LLVM error "ran out of registers during register allocation" on 32-bit x86 with code-coverage #42200
Description
This comes from the discussion on Discourse.
I found a strange error "ran out of registers during register allocation" in the CI log of a package.
The error occurs on 32-bit x86, but I couldn't figure out the trigger for a long time. Later, @ChenNingCong found that the code for code coverage had something to do with it.
The following is an MWE. I think the easiest way to reproduce the error is to write it in runtests.jl and run it with pkg> test --coverage.
struct RGB8
r::UInt8
g::UInt8
b::UInt8
end
Base.broadcastable(c::RGB8) = Ref(c)
_mapc(f, x::RGB8, y::RGB8) = RGB8(f(x.r, y.r), f(x.g, y.g), f(x.b, y.b))
function lighten(c1::RGB8, c2::RGB8, opacity::AbstractFloat)
mixed = _mapc(max, c1, c2)
w(v1, v2) = round(UInt8, v1 * (1 - opacity) + v2 * opacity)
_mapc(w, c1, mixed)
end
c = RGB8(255, 0, 0)
cs = (RGB8(255, 0, 0), RGB8(0, 255, 0))
f(c1, c2, opacity) = lighten.(c1, c2, opacity)
using InteractiveUtils
versioninfo()
@code_llvm f(c, cs, 0.8)
@code_native f(c, cs, 0.8) # errorOutput of v1.6.2 on Windows (i686-w64-mingw32)
Julia Version 1.6.2
Commit 1b93d53fc4 (2021-07-14 15:36 UTC)
Platform Info:
OS: Windows (i686-w64-mingw32)
CPU: 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
WORD_SIZE: 32
LIBM: libopenlibm
LLVM: libLLVM-11.0.1 (ORCJIT, tigerlake)
Environment:
JULIA_LOAD_PATH = @;C:\Users\username\AppData\Local\Temp\jl_iMZPOq
; @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:18 within `f'
; Function Attrs: alignstack(16)
define void @julia_f_224([2 x [3 x i8]]* noalias nocapture sret %0, [3 x i8]* nocapture nonnull readonly align 1 deref
erenceable(3) %1, [2 x [3 x i8]]* nocapture nonnull readonly align 1 dereferenceable(6) %2, double %3) #0 {
top:
%4 = alloca [3 x i8], align 1
%5 = alloca [3 x i8], align 1
%6 = alloca [1 x double], align 8
%7 = alloca [3 x i8], align 1
%8 = alloca [3 x i8], align 1
%9 = alloca [1 x double], align 8
%lcnt = load volatile i64, i64* inttoptr (i32 390202632 to i64*), align 8
%10 = add i64 %lcnt, 1
store volatile i64 %10, i64* inttoptr (i32 390202632 to i64*), align 8
; ┌ @ broadcast.jl:1312 within `broadcasted'
; │┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:6 within `broadcastable'
%lcnt1 = load volatile i64, i64* inttoptr (i32 390202536 to i64*), align 8
%11 = add i64 %lcnt1, 1
store volatile i64 %11, i64* inttoptr (i32 390202536 to i64*), align 8
; ││┌ @ refpointer.jl:134 within `Ref'
; │││┌ @ refvalue.jl:10 within `RefValue' @ refvalue.jl:8
%.elt = getelementptr inbounds [3 x i8], [3 x i8]* %1, i32 0, i32 0
%.unpack = load i8, i8* %.elt, align 1
%.elt35 = getelementptr inbounds [3 x i8], [3 x i8]* %1, i32 0, i32 1
%.unpack36 = load i8, i8* %.elt35, align 1
%.elt37 = getelementptr inbounds [3 x i8], [3 x i8]* %1, i32 0, i32 2
%.unpack38 = load i8, i8* %.elt37, align 1
; └└└└
; ┌ @ broadcast.jl:883 within `materialize'
; │┌ @ broadcast.jl:1098 within `copy'
; ││┌ @ ntuple.jl:49 within `ntuple'
; │││┌ @ broadcast.jl:1098 within `#19'
; ││││┌ @ broadcast.jl:620 within `_broadcast_getindex'
; │││││┌ @ broadcast.jl:644 within `_getindex'
; ││││││┌ @ broadcast.jl:595 within `_broadcast_getindex'
; │││││││┌ @ refvalue.jl:56 within `getindex'
; ││││││││┌ @ Base.jl:33 within `getproperty'
%.fca.0.gep30 = getelementptr inbounds [3 x i8], [3 x i8]* %4, i32 0, i32 0
store i8 %.unpack, i8* %.fca.0.gep30, align 1
%.fca.1.gep32 = getelementptr inbounds [3 x i8], [3 x i8]* %4, i32 0, i32 1
store i8 %.unpack36, i8* %.fca.1.gep32, align 1
%.fca.2.gep34 = getelementptr inbounds [3 x i8], [3 x i8]* %4, i32 0, i32 2
store i8 %.unpack38, i8* %.fca.2.gep34, align 1
; ││││└└└└└
; ││││┌ @ broadcast.jl:621 within `_broadcast_getindex'
; │││││┌ @ broadcast.jl:648 within `_broadcast_getindex_evalf'
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:11 within `lighten'
%lcnt2 = load volatile i64, i64* inttoptr (i32 390202576 to i64*), align 16
%12 = add i64 %lcnt2, 1
