llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
26.77k stars 10.97k forks source link

[SPIR-V] Incorrect builtin implementation work_group_any #97312

Open michalpaszkowski opened 4 days ago

michalpaszkowski commented 4 days ago
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-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v12
8:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024"
target triple = "spir64-unknown-unknown"

; Function Attrs: nounwind
define spir_kernel void @test_wg_any(ptr addrspace(1) %input, ptr addrspace(1) %output) #0 !kernel_arg_addr_space !7 !kernel_arg_access_qual !8 !kernel_arg_type !9
 !kernel_arg_type_qual !10 !kernel_arg_base_type !9 !spirv.ParameterDecorations !11 {
entry:
  %0 = call spir_func i64 @_Z13get_global_idj(i32 0) #1
  %1 = insertelement <3 x i64> undef, i64 %0, i32 0
  %2 = call spir_func i64 @_Z13get_global_idj(i32 1) #1
  %3 = insertelement <3 x i64> %1, i64 %2, i32 1
  %4 = call spir_func i64 @_Z13get_global_idj(i32 2) #1
  %5 = insertelement <3 x i64> %3, i64 %4, i32 2
  %call = extractelement <3 x i64> %5, i32 0
  %conv = trunc i64 %call to i32
  %idxprom = sext i32 %conv to i64
  %arrayidx = getelementptr inbounds float, ptr addrspace(1) %input, i64 %idxprom
  %6 = load float, ptr addrspace(1) %arrayidx, align 4
  %add = add nsw i32 %conv, 1
  %idxprom1 = sext i32 %add to i64
  %arrayidx2 = getelementptr inbounds float, ptr addrspace(1) %input, i64 %idxprom1
  %7 = load float, ptr addrspace(1) %arrayidx2, align 4
  %cmp = fcmp ogt float %6, %7
  %conv3 = select i1 %cmp, i32 1, i32 0
  %8 = icmp ne i32 %conv3, 0
  %9 = zext i1 %8 to i32
  %10 = call spir_func i32 @_Z14work_group_anyi(i32 %9) #2
  %call41 = icmp ne i32 %10, 0
  %call4 = select i1 %call41, i32 1, i32 0
  %idxprom5 = sext i32 %conv to i64
  %arrayidx6 = getelementptr inbounds i32, ptr addrspace(1) %output, i64 %idxprom5
  store i32 %call4, ptr addrspace(1) %arrayidx6, align 4
  ret void
}

; Function Attrs: nounwind willreturn memory(none)
declare spir_func i64 @_Z13get_global_idj(i32) #1

; Function Attrs: convergent nounwind
declare spir_func i32 @_Z14work_group_anyi(i32) #2

attributes #0 = { nounwind }
attributes #1 = { nounwind willreturn memory(none) }
attributes #2 = { convergent nounwind }

!spirv.MemoryModel = !{!0}
!opencl.enable.FP_CONTRACT = !{}
!spirv.Source = !{!1}
!opencl.spir.version = !{!2}
!opencl.ocl.version = !{!3}
!opencl.used.extensions = !{!4}
!opencl.used.optional.core.features = !{!5}
!spirv.Generator = !{!6}

!0 = !{i32 2, i32 2}
!1 = !{i32 3, i32 300000}
!2 = !{i32 2, i32 0}
!3 = !{i32 3, i32 0}
!4 = !{!"cl_khr_subgroups"}
!5 = !{}
!6 = !{i16 6, i16 14}
!7 = !{i32 1, i32 1}
!8 = !{!"none", !"none"}
!9 = !{!"float*", !"int*"}
!10 = !{!"", !""}
!11 = !{!5, !5}

Assertion hit when compiling with SPIR-V backend:

SPIRVBuiltins.cpp:1058: bool llvm::generateGroupInst(const llvm::SPIRV::IncomingCall*, llvm::MachineIRBuilder&, llvm::SPIRVGlobalRegistry*): Assertion `ArgInstruction->getOpcode() == TargetOpcode::G_CONSTANT && "Only constant bool value args are supported"' failed.
michalpaszkowski commented 4 days ago

The above code comes from recent SPIR-V Translator (back translation) from LLVM IR dumped by most recent IGC. The test passes in IGC, since the original LLVM IR coming from the OpenCL CTS test suite has a different builtin call:

%call41 = call spir_func i1 @OpGroupAny_i32_i1(i32 2, i1 %3)
...
declare spir_func i1 @OpGroupAny_i32_i1(i32, i1)