exp on SIMD.Vec triggers LLVM error

[pxl-th]

To reproduce

The Minimal Working Example (MWE) for this bug:

using CUDA
using Adapt
using SIMD
using KernelAbstractions

@inline function vload(::Type{SIMD.Vec{N, T}}, ptr::Core.LLVMPtr{T, AS}) where {N, T, AS}
    alignment = sizeof(T) * N
    vec_ptr = Base.bitcast(Core.LLVMPtr{SIMD.Vec{N, T}, AS}, ptr)
    return unsafe_load(vec_ptr, 1, Val(alignment))
end

@inline function vstore!(ptr::Core.LLVMPtr{T, AS}, x::SIMD.Vec{N, T}) where {N, T, AS}
    alignment = sizeof(T) * N
    vec_ptr = Base.bitcast(Core.LLVMPtr{SIMD.Vec{N, T}, AS}, ptr)
    unsafe_store!(vec_ptr, x, 1, Val(alignment))
    return
end

@kernel function ker!(y, x)
    i = @index(Global)
    v = vload(SIMD.Vec{4, Float32}, pointer(x))
    v = exp(v)
    vstore!(pointer(y), v)
end

function tt(kab)
    x = Adapt.adapt(kab, ones(Float32, 4))
    y = Adapt.adapt(kab, ones(Float32, 4))
    ker!(kab)(y, x; ndrange=1)
    @show y
    return
end
tt(CUDABackend())

Error:

ERROR: LLVM error: Undefined external symbol "expf"
Stacktrace:
  [1] handle_error(reason::Cstring)
    @ LLVM ~/.julia/packages/LLVM/b3kFs/src/core/context.jl:194
  [2] LLVMTargetMachineEmitToMemoryBuffer(T::LLVM.TargetMachine, M::LLVM.Module, codegen::LLVM.API.LLVMCodeGenFileType, ErrorMessage::Base.RefValue{Cstring}, OutMemBuf::Base.RefValue{Ptr{…}})
    @ LLVM.API ~/.julia/packages/LLVM/b3kFs/lib/16/libLLVM.jl:11138
  [3] emit(tm::LLVM.TargetMachine, mod::LLVM.Module, filetype::LLVM.API.LLVMCodeGenFileType)
    @ LLVM ~/.julia/packages/LLVM/b3kFs/src/targetmachine.jl:118
  [4] mcgen(job::GPUCompiler.CompilerJob, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/OGnEB/src/mcgen.jl:75
  [5] mcgen(job::GPUCompiler.CompilerJob{GPUCompiler.PTXCompilerTarget, CUDA.CUDACompilerParams}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ CUDA ~/.julia/packages/CUDA/1kIOw/src/compiler/compilation.jl:127
  [6] macro expansion
    @ ~/.julia/packages/GPUCompiler/OGnEB/src/driver.jl:400 [inlined]
  [7] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module; strip::Bool, validate::Bool, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/OGnEB/src/utils.jl:110
  [8] emit_asm
    @ ~/.julia/packages/GPUCompiler/OGnEB/src/utils.jl:108 [inlined]
  [9] codegen(output::Symbol, job::GPUCompiler.CompilerJob; toplevel::Bool, libraries::Bool, optimize::Bool, cleanup::Bool, validate::Bool, strip::Bool, only_entry::Bool, parent_job::Nothing)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/OGnEB/src/driver.jl:120
...

[maleadt]

Generated LLVM IR:

; ModuleID = 'start'
source_filename = "start"
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"

; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare <4 x float> @llvm.exp.v4f32(<4 x float>) #0

; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.x() #0

; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() #0

define ptx_kernel void @_Z8gpu_ker_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILi1E5TupleI5OneToI5Int64EEE7NDRangeILi1ES0_S0_S8_S8_EE13CuDeviceArrayI7Float32Li1ELi1EESE_({ i64, i32 } %state, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, { i8 addrspace(1)*, i64, [1 x i64], i64 } %1, { i8 addrspace(1)*, i64, [1 x i64], i64 } %2) local_unnamed_addr !dbg !9 {
conversion:
  %.fca.0.0.0.0.extract = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, 0, 0, 0, 0
  %.fca.1.1.0.0.0.extract = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, 1, 1, 0, 0, 0
  %3 = call i32 @llvm.nvvm.read.ptx.sreg.ctaid.x(), !dbg !13, !range !32
  %4 = call i32 @llvm.nvvm.read.ptx.sreg.tid.x(), !dbg !33, !range !40
  %5 = add nuw nsw i32 %4, 1, !dbg !41
  %6 = zext i32 %5 to i64, !dbg !44
  %7 = zext i32 %3 to i64, !dbg !65
  %8 = mul i64 %.fca.1.1.0.0.0.extract, %7, !dbg !73
  %9 = add i64 %8, %6, !dbg !75
  %10 = icmp slt i64 %9, 1, !dbg !76
  %11 = icmp sgt i64 %9, %.fca.0.0.0.0.extract, !dbg !76
  %.not16 = or i1 %10, %11, !dbg !28
  br i1 %.not16, label %L246, label %L113, !dbg !28

