@fastmath maximum segfaults for Float16 on master
#49907
Labels
compiler:llvm
For issues that relate to LLVM
domain:float16
domain:fold
sum, maximum, reduce, foldl, etc.
kind:regression
Regression in behavior compared to a previous version
kind:upstream
The issue is with an upstream dependency, e.g. LLVM
Milestone
Comments
|
Can reproduce. Note that it can be triggered by julia> versioninfo()
Julia Version 1.10.0-DEV.1351
Commit a6ad9ea099f (2023-05-21 08:01 UTC)
Platform Info:
OS: Linux (x86_64-linux-gnu)
CPU: 12 × Intel(R) Xeon(R) CPU E5-2603 v4 @ 1.70GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-15.0.7 (ORCJIT, broadwell)
Threads: 5 on 12 virtual cores
Environment:
LD_LIBRARY_PATH = /usr/local/cuda/lib64
JULIA_NUM_THREADS = 4
julia> @fastmath max(Float16(1), Float16(2))
Float16(2.0)
julia> @fastmath reduce(max, Float16[1,2,3])
Float16(3.0)
julia> @fastmath reduce(max, Float16[1,2,3]; init = Float16(0))
LLVM ERROR: Cannot select: 0x204fef8: v16f16 = X86ISD::FMAX nnan ninf nsz arcp contract afn reassoc 0x2031f30, 0x1eec5d0, array.jl:938 @[ reduce.jl:60 @[ reduce.jl:48 @[ reduce.jl:44 ] ] ]
0x2031f30: v16f16,ch = CopyFromReg 0x1d83098, Register:v16f16 %9, array.jl:938 @[ reduce.jl:60 @[ reduce.jl:48 @[ reduce.jl:44 ] ] ]Also triggered by some julia> foldl(Base.FastMath.max_fast, Float16[1, 2, 3])
LLVM ERROR: Cannot select: 0x248ba98: v16f16 = X86ISD::FMAX nnan ninf nsz arcp contract afn reassoc 0x24abec0, 0x24819c8, array.jl:938 @[ reduce.jl:60 @[ reduce.jl:48 @[ reduce.jl:44 ] ] ]On the same machine, a version from before #48153 does not have the problem: julia> @fastmath reduce(max, Float16[1,2,3]; init = Float16(0))
Float16(3.0)
julia> versioninfo()
Julia Version 1.10.0-DEV.220
Commit 9ded051e9f8 (2022-12-29 10:05 UTC)
Platform Info:
OS: Linux (x86_64-linux-gnu)On an M1 mac, the problem does not seem to occur: julia> @fastmath reduce(max, Float16[1,2,3]; init = Float16(0))
Float16(3.0)
julia> versioninfo()
Julia Version 1.10.0-DEV.1351
Commit a6ad9ea099 (2023-05-21 08:01 UTC)
Platform Info:
OS: macOS (arm64-apple-darwin21.6.0)
CPU: 8 × Apple M1
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-15.0.7 (ORCJIT, apple-m1)
Threads: 5 on 4 virtual cores |
|
This is probably an issue with the demote float16 pass. It would be cool to see the LLVM IR generated on the function that crashes. |
julia> @code_llvm Base.FastMath.maximum_fast(Float16[1, 2, 3]; init = Float16(0)); @ fastmath.jl:380 within `maximum_fast`
define half @julia_maximum_fast_289([1 x half]* nocapture noundef nonnull readonly align 2 dereferenceable(2) %0, {}* noundef nonnull align 16 dereferenceable(40) %1) #0 {
top:
%thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #9
%ppgcstack_i8 = getelementptr i8, i8* %thread_ptr, i64 -8
%ppgcstack = bitcast i8* %ppgcstack_i8 to {}****
%pgcstack = load {}***, {}**** %ppgcstack, align 8
%ptls_field16 = getelementptr inbounds {}**, {}*** %pgcstack, i64 2
%2 = bitcast {}*** %ptls_field16 to i64***
%ptls_load1718 = load i64**, i64*** %2, align 8
%3 = getelementptr inbounds i64*, i64** %ptls_load1718, i64 2
