fix test_utf8.j #9
Assignees
Labels
bug
Indicates an unexpected problem or unintended behavior
Comments
ghost
assigned 
Closed
Closed
Merged
burrowsa
pushed a commit
to burrowsa/julia
that referenced
this issue
Mar 24, 2014
update to 05c323d [ci skip]
jiahao
added a commit
that referenced
this issue
Jul 18, 2014
- A_mul_B - Ac_mul_b_RFP - qrp - cholpfact, cholpfact! - qrfact, qrfact! - LinAlg.LUTridiagonal - LinAlg.solve Previously, callling any of these resulted in an `ERROR: ____ not defined`. HT astrieanna/TypeCheck.jl/#9
5 tasks
StefanKarpinski
pushed a commit
that referenced
this issue
Feb 8, 2018
Add ＠compat as discussed in #8
StefanKarpinski
added a commit
that referenced
this issue
Feb 8, 2018
Keno
added a commit
that referenced
this issue
Jul 17, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
Keno
added a commit
that referenced
this issue
Jul 17, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
Keno
added a commit
that referenced
this issue
Jul 17, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
KristofferC
pushed a commit
that referenced
this issue
Jul 20, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
(cherry picked from commit 0a12944)
KristofferC
pushed a commit
that referenced
this issue
Aug 26, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
(cherry picked from commit 0a12944)
KristofferC
pushed a commit
that referenced
this issue
Aug 26, 2019
The bug here is a bit subtle, but perhaps best illustrated with
the included test case:
```
function f32579(x::Int64, b::Bool)
if b
x = nothing
end
if isa(x, Int64)
y = x
else
y = x
end
if isa(y, Nothing)
z = y
else
z = y
end
return z === nothing
end
```
The code just after SSA conversion looks like:
```
2 1 ─ goto #3 if not _3
3 2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
5 3 ┄ %3 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %4 = (%3 isa Main.Int64)::Bool
└── goto #5 if not %4
6 4 ─ %6 = π (%3, Int64)
└── goto #6
8 5 ─ %8 = π (%3, Nothing)
10 6 ┄ %9 = φ (#4 => %6, #5 => %8)::Union{Nothing, Int64}
│ %10 = (%9 isa Main.Nothing)::Bool
└── goto #8 if not %10
11 7 ─ %12 = π (%9, Nothing)
└── goto #9
13 8 ─ %14 = π (%9, Int64)
15 9 ┄ %15 = φ (#7 => %12, #8 => %14)::Union{Nothing, Int64}
│ %16 = (%15 === Main.nothing)::Bool
└── return %16
```
Now, we have special code in SROA (despite it not really being an
SROA transform) that looks at `===` and replaces
it by a nest of phis of booleans. The reasoning for this transform
is that it eliminates a use of a value where we only care about the
type and not the content, thus making it more likely that the value
will subsequently be eligible for SROA. In addition, while it goes
along resolving which values feed into any particular phi, it
accumulates and type conditions it encounters along the way.
Thus in the example above, something like the following happens:
- We look at %14, which πs to %9 with an Int64 constraint, so we only
consider the #4 predecessor for %9 (due to the constraint), until
we get to %3, where we again only consider the #1 predecessor,
where we find the argument (of type Int64) and conclude the result
is always false
- Now we pop the next item of the stack from our original phi, look
at %12, which πs to %9 with a Nothing constraint.