L113:                                             ; preds = %conversion
  %.fca.0.extract4 = extractvalue { i8 addrspace(1)*, i64, [1 x i64], i64 } %1, 0
  %12 = bitcast i8 addrspace(1)* %.fca.0.extract4 to <4 x float> addrspace(1)*, !dbg !88
  %.fca.0.extract2 = extractvalue { i8 addrspace(1)*, i64, [1 x i64], i64 } %2, 0
  %13 = bitcast i8 addrspace(1)* %.fca.0.extract2 to <4 x float> addrspace(1)*, !dbg !103
  %.unpack = load <4 x float>, <4 x float> addrspace(1)* %13, align 16, !dbg !103, !tbaa !113
  %14 = call <4 x float> @llvm.exp.v4f32(<4 x float> %.unpack), !dbg !116
  store <4 x float> %14, <4 x float> addrspace(1)* %12, align 16, !dbg !88, !tbaa !113
  br label %L246, !dbg !88

L246:                                             ; preds = %L113, %conversion
  ret void, !dbg !125
}

attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

!llvm.module.flags = !{!0, !1}
!llvm.dbg.cu = !{!2, !4, !5, !6}
!julia.kernel = !{!7}
!nvvm.annotations = !{!8}

!0 = !{i32 2, !"Dwarf Version", i32 2}
!1 = !{i32 2, !"Debug Info Version", i32 3}
!2 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, nameTableKind: None)
!3 = !DIFile(filename: "julia", directory: ".")
!4 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, nameTableKind: None)
!5 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, nameTableKind: None)
!6 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, nameTableKind: None)
!7 = !{void ({ i64, i32 }, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] }, { i8 addrspace(1)*, i64, [1 x i64], i64 }, { i8 addrspace(1)*, i64, [1 x i64], i64 })* @_Z8gpu_ker_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILi1E5TupleI5OneToI5Int64EEE7NDRangeILi1ES0_S0_S8_S8_EE13CuDeviceArrayI7Float32Li1ELi1EESE_}
!8 = !{void ({ i64, i32 }, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] }, { i8 addrspace(1)*, i64, [1 x i64], i64 }, { i8 addrspace(1)*, i64, [1 x i64], i64 })* @_Z8gpu_ker_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILi1E5TupleI5OneToI5Int64EEE7NDRangeILi1ES0_S0_S8_S8_EE13CuDeviceArrayI7Float32Li1ELi1EESE_, !"kernel", i32 1}
!9 = distinct !DISubprogram(name: "gpu_ker!", linkageName: "julia_gpu_ker!_25692", scope: null, file: !10, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!10 = !DIFile(filename: "none", directory: ".")
!11 = !DISubroutineType(types: !12)
!12 = !{}
!13 = !DILocation(line: 39, scope: !14, inlinedAt: !16)
!14 = distinct !DISubprogram(name: "macro expansion;", linkageName: "macro expansion", scope: !15, file: !15, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!15 = !DIFile(filename: "/home/tim/Julia/pkg/LLVM/src/interop/base.jl", directory: ".")
!16 = !DILocation(line: 7, scope: !17, inlinedAt: !19)
!17 = distinct !DISubprogram(name: "macro expansion;", linkageName: "macro expansion", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!18 = !DIFile(filename: "/home/tim/Julia/pkg/CUDA/src/device/intrinsics/indexing.jl", directory: ".")
!19 = !DILocation(line: 7, scope: !20, inlinedAt: !21)
!20 = distinct !DISubprogram(name: "_index;", linkageName: "_index", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!21 = !DILocation(line: 56, scope: !22, inlinedAt: !23)
!22 = distinct !DISubprogram(name: "blockIdx_x;", linkageName: "blockIdx_x", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!23 = !DILocation(line: 78, scope: !24, inlinedAt: !25)
!24 = distinct !DISubprogram(name: "#blockIdx;", linkageName: "#blockIdx", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!25 = !DILocation(line: 168, scope: !26, inlinedAt: !28)
!26 = distinct !DISubprogram(name: "#__validindex;", linkageName: "#__validindex", scope: !27, file: !27, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!27 = !DIFile(filename: "/home/tim/Julia/pkg/CUDA/src/CUDAKernels.jl", directory: ".")
!28 = !DILocation(line: 96, scope: !29, inlinedAt: !31)
!29 = distinct !DISubprogram(name: "gpu_ker!;", linkageName: "gpu_ker!", scope: !30, file: !30, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!30 = !DIFile(filename: "/home/tim/Julia/pkg/KernelAbstractions/src/macros.jl", directory: ".")
!31 = !DILocation(line: 0, scope: !9)
!32 = !{i32 0, i32 2147483646}