%safepoint = load i64*, i64** %3, align 8
fence syncscope("singlethread") seq_cst
%4 = load volatile i64, i64* %safepoint, align 8
fence syncscope("singlethread") seq_cst
; ┌ @ fastmath.jl:380 within `#maximum_fast#1`
; │┌ @ reducedim.jl:406 within `reduce`
; ││┌ @ reducedim.jl:406 within `#reduce#811`
; │││┌ @ reducedim.jl:357 within `mapreduce`
%5 = getelementptr inbounds [1 x half], [1 x half]* %0, i64 0, i64 0
; ││││┌ @ reducedim.jl:357 within `#mapreduce#809`
; │││││┌ @ reducedim.jl:362 within `_mapreduce_dim`
; ││││││┌ @ reduce.jl:44 within `mapfoldl_impl`
; │││││││┌ @ reduce.jl:48 within `foldl_impl`
; ││││││││┌ @ reduce.jl:56 within `_foldl_impl`
; │││││││││┌ @ array.jl:938 within `iterate` @ array.jl:938
; ││││││││││┌ @ essentials.jl:10 within `length`
%6 = bitcast {}* %1 to { i8*, i64, i16, i16, i32 }*
%7 = getelementptr inbounds { i8*, i64, i16, i16, i32 }, { i8*, i64, i16, i16, i32 }* %6, i64 0, i32 1
%8 = load i64, i64* %7, align 8
; ││││││││││└
; ││││││││││┌ @ int.jl:520 within `<` @ int.jl:513
%.not = icmp eq i64 %8, 0
; ││││││││││└
br i1 %.not, label %L19, label %L20
L19: ; preds = %top
%9 = load half, half* %5, align 2
br label %L55
L20: ; preds = %top
; ││││││││││┌ @ essentials.jl:13 within `getindex`
%10 = bitcast {}* %1 to half**
%11 = load half*, half** %10, align 8
%12 = load half, half* %11, align 2
; │││││││││└└
; │││││││││ @ reduce.jl:58 within `_foldl_impl`
; │││││││││┌ @ reduce.jl:86 within `BottomRF`
; ││││││││││┌ @ fastmath.jl:251 within `max_fast`
; │││││││││││┌ @ fastmath.jl:191 within `gt_fast`
; ││││││││││││┌ @ fastmath.jl:189 within `lt_fast`
%13 = load half, half* %5, align 2
; │││││││││││└└
; │││││││││││┌ @ essentials.jl:621 within `ifelse`
%.inv = fcmp fast olt half %13, %12
%14 = select fast i1 %.inv, half %12, half %13
; │││││││││└└└
; │││││││││ @ reduce.jl:60 within `_foldl_impl`
; │││││││││┌ @ array.jl:938 within `iterate`
; ││││││││││┌ @ int.jl:520 within `<` @ int.jl:513
%.not1926.not = icmp eq i64 %8, 1
; ││││││││││└
br i1 %.not1926.not, label %L55, label %iter.check
iter.check: ; preds = %L20
%15 = add nsw i64 %8, -1
%min.iters.check = icmp ult i64 %15, 8
br i1 %min.iters.check, label %vec.epilog.scalar.ph, label %vector.main.loop.iter.check
vector.main.loop.iter.check: ; preds = %iter.check
%min.iters.check29 = icmp ult i64 %15, 32
br i1 %min.iters.check29, label %vec.epilog.ph, label %vector.ph
vector.ph: ; preds = %vector.main.loop.iter.check
%n.vec = and i64 %15, -32
%minmax.ident.splatinsert = insertelement <16 x half> poison, half %14, i64 0
%minmax.ident.splat = shufflevector <16 x half> %minmax.ident.splatinsert, <16 x half> poison, <16 x i32> zeroinitializer
br label %vector.body
vector.body: ; preds = %vector.body, %vector.ph
%index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
%vec.phi = phi <16 x half> [ %minmax.ident.splat, %vector.ph ], [ %22, %vector.body ]
%vec.phi30 = phi <16 x half> [ %minmax.ident.splat, %vector.ph ], [ %23, %vector.body ]
%offset.idx = or i64 %index, 1
; ││││││││││┌ @ essentials.jl:13 within `getindex`
%16 = getelementptr inbounds half, half* %11, i64 %offset.idx