At this point we used to terminate the search because we already looked
at %9. However, crucially, we looked at %9 only with an Int64 constraint,
so we missed the fact that `nothing` was in fact a possible value for this
phi. The result was a missing entry in the generated phi node:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
(note the missing #2 predecessor in phi node %3), which would result
in an undefined value at runtime, though in this case LLVM would
have taken advantage of that to just return 0:
```
define i8 @julia_f32579_16051(i64, i8) {
top:
; @ REPL[1]:15 within `f32579'
ret i8 0
}
```
Compare this now to the optimized IR with this patch:
```
1 ─ goto #3 if not b
2 ─ %2 = Main.nothing::Core.Compiler.Const(nothing, false)
3 ┄ %3 = φ (#2 => true, #1 => false)::Bool
│ %4 = φ (#2 => %2, #1 => _2)::Union{Nothing, Int64}
│ %5 = (%4 isa Main.Int64)::Bool
└── goto #5 if not %5
4 ─ %7 = π (%4, Int64)
└── goto #6
5 ─ %9 = π (%4, Nothing)
6 ┄ %10 = φ (#4 => %3, #5 => %3)::Bool
│ %11 = φ (#4 => %7, #5 => %9)::Union{Nothing, Int64}
│ %12 = (%11 isa Main.Nothing)::Bool
└── goto #8 if not %12
7 ─ goto #9
8 ─ nothing::Nothing
9 ┄ %16 = φ (#7 => %10, #8 => %10)::Bool
└── return %16
```
The %3 phi node has its missing entry and the generated LLVM code
correctly returns `b`:
```
define i8 @julia_f32579_16112(i64, i8) {
top:
%2 = and i8 %1, 1
; @ REPL[1]:15 within `f32579'
ret i8 %2
}
```
(cherry picked from commit 0a12944)
aviatesk
added a commit
that referenced
this issue
Oct 5, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- add special SROA handling for `NamedTuple` generated by keyword sorter:
With the change on `structdiff`, IR for a call with type-unstable
keyword arguments after inlining would look like:
```
%1 = tuple(a, b, c)::Tuple{Any, Any, Any}
%2 = NamedTuple{(:a, :b, :c)(%1)::NamedTuple{(:a, :b, :c), _A} where _A<:Tuple{Any, Any, Any}
%3 = Core.getfield(%2, :a)::Any
%4 = Core.getfield(%2, :b)::Any
%5 = Core.getfield(%2, :c)::Any
[... other body of the keyword func ...]
```
We can implement a bit hacky special handling within our SROA pass
that checks if this definition (%2) is partly well-known `NamedTuple`
construction, where its names are fully known, and also checks if its
call argument (%1) is fully-known `tuple` call. In a case when the
length of the `NamedTuple` names and the length of the arguments for
the `tuple` call, we can safely replace those `getfield` calls with
the corresponding `tuple` call argument, while letting the later DCE
pass to delete the constructions of tuple and named-tuple altogether.
With these changes, the IR for the example `NewInstruction` constructor
is fairly optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 5, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- add special SROA handling for `NamedTuple` generated by keyword sorter:
With the change on `structdiff`, IR for a call with type-unstable
keyword arguments after inlining would look like:
```
%1 = tuple(a, b, c)::Tuple{Any, Any, Any}
%2 = NamedTuple{(:a, :b, :c)(%1)::NamedTuple{(:a, :b, :c), _A} where _A<:Tuple{Any, Any, Any}
%3 = Core.getfield(%2, :a)::Any
%4 = Core.getfield(%2, :b)::Any
%5 = Core.getfield(%2, :c)::Any
[... other body of the keyword func ...]
```
We can implement a bit hacky special handling within our SROA pass
that checks if this definition (%2) is partly well-known `NamedTuple`
construction, where its names are fully known, and also checks if its
call argument (%1) is fully-known `tuple` call. In a case when the
length of the `NamedTuple` names and the length of the arguments for
the `tuple` call, we can safely replace those `getfield` calls with
the corresponding `tuple` call argument, while letting the later DCE
pass to delete the constructions of tuple and named-tuple altogether.
With these changes, the IR for the example `NewInstruction` constructor
is fairly optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 7, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- add special SROA handling for `NamedTuple` generated by keyword sorter:
With the change on `structdiff`, IR for a call with type-unstable
keyword arguments after inlining would look like:
```
%1 = tuple(a, b, c)::Tuple{Any, Any, Any}
%2 = NamedTuple{(:a, :b, :c)(%1)::NamedTuple{(:a, :b, :c), _A} where _A<:Tuple{Any, Any, Any}
%3 = Core.getfield(%2, :a)::Any
%4 = Core.getfield(%2, :b)::Any
%5 = Core.getfield(%2, :c)::Any
[... other body of the keyword func ...]
```
We can implement a bit hacky special handling within our SROA pass
that checks if this definition (%2) is partly well-known `NamedTuple`
construction, where its names are fully known, and also checks if its
call argument (%1) is fully-known `tuple` call. In a case when the
length of the `NamedTuple` names and the length of the arguments for
the `tuple` call, we can safely replace those `getfield` calls with
the corresponding `tuple` call argument, while letting the later DCE
pass to delete the constructions of tuple and named-tuple altogether.