store volatile i64 %12, i64* inttoptr (i32 390202576 to i64*), align 16
; │││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:8 within `_mapc'
%lcnt3 = load volatile i64, i64* inttoptr (i32 390202552 to i64*), align 8
%13 = add i64 %lcnt3, 1
store volatile i64 %13, i64* inttoptr (i32 390202552 to i64*), align 8
; ││││││││┌ @ Base.jl:33 within `getproperty'
%14 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 0, i32 0
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%15 = load i8, i8* %14, align 1
%.not = icmp ult i8 %15, %.unpack
; │││││││││└
%16 = select i1 %.not, i8 %.unpack, i8 %15
; ││││││││└
; ││││││││┌ @ Base.jl:33 within `getproperty'
%17 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 0, i32 1
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%18 = load i8, i8* %17, align 1
%.not50 = icmp ult i8 %18, %.unpack36
; │││││││││└
%19 = select i1 %.not50, i8 %.unpack36, i8 %18
; ││││││││└
; ││││││││┌ @ Base.jl:33 within `getproperty'
%20 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 0, i32 2
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%21 = load i8, i8* %20, align 1
%.not51 = icmp ult i8 %21, %.unpack38
; │││││││││└
%22 = select i1 %.not51, i8 %.unpack38, i8 %21
; ││││││││└
; ││││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:2 within `RGB8'
%lcnt4 = load volatile i64, i64* inttoptr (i32 390202504 to i64*), align 8
%23 = add i64 %lcnt4, 1
store volatile i64 %23, i64* inttoptr (i32 390202504 to i64*), align 8
%24 = getelementptr inbounds [3 x i8], [3 x i8]* %5, i32 0, i32 0
store i8 %16, i8* %24, align 1
%25 = getelementptr inbounds [3 x i8], [3 x i8]* %5, i32 0, i32 1
store i8 %19, i8* %25, align 1
%26 = getelementptr inbounds [3 x i8], [3 x i8]* %5, i32 0, i32 2
store i8 %22, i8* %26, align 1
; ││││││└└└
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:12 within `lighten'
%lcnt5 = load volatile i64, i64* inttoptr (i32 390202584 to i64*), align 8
%27 = add i64 %lcnt5, 1
store volatile i64 %27, i64* inttoptr (i32 390202584 to i64*), align 8
%28 = getelementptr inbounds [1 x double], [1 x double]* %6, i32 0, i32 0
store double %3, double* %28, align 8
; ││││││└
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:13 within `lighten'
%lcnt6 = load volatile i64, i64* inttoptr (i32 390202592 to i64*), align 32
%29 = add i64 %lcnt6, 1
store volatile i64 %29, i64* inttoptr (i32 390202592 to i64*), align 32
%30 = call [3 x i8] @j__mapc_226([1 x double]* nocapture readonly %6, [3 x i8]* nocapture readonly %4, [3 x i
8]* nocapture readonly %5) #0
%.fca.0.extract20 = extractvalue [3 x i8] %30, 0
%.fca.1.extract22 = extractvalue [3 x i8] %30, 1
%.fca.2.extract24 = extractvalue [3 x i8] %30, 2
; ││││└└└
; ││││┌ @ broadcast.jl:620 within `_broadcast_getindex'
; │││││┌ @ broadcast.jl:644 within `_getindex'
; ││││││┌ @ broadcast.jl:595 within `_broadcast_getindex'
; │││││││┌ @ refvalue.jl:56 within `getindex'
; ││││││││┌ @ Base.jl:33 within `getproperty'
%.fca.0.gep = getelementptr inbounds [3 x i8], [3 x i8]* %7, i32 0, i32 0
store i8 %.unpack, i8* %.fca.0.gep, align 1
%.fca.1.gep = getelementptr inbounds [3 x i8], [3 x i8]* %7, i32 0, i32 1
store i8 %.unpack36, i8* %.fca.1.gep, align 1
%.fca.2.gep = getelementptr inbounds [3 x i8], [3 x i8]* %7, i32 0, i32 2
store i8 %.unpack38, i8* %.fca.2.gep, align 1
; ││││└└└└└
; ││││┌ @ broadcast.jl:621 within `_broadcast_getindex'
; │││││┌ @ broadcast.jl:648 within `_broadcast_getindex_evalf'
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:11 within `lighten'
%lcnt7 = load volatile i64, i64* inttoptr (i32 390202576 to i64*), align 16
%31 = add i64 %lcnt7, 1
store volatile i64 %31, i64* inttoptr (i32 390202576 to i64*), align 16
; │││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:8 within `_mapc'
%lcnt8 = load volatile i64, i64* inttoptr (i32 390202552 to i64*), align 8
%32 = add i64 %lcnt8, 1
store volatile i64 %32, i64* inttoptr (i32 390202552 to i64*), align 8
; ││││││││┌ @ Base.jl:33 within `getproperty'