%17 = bitcast half* %16 to <16 x half>*
%wide.load = load <16 x half>, <16 x half>* %17, align 2
%18 = getelementptr inbounds half, half* %16, i64 16
%19 = bitcast half* %18 to <16 x half>*
%wide.load31 = load <16 x half>, <16 x half>* %19, align 2
; │││││││││└└
; │││││││││ @ reduce.jl:62 within `_foldl_impl`
; │││││││││┌ @ reduce.jl:86 within `BottomRF`
; ││││││││││┌ @ fastmath.jl:251 within `max_fast`
; │││││││││││┌ @ essentials.jl:621 within `ifelse`
%20 = fcmp fast olt <16 x half> %vec.phi, %wide.load
%21 = fcmp fast olt <16 x half> %vec.phi30, %wide.load31
%22 = select <16 x i1> %20, <16 x half> %wide.load, <16 x half> %vec.phi
%23 = select <16 x i1> %21, <16 x half> %wide.load31, <16 x half> %vec.phi30
%index.next = add nuw i64 %index, 32
%24 = icmp eq i64 %index.next, %n.vec
br i1 %24, label %middle.block, label %vector.body
middle.block: ; preds = %vector.body
; │││││││││└└└
; │││││││││ @ reduce.jl:60 within `_foldl_impl`
; │││││││││┌ @ array.jl:938 within `iterate`
%25 = call fast <16 x half> @llvm.maxnum.v16f16(<16 x half> %22, <16 x half> %23)
%26 = call fast half @llvm.vector.reduce.fmax.v16f16(<16 x half> %25)
%cmp.n = icmp eq i64 %15, %n.vec
br i1 %cmp.n, label %L55, label %vec.epilog.iter.check
vec.epilog.iter.check: ; preds = %middle.block
%ind.end36 = or i64 %n.vec, 2
%ind.end34 = or i64 %n.vec, 1
%n.vec.remaining = and i64 %15, 24
%min.epilog.iters.check = icmp eq i64 %n.vec.remaining, 0
br i1 %min.epilog.iters.check, label %vec.epilog.scalar.ph, label %vec.epilog.ph
vec.epilog.ph: ; preds = %vec.epilog.iter.check, %vector.main.loop.iter.check
%bc.merge.rdx = phi half [ %14, %vector.main.loop.iter.check ], [ %26, %vec.epilog.iter.check ]
%vec.epilog.resume.val = phi i64 [ 0, %vector.main.loop.iter.check ], [ %n.vec, %vec.epilog.iter.check ]
%n.vec33 = and i64 %15, -8
%ind.end = or i64 %n.vec33, 1
%ind.end35 = or i64 %n.vec33, 2
%minmax.ident.splatinsert41 = insertelement <8 x half> poison, half %bc.merge.rdx, i64 0
%minmax.ident.splat42 = shufflevector <8 x half> %minmax.ident.splatinsert41, <8 x half> poison, <8 x i32> zeroinitializer
br label %vec.epilog.vector.body
vec.epilog.vector.body: ; preds = %vec.epilog.vector.body, %vec.epilog.ph
%index39 = phi i64 [ %vec.epilog.resume.val, %vec.epilog.ph ], [ %index.next45, %vec.epilog.vector.body ]
%vec.phi40 = phi <8 x half> [ %minmax.ident.splat42, %vec.epilog.ph ], [ %30, %vec.epilog.vector.body ]
%offset.idx43 = or i64 %index39, 1
; ││││││││││┌ @ essentials.jl:13 within `getindex`
%27 = getelementptr inbounds half, half* %11, i64 %offset.idx43
%28 = bitcast half* %27 to <8 x half>*
%wide.load44 = load <8 x half>, <8 x half>* %28, align 2
; │││││││││└└
; │││││││││ @ reduce.jl:62 within `_foldl_impl`
; │││││││││┌ @ reduce.jl:86 within `BottomRF`
; ││││││││││┌ @ fastmath.jl:251 within `max_fast`
; │││││││││││┌ @ essentials.jl:621 within `ifelse`
%29 = fcmp fast olt <8 x half> %vec.phi40, %wide.load44
%30 = select <8 x i1> %29, <8 x half> %wide.load44, <8 x half> %vec.phi40
%index.next45 = add nuw i64 %index39, 8
%31 = icmp eq i64 %index.next45, %n.vec33
