.log2(%v2f64 %a) {
   ; CHECK: log
+  ; GISEL: log
   %1 = call %v2f64 @llvm.log2.v2f64(%v2f64 %a)
   ret %v2f64 %1
 }
 ; CHECK: test_v2f64.fma:
 define %v2f64 @test_v2f64.fma(%v2f64 %a, %v2f64 %b, %v2f64 %c) {
   ; CHECK: fma
   %1 = call %v2f64 @llvm.fma.v2f64(%v2f64 %a, %v2f64 %b, %v2f64 %c)
   ret %v2f64 %1

test/CodeGen/AArch64/f16-instructions.ll

 ; GISEL-NEXT: ldp x29, x30, [sp], #16
 ; GISEL-NEXT: ret
 
 define half @test_log10(half %a) #0 {
   %r = call half @llvm.log10.f16(half %a)
   ret half %r
 }
 
+; FALLBACK-NOT: remark:{{.*}}test_log2
+; FALLBACK-FP16-NOT: remark:{{.*}}test_log2
+
 ; CHECK-COMMON-LABEL: test_log2:
 ; CHECK-COMMON-NEXT: stp x29, x30, [sp, #-16]!
 ; CHECK-COMMON-NEXT: mov  x29, sp
 ; CHECK-COMMON-NEXT: fcvt s0, h0
 ; CHECK-COMMON-NEXT: bl {{_?}}log2f
 ; CHECK-COMMON-NEXT: fcvt h0, s0
 ; CHECK-COMMON-NEXT: ldp x29, x30, [sp], #16
 ; CHECK-COMMON-NEXT: ret
+
+; GISEL-LABEL: test_log2:
+; GISEL-NEXT: stp x29, x30, [sp, #-16]!
+; GISEL-NEXT: mov  x29, sp
+; GISEL-NEXT: fcvt s0, h0
+; GISEL-NEXT: bl {{_?}}log2f
+; GISEL-NEXT: fcvt h0, s0
+; GISEL-NEXT: ldp x29, x30, [sp], #16
+; GISEL-NEXT: ret
+
 define half @test_log2(half %a) #0 {
   %r = call half @llvm.log2.f16(half %a)
   ret half %r
 }
 
 ; CHECK-CVT-LABEL: test_fma:
 ; CHECK-CVT-NEXT: fcvt s2, h2
 ; CHECK-CVT-NEXT: fcvt s1, h1