With these changes, the IR for the example `NewInstruction` constructor
is fairly optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 8, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- Add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it
directly forms `:splatnew` allocation rather than redirects to the
general `NamedTuple` constructor, that could be confused for abstract
input tuple type.
- Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types.
This improvement lets `inline_splatnew` to handle more abstract
`NamedTuple`s, especially whose names are fully known but its fields
tuple type is abstract.
Those improvements are combined to allow our SROA pass to optimize away
`NamedTuple` and `tuple` calls generated for keyword argument handling.
E.g. the IR for the example `NewInstruction` constructor is now fairly
optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 8, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- Add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it
directly forms `:splatnew` allocation rather than redirects to the
general `NamedTuple` constructor, that could be confused for abstract
input tuple type.
- Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types.
This improvement lets `inline_splatnew` to handle more abstract
`NamedTuple`s, especially whose names are fully known but its fields
tuple type is abstract.
Those improvements are combined to allow our SROA pass to optimize away
`NamedTuple` and `tuple` calls generated for keyword argument handling.
E.g. the IR for the example `NewInstruction` constructor is now fairly
optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 8, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- Add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it
directly forms `:splatnew` allocation rather than redirects to the
general `NamedTuple` constructor, that could be confused for abstract
input tuple type.
- Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types.
This improvement lets `inline_splatnew` to handle more abstract
`NamedTuple`s, especially whose names are fully known but its fields
tuple type is abstract.
Those improvements are combined to allow our SROA pass to optimize away
`NamedTuple` and `tuple` calls generated for keyword argument handling.
E.g. the IR for the example `NewInstruction` constructor is now fairly
optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 8, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- Add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it
directly forms `:splatnew` allocation rather than redirects to the
general `NamedTuple` constructor, that could be confused for abstract
input tuple type.
- Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types.
This improvement lets `inline_splatnew` to handle more abstract
`NamedTuple`s, especially whose names are fully known but its fields
tuple type is abstract.
Those improvements are combined to allow our SROA pass to optimize away
`NamedTuple` and `tuple` calls generated for keyword argument handling.
E.g. the IR for the example `NewInstruction` constructor is now fairly
optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
aviatesk
added a commit
that referenced
this issue
Oct 8, 2022
This commit tries to fix and improve performance for calling keyword
funcs whose arguments types are not fully known but `@nospecialize`-d.
The final result would look like (this particular example is taken from
our Julia-level compiler implementation):
```julia
abstract type CallInfo end
struct NoCallInfo <: CallInfo end
struct NewInstruction
stmt::Any
type::Any
info::CallInfo
line::Union{Int32,Nothing} # if nothing, copy the line from previous statement in the insertion location
flag::Union{UInt8,Nothing} # if nothing, IR flags will be recomputed on insertion
function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo),
line::Union{Int32,Nothing}, flag::Union{UInt8,Nothing})
return new(stmt, type, info, line, flag)
end
end
@nospecialize
function NewInstruction(newinst::NewInstruction;
stmt=newinst.stmt,
type=newinst.type,
info::CallInfo=newinst.info,
line::Union{Int32,Nothing}=newinst.line,
flag::Union{UInt8,Nothing}=newinst.flag)
return NewInstruction(stmt, type, info, line, flag)
end
@Specialize
using BenchmarkTools
struct VirtualKwargs
stmt::Any
type::Any
info::CallInfo
end
vkws = VirtualKwargs(nothing, Any, NoCallInfo())
newinst = NewInstruction(nothing, Any, NoCallInfo(), nothing, nothing)
runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info)
@benchmark runner($newinst, $vkws)
```
> on master
```
BenchmarkTools.Trial: 10000 samples with 186 evaluations.
Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29%
Time (median): 605.608 ns ┊ GC (median): 0.00%
Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85%
█▇▂▆▄ ▁█▇▄▂ ▂
██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █
560 ns Histogram: log(frequency) by time 1.23 μs <
Memory estimate: 32 bytes, allocs estimate: 2.
```
> on this commit
```julia
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00%
Time (median): 3.098 ns ┊ GC (median): 0.00%
Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇█▆▅▄▂
▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
3.08 ns Histogram: frequency by time 3.19 ns <
Memory estimate: 0 bytes, allocs estimate: 0.
```
So for this particular case it achieves roughly 200x speed up.
This is because this commit allows inlining of a call to keyword sorter
as well as removal of `NamedTuple` call.
Especially this commit is composed of the following improvements:
- Add early return case for `structdiff`:
This change improves the return type inference for a case when
compared `NamedTuple`s are type unstable but there is no difference
in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s.
And in such case the optimizer will remove `structdiff` and succeeding
`pairs` calls, letting the keyword sorter to be inlined.
- Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it
directly forms `:splatnew` allocation rather than redirects to the
general `NamedTuple` constructor, that could be confused for abstract
input tuple type.
- Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types.
This improvement lets `inline_splatnew` to handle more abstract
`NamedTuple`s, especially whose names are fully known but its fields
tuple type is abstract.
Those improvements are combined to allow our SROA pass to optimize away
`NamedTuple` and `tuple` calls generated for keyword argument handling.
E.g. the IR for the example `NewInstruction` constructor is now fairly
optimized, like:
```julia
julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info
NewInstruction(newinst; stmt, type, info)
end |> only
2 1 ── %1 = Base.getfield(_2, :line)::Union{Nothing, Int32} │╻╷ Type##kw
│ %2 = Base.getfield(_2, :flag)::Union{Nothing, UInt8} ││┃ getproperty
│ %3 = (isa)(%1, Nothing)::Bool ││
│ %4 = (isa)(%2, Nothing)::Bool ││
│ %5 = (Core.Intrinsics.and_int)(%3, %4)::Bool ││
└─── goto #3 if not %5 ││
2 ── %7 = %new(Main.NewInstruction, _3, _4, _5, nothing, nothing)::NewInstruction NewInstruction
└─── goto #10 ││
3 ── %9 = (isa)(%1, Int32)::Bool ││
│ %10 = (isa)(%2, Nothing)::Bool ││
│ %11 = (Core.Intrinsics.and_int)(%9, %10)::Bool ││
└─── goto #5 if not %11 ││
4 ── %13 = π (%1, Int32) ││
│ %14 = %new(Main.NewInstruction, _3, _4, _5, %13, nothing)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
5 ── %16 = (isa)(%1, Nothing)::Bool ││
│ %17 = (isa)(%2, UInt8)::Bool ││
│ %18 = (Core.Intrinsics.and_int)(%16, %17)::Bool ││
└─── goto #7 if not %18 ││
6 ── %20 = π (%2, UInt8) ││
│ %21 = %new(Main.NewInstruction, _3, _4, _5, nothing, %20)::NewInstruction│││╻ NewInstruction
└─── goto #10 ││
7 ── %23 = (isa)(%1, Int32)::Bool ││
│ %24 = (isa)(%2, UInt8)::Bool ││
│ %25 = (Core.Intrinsics.and_int)(%23, %24)::Bool ││
└─── goto #9 if not %25 ││
8 ── %27 = π (%1, Int32) ││
│ %28 = π (%2, UInt8) ││
│ %29 = %new(Main.NewInstruction, _3, _4, _5, %27, %28)::NewInstruction │││╻ NewInstruction
└─── goto #10 ││
9 ── Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└─── unreachable ││
10 ┄ %33 = φ (#2 => %7, #4 => %14, #6 => %21, #8 => %29)::NewInstruction ││
└─── goto #11 ││
11 ─ return %33 │
=> NewInstruction
```
github-merge-queue bot
pushed a commit
that referenced
this issue
Jul 15, 2023
…d and inlined) (#43322) A follow up attemp to fix #27988. (close #47493 close #50554) Examples: ```julia julia> using LazyArrays julia> bc = @~ @. 1*(1 + 1) + 1*1; julia> bc2 = @~ 1 .* 1 .- 1 .* 1 .^2 .+ 1 .* 1 .+ 1 .