br i1 %31, label %vec.epilog.middle.block, label %vec.epilog.vector.body
vec.epilog.middle.block: ; preds = %vec.epilog.vector.body
; │││││││││└└└
; │││││││││ @ reduce.jl:60 within `_foldl_impl`
; │││││││││┌ @ array.jl:938 within `iterate`
%32 = call fast half @llvm.vector.reduce.fmax.v8f16(<8 x half> %30)
%cmp.n38 = icmp eq i64 %15, %n.vec33
br i1 %cmp.n38, label %L55, label %vec.epilog.scalar.ph
vec.epilog.scalar.ph: ; preds = %vec.epilog.middle.block, %vec.epilog.iter.check, %iter.check
%bc.resume.val = phi i64 [ %ind.end, %vec.epilog.middle.block ], [ %ind.end34, %vec.epilog.iter.check ], [ 1, %iter.check ]
%bc.resume.val37 = phi i64 [ %ind.end35, %vec.epilog.middle.block ], [ %ind.end36, %vec.epilog.iter.check ], [ 2, %iter.check ]
%bc.merge.rdx46 = phi half [ %32, %vec.epilog.middle.block ], [ %26, %vec.epilog.iter.check ], [ %14, %iter.check ]
br label %L42
L42: ; preds = %L42, %vec.epilog.scalar.ph
%33 = phi i64 [ %value_phi628, %L42 ], [ %bc.resume.val, %vec.epilog.scalar.ph ]
%value_phi628 = phi i64 [ %36, %L42 ], [ %bc.resume.val37, %vec.epilog.scalar.ph ]
%value_phi527 = phi half [ %37, %L42 ], [ %bc.merge.rdx46, %vec.epilog.scalar.ph ]
; ││││││││││┌ @ essentials.jl:13 within `getindex`
%34 = getelementptr inbounds half, half* %11, i64 %33
%35 = load half, half* %34, align 2
; ││││││││││└
; ││││││││││┌ @ int.jl:87 within `+`
%36 = add nuw nsw i64 %value_phi628, 1
; │││││││││└└
; │││││││││ @ reduce.jl:62 within `_foldl_impl`
; │││││││││┌ @ reduce.jl:86 within `BottomRF`
; ││││││││││┌ @ fastmath.jl:251 within `max_fast`
; │││││││││││┌ @ essentials.jl:621 within `ifelse`
%.inv20 = fcmp fast olt half %value_phi527, %35
%37 = select fast i1 %.inv20, half %35, half %value_phi527
; │││││││││└└└
; │││││││││ @ reduce.jl:60 within `_foldl_impl`
; │││││││││┌ @ array.jl:938 within `iterate`
; ││││││││││┌ @ int.jl:520 within `<` @ int.jl:513
%exitcond.not = icmp eq i64 %value_phi628, %8
; ││││││││││└
br i1 %exitcond.not, label %L55, label %L42
L55: ; preds = %L42, %vec.epilog.middle.block, %middle.block, %L20, %L19
%value_phi4 = phi half [ %9, %L19 ], [ %14, %L20 ], [ %26, %middle.block ], [ %32, %vec.epilog.middle.block ], [ %37, %L42 ]
; └└└└└└└└└└
ret half %value_phi4
}So the problem is that those |
|
I love that llvm creates an intrinsic that it doesn't know how to lower. |
Closed
|
This is llvm/llvm-project#59258, which got fixed in https://reviews.llvm.org/D139078. I saw that @maleadt was adding some patches so could we get this on as well? |
|
I just finished rebuilding all of LLVM 😅 |
|
I'm so sorry |
maleadt
added a commit
that referenced
this issue
Jul 23, 2023
KristofferC
pushed a commit
that referenced
this issue
Jul 24, 2023
- llvm/llvm-project@af39acd closing #50448 - https://reviews.llvm.org/D139078 closing #49907 (cherry picked from commit 092231c)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
On master I get
Complete output
It works fine with Julia 1.9.0.
Float32andFloat64don't seem to be affected, and without@fastmathit also works forFloat16.The text was updated successfully, but these errors were encountered: