@time in global while loop causes internal error

[KristofferC]

julia> while true
           @time 1
       end
Internal error: encountered unexpected error in runtime:
BoundsError(a=Array{Core.Compiler.BasicBlock, (10,)}[Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=1, stop=8), preds=Array{Int64, (1,)}[10], succs=Array{Int64, (1,)}[2]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=9, stop=14), preds=Array{Int64, (1,)}[1], succs=Array{Int64, (2,)}[5, 3]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=15, stop=17), preds=Array{Int64, (1,)}[2], succs=Array{Int64, (1,)}[4]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=18, stop=18), preds=Array{Int64, (1,)}[3], succs=Array{Int64, (1,)}[7]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=19, stop=24), preds=Array{Int64, (1,)}[2], succs=Array{Int64, (1,)}[6]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=25, stop=25), preds=Array{Int64, (1,)}[5], succs=Array{Int64, (1,)}[7]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=26, stop=42), preds=Array{Int64, (2,)}[4, 6], succs=Array{Int64, (2,)}[9, 8]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=43, stop=44), preds=Array{Int64, (1,)}[7], succs=Array{Int64, (0,)}[]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=45, stop=72), preds=Array{Int64, (1,)}[7], succs=Array{Int64, (1,)}[10]), Core.Compiler.BasicBlock(stmts=Core.Compiler.StmtRange(start=73, stop=77), preds=Array{Int64, (1,)}[9], succs=Array{Int64, (1,)}[1])], i=(0,))
ijl_bounds_error_ints at /home/kc/julia/src/rtutils.c:194
getindex at ./essentials.jl:13 [inlined]
getindex at ./abstractarray.jl:1263 [inlined]
replace_code_newstyle! at ./compiler/ssair/legacy.jl:89
ir_to_codeinf! at ./compiler/optimize.jl:154 [inlined]
transform_result_for_cache at ./compiler/typeinfer.jl:364 [inlined]
cache_result! at ./compiler/typeinfer.jl:395
_typeinf at ./compiler/typeinfer.jl:277
typeinf at ./compiler/typeinfer.jl:213
typeinf_ext_toplevel at ./compiler/typeinfer.jl:1073
typeinf_ext_toplevel at ./compiler/typeinfer.jl:1060
jfptr_typeinf_ext_toplevel_18438 at /home/kc/julia/usr/lib/julia/sys.so (unknown line)

Seems to have been introduced in 1.7.

[vtjnash]

I think @ianatol had looked into this a few times, and was just deciding how to best fix this?

[ianatol]

Ref: #37154

I haven't really worked on it since last year, but happy to check it out soon

[inkydragon]

versioninfo()
while true
    @time 1
end

• ✔️ 1.6.4 (2021-11-19)
• ✔️ 1.7.0-rc1 (2021-09-12)
• ✔️ 1.7.0-rc1.33 (2021-09-23) cd462dc
• ❌ 1.7.0-rc1.34 (2021-09-23) 8ebf8fd
  Fix bug in @time compilation time measurement, using _tryfinally macro. #41923 ping @NHDaly
• ❌ 1.7.3 (2022-05-06)
• ❌ 1.8.0-rc4 (2022-08-08)
• ❌ 1.9.0-DEV.1125 (2022-08-13)

[NHDaly]

CC: @IanButterworth alas, this will never end.

Thanks for the careful report, @inkydragon! @IanButterworth: do you have any ideas? That PR was mostly just reopening your old PR; I don't immediately see anything wrong.

[IanButterworth]

@ianatol from the above I get the impression this is a compiler issue, and so the fix isn't to change the time macro code?

[ianatol]

@IanButterworth I can't confirm for sure, but it looks very similar to the issues in #37154. It may be possible to change the @time macro to not trigger it, but the underlying internal error is what really needs to be fixed

[martinholters]

MWE:

julia> while true
           x = try finally end
       end

[martinholters]

So what's happening is that we produce IR code where the first instructions are PhiNode with one of their edges being 0.
Indeed, this simple modificiotn of above MWE does not produce the problem:

julia> begin
           nothing
           while true
               x = try finally end
           end
       end

With this patch (which probably is not the right thing to do):

--- a/base/compiler/ssair/legacy.jl
+++ b/base/compiler/ssair/legacy.jl
@@ -86,7 +86,7 @@ function replace_code_newstyle!(ci::CodeInfo, ir::IRCode, nargs::Int)
         elseif isa(stmt, GotoIfNot)
             code[i] = GotoIfNot(stmt.cond, first(ir.cfg.blocks[stmt.dest].stmts))
         elseif isa(stmt, PhiNode)
-            code[i] = PhiNode(Int32[last(ir.cfg.blocks[edge].stmts) for edge in stmt.edges], stmt.values)
+            code[i] = PhiNode(Int32[edge == 0 ? 0 : last(ir.cfg.blocks[edge].stmts) for edge in stmt.edges], stmt.values)
         elseif isexpr(stmt, :enter)
             stmt.args[1] = first(ir.cfg.blocks[stmt.args[1]::Int].stmts)
             code[i] = stmt

I obtain the following:

julia> f() = while true x = try finally end end
f (generic function with 1 method)

julia> @code_typed f()
CodeInfo(
1 ─ %1  = φ (#1 => false, #9 => %15)::Bool
└── %2  = φ (#1 => #undef, #9 => %16)::Nothing
2 ─       $(Expr(:enter, #5))
3 ─ %4  = ϒ (%1)::Bool
│   %5  = ϒ (%2)::Nothing
│   %6  = ϒ (true)::Bool
│   %7  = ϒ (nothing)::Nothing
└──       $(Expr(:leave, 1))
4 ─       goto #7
5 ┄ %10 = φᶜ (%4, %6)::Bool
│   %11 = φᶜ (%5, %7)::Nothing
└──       $(Expr(:leave, 1))
6 ─       nothing::Nothing
7 ┄ %14 = φ (#4 => -1, #6 => 1)::Int64
│   %15 = φ (#4 => true, #6 => %10)::Bool
│   %16 = φ (#4 => nothing, #6 => %11)::Nothing
│   %17 = (%14 === 1)::Bool
└──       goto #9 if not %17
8 ─       invoke Base.rethrow()::Union{}
└──       unreachable
9 ─       $(Expr(:throw_undef_if_not, Symbol(""), :(%15)))::Any
└──       goto #1
) => Union{}

Can our PhiNodes represent "code entry" as an edge, or would we have to synthesize a nothing::Nothing block as first instruction? Like for

julia> g() = (nothing; while true x = try finally end end)
g (generic function with 1 method)

julia> @code_typed g()
CodeInfo(
1 ──       nothing::Nothing
2 ┄─ %2  = φ (#1 => false, #10 => %16)::Bool
└─── %3  = φ (#1 => #undef, #10 => %17)::Nothing
3 ──       $(Expr(:enter, #6))
# ... as above, shifted by one instruction/one block

[martinholters]

With this patch (which probably is not the right thing to do):

... or maybe it is? I do get the same output with @code_llvm for f() and g(), which is rather promising...