^ 3; ``` On master: <details><summary> click for details </summary> <p> ```julia julia> @code_typed Broadcast.flatten(bc).f(1,1,1,1,1) CodeInfo( 1 ─ %1 = Core.getfield(args, 1)::Int64 │ %2 = Core.getfield(args, 2)::Int64 │ %3 = Core.getfield(args, 3)::Int64 │ %4 = Core.getfield(args, 4)::Int64 │ %5 = Core.getfield(args, 5)::Int64 │ %6 = invoke Base.Broadcast.var"#13#14"{Base.Broadcast.var"#16#18"{Base.Broadcast.var"#15#17", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}, typeof(+)}}(Base.Broadcast.var"#16#18"{Base.Broadcast.var"#15#17", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}, typeof(+)}(Base.Broadcast.var"#15#17"(), Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}(Base.Broadcast.var"#15#17"())), Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), +))(%1::Int64, %2::Int64, %3::Vararg{Int64}, %4, %5)::Tuple{Int64, Int64, Vararg{Int64}} │ %7 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), %6)::Tuple{Int64, Int64} │ %8 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), %6)::Tuple{Vararg{Int64}} │ %9 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#16#18"{Base.Broadcast.var"#9#11", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}, typeof(*)}(Base.Broadcast.var"#9#11"(), Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}(Base.Broadcast.var"#15#17"())), Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), *), %8)::Tuple{Int64} │ %10 = Core.getfield(%7, 1)::Int64 │ %11 = Core.getfield(%7, 2)::Int64 │ %12 = Base.mul_int(%10, %11)::Int64 │ %13 = Core.getfield(%9, 1)::Int64 │ %14 = Base.add_int(%12, %13)::Int64 └── return %14 ) => Int64 julia> @code_typed Broadcast.flatten(bc2).f(1,1,1,^,1,Val(2),1,1,^,1,Val(3)) CodeInfo( 1 ─ %1 = Core.getfield(args, 1)::Int64 │ %2 = Core.getfield(args, 2)::Int64 │ %3 = Core.getfield(args, 3)::Int64 │ %4 = Core.getfield(args, 5)::Int64 │ %5 = Core.getfield(args, 7)::Int64 │ %6 = Core.getfield(args, 8)::Int64 │ %7 = Core.getfield(args, 10)::Int64 │ %8 = invoke Base.Broadcast.var"#13#14"{Base.Broadcast.var"#16#18"{Base.Broadcast.var"#15#17", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}}, typeof(Base.literal_pow)}}(Base.Broadcast.var"#16#18"{Base.Broadcast.var"#15#17", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}}, typeof(Base.literal_pow)}(Base.Broadcast.var"#15#17"(), Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}(Base.Broadcast.var"#15#17"()))), Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"()))), Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"()))), Base.literal_pow))(%3::Int64, ^::Function, %4::Vararg{Any}, $(QuoteNode(Val{2}())), %5, %6, ^, %7, $(QuoteNode(Val{3}())))::Tuple{Int64, Any, Vararg{Any}} │ %9 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), %8)::Tuple{Int64, Any} │ %10 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), %8)::Tuple │ %11 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#15#17"(), %10)::Tuple │ %12 = Core.getfield(%9, 1)::Int64 │ %13 = Core.getfield(%9, 2)::Any │ %14 = (*)(%12, %13)::Any │ %15 = Core.tuple(%14)::Tuple{Any} │ %16 = Core._apply_iterate(Base.iterate, Core.tuple, %15, %11)::Tuple{Any, Vararg{Any}} │ %17 = Base.mul_int(%1, %2)::Int64 │ %18 = Core.tuple(%17)::Tuple{Int64} │ %19 = Core._apply_iterate(Base.iterate, Core.tuple, %18, %16)::Tuple{Int64, Any, Vararg{Any}} │ %20 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), %19)::Tuple{Int64, Any} │ %21 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), %19)::Tuple │ %22 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#16#18"{Base.Broadcast.var"#15#17", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}, typeof(*)}(Base.Broadcast.var"#15#17"(), Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}(Base.Broadcast.var"#15#17"())), Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), *), %21)::Tuple{Any, Vararg{Any}} │ %23 = Core.getfield(%20, 1)::Int64 │ %24 = Core.getfield(%20, 2)::Any │ %25 = (-)(%23, %24)::Any │ %26 = Core.tuple(%25)::Tuple{Any} │ %27 = Core._apply_iterate(Base.iterate, Core.tuple, %26, %22)::Tuple{Any, Any, Vararg{Any}} │ %28 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"())), %27)::Tuple{Any, Any} │ %29 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"())), %27)::Tuple │ %30 = Core._apply_iterate(Base.iterate, Base.Broadcast.var"#16#18"{Base.Broadcast.var"#9#11", Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}}, Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}}, Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}}, typeof(Base.literal_pow)}(Base.Broadcast.var"#9#11"(), Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}}(Base.Broadcast.var"#13#14"{Base.Broadcast.var"#15#17"}(Base.Broadcast.var"#15#17"()))), Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}}(Base.Broadcast.var"#23#24"{Base.Broadcast.var"#25#26"}(Base.Broadcast.var"#25#26"()))), Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}}(Base.Broadcast.var"#19#20"{Base.Broadcast.var"#21#22"}(Base.Broadcast.var"#21#22"()))), Base.literal_pow), %29)::Tuple{Any} │ %31 = Core.getfield(%28, 1)::Any │ %32 = Core.getfield(%28, 2)::Any │ %33 = (+)(%31, %32)::Any │ %34 = Core.getfield(%30, 1)::Any │ %35 = (+)(%33, %34)::Any └── return %35 ) => Any ``` </p> </details> On this PR ```julia julia> @code_typed Broadcast.flatten(bc).f(1,1,1,1,1) CodeInfo( 1 ─ %1 = Core.getfield(args, 1)::Int64 │ %2 = Core.getfield(args, 2)::Int64 │ %3 = Core.getfield(args, 3)::Int64 │ %4 = Core.getfield(args, 4)::Int64 │ %5 = Core.getfield(args, 5)::Int64 │ %6 = Base.add_int(%2, %3)::Int64 │ %7 = Base.mul_int(%1, %6)::Int64 │ %8 = Base.mul_int(%4, %5)::Int64 │ %9 = Base.add_int(%7, %8)::Int64 └── return %9 ) => Int64 julia> @code_typed Broadcast.flatten(bc2).f(1,1,1,^,1,Val(2),1,1,^,1,Val(3)) CodeInfo( 1 ─ %1 = Core.getfield(args, 1)::Int64 │ %2 = Core.getfield(args, 2)::Int64 │ %3 = Core.getfield(args, 3)::Int64 │ %4 = Core.getfield(args, 5)::Int64 │ %5 = Core.getfield(args, 7)::Int64 │ %6 = Core.getfield(args, 8)::Int64 │ %7 = Core.getfield(args, 10)::Int64 │ %8 = Base.mul_int(%1, %2)::Int64 │ %9 = Base.mul_int(%4, %4)::Int64 │ %10 = Base.mul_int(%3, %9)::Int64 │ %11 = Base.sub_int(%8, %10)::Int64 │ %12 = Base.mul_int(%5, %6)::Int64 │ %13 = Base.add_int(%11, %12)::Int64 │ %14 = Base.mul_int(%7, %7)::Int64 │ %15 = Base.mul_int(%14, %7)::Int64 │ %16 = Base.add_int(%13, %15)::Int64 └── return %16 ) => Int64 ```
quinnj
pushed a commit
that referenced
this issue
Jan 26, 2024
`@something` eagerly unwraps any `Some` given to it, while keeping the
variable between its arguments the same. This can be an issue if a
previously unpacked value is used as input to `@something`, leading to a
type instability on more than two arguments (e.g. because of a fallback
to `Some(nothing)`). By using different variables for each argument,
type inference has an easier time handling these cases that are isolated
to single branches anyway.
This also adds some comments to the macro, since it's non-obvious what
it does.
Benchmarking the specific case I encountered this in led to a ~2x
performance improvement on multiple machines.
1.10-beta3/master:
```
[sukera@tower 01]$ jl1100 -q --project=. -L 01.jl -e 'bench()'
v"1.10.0-beta3"
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
Range (min … max): 38.670 μs … 70.350 μs ┊ GC (min … max): 0.00% … 0.00%
Time (median): 43.340 μs ┊ GC (median): 0.00%
Time (mean ± σ): 43.395 μs ± 1.518 μs ┊ GC (mean ± σ): 0.00% ± 0.00%
▆█▂ ▁▁
▂▂▂▂▂▂▂▂▂▁▂▂▂▃▃▃▂▂▃▃▃▂▂▂▂▂▄▇███▆██▄▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂ ▃
38.7 μs Histogram: frequency by time 48 μs <
Memory estimate: 0 bytes, allocs estimate: 0.