!33 = !DILocation(line: 39, scope: !14, inlinedAt: !34)
!34 = !DILocation(line: 7, scope: !17, inlinedAt: !35)
!35 = !DILocation(line: 7, scope: !20, inlinedAt: !36)
!36 = !DILocation(line: 46, scope: !37, inlinedAt: !38)
!37 = distinct !DISubprogram(name: "threadIdx_x;", linkageName: "threadIdx_x", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!38 = !DILocation(line: 92, scope: !39, inlinedAt: !25)
!39 = distinct !DISubprogram(name: "#threadIdx;", linkageName: "#threadIdx", scope: !18, file: !18, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!40 = !{i32 0, i32 1023}
!41 = !DILocation(line: 87, scope: !42, inlinedAt: !36)
!42 = distinct !DISubprogram(name: "+;", linkageName: "+", scope: !43, file: !43, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!43 = !DIFile(filename: "int.jl", directory: ".")
!44 = !DILocation(line: 811, scope: !45, inlinedAt: !47)
!45 = distinct !DISubprogram(name: "toInt64;", linkageName: "toInt64", scope: !46, file: !46, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!46 = !DIFile(filename: "boot.jl", directory: ".")
!47 = !DILocation(line: 892, scope: !48, inlinedAt: !49)
!48 = distinct !DISubprogram(name: "Int64;", linkageName: "Int64", scope: !46, file: !46, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!49 = !DILocation(line: 7, scope: !50, inlinedAt: !52)
!50 = distinct !DISubprogram(name: "convert;", linkageName: "convert", scope: !51, file: !51, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!51 = !DIFile(filename: "number.jl", directory: ".")
!52 = !DILocation(line: 307, scope: !53, inlinedAt: !55)
!53 = distinct !DISubprogram(name: "to_index;", linkageName: "to_index", scope: !54, file: !54, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!54 = !DIFile(filename: "indices.jl", directory: ".")
!55 = !DILocation(line: 292, scope: !53, inlinedAt: !56)
!56 = !DILocation(line: 368, scope: !57, inlinedAt: !58)
!57 = distinct !DISubprogram(name: "to_indices;", linkageName: "to_indices", scope: !54, file: !54, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!58 = !DILocation(line: 365, scope: !57, inlinedAt: !59)
!59 = !DILocation(line: 1312, scope: !60, inlinedAt: !62)
!60 = distinct !DISubprogram(name: "getindex;", linkageName: "getindex", scope: !61, file: !61, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!61 = !DIFile(filename: "abstractarray.jl", directory: ".")
!62 = !DILocation(line: 84, scope: !63, inlinedAt: !25)
!63 = distinct !DISubprogram(name: "expand;", linkageName: "expand", scope: !64, file: !64, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!64 = !DIFile(filename: "/home/tim/Julia/pkg/KernelAbstractions/src/nditeration.jl", directory: ".")
!65 = !DILocation(line: 86, scope: !66, inlinedAt: !67)
!66 = distinct !DISubprogram(name: "-;", linkageName: "-", scope: !43, file: !43, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!67 = !DILocation(line: 78, scope: !68, inlinedAt: !69)
!68 = distinct !DISubprogram(name: "#1;", linkageName: "#1", scope: !64, file: !64, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!69 = !DILocation(line: 48, scope: !70, inlinedAt: !72)
!70 = distinct !DISubprogram(name: "ntuple;", linkageName: "ntuple", scope: !71, file: !71, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!71 = !DIFile(filename: "ntuple.jl", directory: ".")
!72 = !DILocation(line: 74, scope: !63, inlinedAt: !62)
!73 = !DILocation(line: 88, scope: !74, inlinedAt: !67)
!74 = distinct !DISubprogram(name: "*;", linkageName: "*", scope: !43, file: !43, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!75 = !DILocation(line: 87, scope: !42, inlinedAt: !67)
!76 = !DILocation(line: 514, scope: !77, inlinedAt: !78)
!77 = distinct !DISubprogram(name: "<=;", linkageName: "<=", scope: !43, file: !43, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!78 = !DILocation(line: 1426, scope: !79, inlinedAt: !81)
!79 = distinct !DISubprogram(name: "in;", linkageName: "in", scope: !80, file: !80, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!80 = !DIFile(filename: "range.jl", directory: ".")
!81 = !DILocation(line: 382, scope: !82, inlinedAt: !84)
!82 = distinct !DISubprogram(name: "map;", linkageName: "map", scope: !83, file: !83, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!83 = !DIFile(filename: "tuple.jl", directory: ".")
!84 = !DILocation(line: 477, scope: !85, inlinedAt: !87)
!85 = distinct !DISubprogram(name: "in;", linkageName: "in", scope: !86, file: !86, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!86 = !DIFile(filename: "multidimensional.jl", directory: ".")
!87 = !DILocation(line: 169, scope: !26, inlinedAt: !28)
!88 = !DILocation(line: 39, scope: !14, inlinedAt: !89)
!89 = !DILocation(line: 0, scope: !90, inlinedAt: !91)
!90 = distinct !DISubprogram(name: "macro expansion;", linkageName: "macro expansion", scope: !10, file: !10, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!91 = !DILocation(line: 0, scope: !92, inlinedAt: !93)
!92 = distinct !DISubprogram(name: "pointerset;", linkageName: "pointerset", scope: !10, file: !10, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!93 = !DILocation(line: 88, scope: !94, inlinedAt: !96)
!94 = distinct !DISubprogram(name: "unsafe_store!;", linkageName: "unsafe_store!", scope: !95, file: !95, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!95 = !DIFile(filename: "/home/tim/Julia/pkg/LLVM/src/interop/pointer.jl", directory: ".")
!96 = !DILocation(line: 4, scope: !97, inlinedAt: !99)
!97 = distinct !DISubprogram(name: "vstore!;", linkageName: "vstore!", scope: !98, file: !98, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!98 = !DIFile(filename: "REPL[15]", directory: ".")
!99 = !DILocation(line: 5, scope: !100, inlinedAt: !102)
!100 = distinct !DISubprogram(name: "macro expansion;", linkageName: "macro expansion", scope: !101, file: !101, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!101 = !DIFile(filename: "REPL[16]", directory: ".")
!102 = !DILocation(line: 97, scope: !29, inlinedAt: !31)
!103 = !DILocation(line: 39, scope: !14, inlinedAt: !104)
!104 = !DILocation(line: 0, scope: !90, inlinedAt: !105)
!105 = !DILocation(line: 0, scope: !106, inlinedAt: !107)
!106 = distinct !DISubprogram(name: "pointerref;", linkageName: "pointerref", scope: !10, file: !10, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!107 = !DILocation(line: 85, scope: !108, inlinedAt: !109)
!108 = distinct !DISubprogram(name: "unsafe_load;", linkageName: "unsafe_load", scope: !95, file: !95, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!109 = !DILocation(line: 4, scope: !110, inlinedAt: !112)
!110 = distinct !DISubprogram(name: "vload;", linkageName: "vload", scope: !111, file: !111, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!111 = !DIFile(filename: "REPL[14]", directory: ".")
!112 = !DILocation(line: 3, scope: !100, inlinedAt: !102)
!113 = !{!114, !114, i64 0, i64 0}
!114 = !{!"custom_tbaa_addrspace(1)", !115, i64 0}
!115 = !{!"custom_tbaa"}
!116 = !DILocation(line: 131, scope: !117, inlinedAt: !119)
!117 = distinct !DISubprogram(name: "macro expansion;", linkageName: "macro expansion", scope: !118, file: !118, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!118 = !DIFile(filename: "/home/tim/.julia/packages/SIMD/hyXY3/src/LLVM_intrinsics.jl", directory: ".")
!119 = !DILocation(line: 127, scope: !120, inlinedAt: !121)
!120 = distinct !DISubprogram(name: "exp;", linkageName: "exp", scope: !118, file: !118, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!121 = !DILocation(line: 156, scope: !122, inlinedAt: !124)
!122 = distinct !DISubprogram(name: "exp;", linkageName: "exp", scope: !123, file: !123, type: !11, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !12)
!123 = !DIFile(filename: "/home/tim/.julia/packages/SIMD/hyXY3/src/simdvec.jl", directory: ".")
!124 = !DILocation(line: 4, scope: !100, inlinedAt: !102)
!125 = !DILocation(line: 99, scope: !29, inlinedAt: !31)

❯ llc /tmp/test.ll
error: no libcall available for fexp

The @llvm.exp.v4f32 call is generated directly (i.e. using llvmcall) by SIMD.jl. Not all back-ends (here, NVPTX) support all combinations of these intrinsics, which is explicitly documented in the LLVM language reference:

You can use llvm.exp on any floating-point or vector of floating-point type. Not all targets support all types however.

... which makes them pretty annoying, as you can't rely on them.