%33 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 1, i32 0
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%34 = load i8, i8* %33, align 1
%.not58 = icmp ult i8 %34, %.unpack
; │││││││││└
%35 = select i1 %.not58, i8 %.unpack, i8 %34
; ││││││││└
; ││││││││┌ @ Base.jl:33 within `getproperty'
%36 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 1, i32 1
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%37 = load i8, i8* %36, align 1
%.not59 = icmp ult i8 %37, %.unpack36
; │││││││││└
%38 = select i1 %.not59, i8 %.unpack36, i8 %37
; ││││││││└
; ││││││││┌ @ Base.jl:33 within `getproperty'
%39 = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %2, i32 0, i32 1, i32 2
; ││││││││└
; ││││││││┌ @ promotion.jl:421 within `max'
; │││││││││┌ @ int.jl:441 within `<'
%40 = load i8, i8* %39, align 1
%.not60 = icmp ult i8 %40, %.unpack38
; │││││││││└
%41 = select i1 %.not60, i8 %.unpack38, i8 %40
; ││││││││└
; ││││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:2 within `RGB8'
%lcnt9 = load volatile i64, i64* inttoptr (i32 390202504 to i64*), align 8
%42 = add i64 %lcnt9, 1
store volatile i64 %42, i64* inttoptr (i32 390202504 to i64*), align 8
%43 = getelementptr inbounds [3 x i8], [3 x i8]* %8, i32 0, i32 0
store i8 %35, i8* %43, align 1
%44 = getelementptr inbounds [3 x i8], [3 x i8]* %8, i32 0, i32 1
store i8 %38, i8* %44, align 1
%45 = getelementptr inbounds [3 x i8], [3 x i8]* %8, i32 0, i32 2
store i8 %41, i8* %45, align 1
; ││││││└└└
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:12 within `lighten'
%lcnt10 = load volatile i64, i64* inttoptr (i32 390202584 to i64*), align 8
%46 = add i64 %lcnt10, 1
store volatile i64 %46, i64* inttoptr (i32 390202584 to i64*), align 8
%47 = getelementptr inbounds [1 x double], [1 x double]* %9, i32 0, i32 0
store double %3, double* %47, align 8
; ││││││└
; ││││││┌ @ C:\Users\username\.julia\dev\RanOutOfRegisters\test\runtests.jl:13 within `lighten'
%lcnt11 = load volatile i64, i64* inttoptr (i32 390202592 to i64*), align 32
%48 = add i64 %lcnt11, 1
store volatile i64 %48, i64* inttoptr (i32 390202592 to i64*), align 32
%49 = call [3 x i8] @j__mapc_227([1 x double]* nocapture readonly %9, [3 x i8]* nocapture readonly %7, [3 x i
8]* nocapture readonly %8) #0
%.fca.0.extract = extractvalue [3 x i8] %49, 0
%.fca.1.extract = extractvalue [3 x i8] %49, 1
%.fca.2.extract = extractvalue [3 x i8] %49, 2
; └└└└└└└
%.sroa.012.sroa.0.0..sroa.012.0..sroa_idx.sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32
0, i32 0, i32 0
store i8 %.fca.0.extract20, i8* %.sroa.012.sroa.0.0..sroa.012.0..sroa_idx.sroa_idx, align 1
%.sroa.012.sroa.2.0..sroa.012.0..sroa_idx.sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32
0, i32 0, i32 1
store i8 %.fca.1.extract22, i8* %.sroa.012.sroa.2.0..sroa.012.0..sroa_idx.sroa_idx, align 1
%.sroa.012.sroa.3.0..sroa.012.0..sroa_idx.sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32
0, i32 0, i32 2
store i8 %.fca.2.extract24, i8* %.sroa.012.sroa.3.0..sroa.012.0..sroa_idx.sroa_idx, align 1
%.sroa.215.0..sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32 0, i32 1, i32 0
store i8 %.fca.0.extract, i8* %.sroa.215.0..sroa_idx, align 1
%.sroa.316.0..sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32 0, i32 1, i32 1
store i8 %.fca.1.extract, i8* %.sroa.316.0..sroa_idx, align 1
%.sroa.4.0..sroa_idx = getelementptr inbounds [2 x [3 x i8]], [2 x [3 x i8]]* %0, i32 0, i32 1, i32 2
store i8 %.fca.2.extract, i8* %.sroa.4.0..sroa_idx, align 1
ret void
}
error: ran out of registers during register allocation
ERROR: Package RanOutOfRegisters errored during testingThis problem occurs in the following versions (on both Linux and Windows):
- v1.5.4 (LLVM 9.0)
- v1.6.2 (LLVM 11.0)
- v1.7.0-beta4 (LLVM 12.0)
- v1.8.0-DEV.498 (LLVM 12.0)
On the other hand, it does not seem to occur on x86-64 or 32-/64-bit ARM.
If you have difficulty preparing a 32-bit x86 environment, you can also use GitHub Actions to reproduce the error.