```
This PR:
```
[sukera@tower 01]$ julia -q --project=. -L 01.jl -e 'bench()'
v"1.11.0-DEV.970"
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
Range (min … max): 22.820 μs … 44.980 μs ┊ GC (min … max): 0.00% … 0.00%
Time (median): 24.300 μs ┊ GC (median): 0.00%
Time (mean ± σ): 24.370 μs ± 832.239 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▂▅▇██▇▆▅▁
▂▂▂▂▂▂▂▂▃▃▄▅▇███████████▅▄▃▃▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▂▂ ▃
22.8 μs Histogram: frequency by time 27.7 μs <
Memory estimate: 0 bytes, allocs estimate: 0.
```
<details>
<summary>Benchmarking code (spoilers for Advent Of Code 2023 Day 01,
Part 01). Running this requires the input of that Advent Of Code
day.</summary>
```julia
using BenchmarkTools
using InteractiveUtils
isdigit(d::UInt8) = UInt8('0') <= d <= UInt8('9')
someDigit(c::UInt8) = isdigit(c) ? Some(c - UInt8('0')) : nothing
function part1(data)
total = 0
may_a = nothing
may_b = nothing
for c in data
digitRes = someDigit(c)
may_a = @something may_a digitRes Some(nothing)
may_b = @something digitRes may_b Some(nothing)
if c == UInt8('\n')
digit_a = may_a::UInt8
digit_b = may_b::UInt8
total += digit_a*0xa + digit_b
may_a = nothing
may_b = nothing
end
end
return total
end
function bench()
data = read("input.txt")
display(VERSION)
println()
display(@benchmark part1($data))
nothing
end
```
</details>
<details>
<summary>`@code_warntype` before</summary>
```julia
julia> @code_warntype part1(data)
MethodInstance for part1(::Vector{UInt8})
from part1(data) @ Main ~/Documents/projects/AOC/2023/01/01.jl:7
Arguments
#self#::Core.Const(part1)
data::Vector{UInt8}
Locals
@_3::Union{Nothing, Tuple{UInt8, Int64}}
may_b::Union{Nothing, UInt8}
may_a::Union{Nothing, UInt8}
total::Int64
c::UInt8
digit_b::UInt8
digit_a::UInt8
val@_10::Any
val@_11::Any
digitRes::Union{Nothing, Some{UInt8}}
@_13::Union{Some{Nothing}, Some{UInt8}, UInt8}
@_14::Union{Some{Nothing}, Some{UInt8}}
@_15::Some{Nothing}
@_16::Union{Some{Nothing}, Some{UInt8}, UInt8}
@_17::Union{Some{Nothing}, UInt8}
@_18::Some{Nothing}
Body::Int64
1 ── (total = 0)
│ (may_a = Main.nothing)
│ (may_b = Main.nothing)
│ %4 = data::Vector{UInt8}
│ (@_3 = Base.iterate(%4))
│ %6 = (@_3 === nothing)::Bool
│ %7 = Base.not_int(%6)::Bool
└─── goto #24 if not %7
2 ┄─ Core.NewvarNode(:(digit_b))
│ Core.NewvarNode(:(digit_a))
│ Core.NewvarNode(:(val@_10))
│ %12 = @_3::Tuple{UInt8, Int64}
│ (c = Core.getfield(%12, 1))
│ %14 = Core.getfield(%12, 2)::Int64
│ (digitRes = Main.someDigit(c))
│ (val@_11 = may_a)
│ %17 = (val@_11::Union{Nothing, UInt8} !== Base.nothing)::Bool
└─── goto #4 if not %17
3 ── (@_13 = val@_11::UInt8)
└─── goto #11
4 ── (val@_11 = digitRes)
│ %22 = (val@_11::Union{Nothing, Some{UInt8}} !== Base.nothing)::Bool
└─── goto #6 if not %22
5 ── (@_14 = val@_11::Some{UInt8})
└─── goto #10
6 ── (val@_11 = Main.Some(Main.nothing))
│ %27 = (val@_11::Core.Const(Some(nothing)) !== Base.nothing)::Core.Const(true)
└─── goto #8 if not %27
7 ── (@_15 = val@_11::Core.Const(Some(nothing)))
└─── goto #9
8 ── Core.Const(:(@_15 = Base.nothing))
9 ┄─ (@_14 = @_15)
10 ┄ (@_13 = @_14)
11 ┄ %34 = @_13::Union{Some{Nothing}, Some{UInt8}, UInt8}
│ (may_a = Base.something(%34))
│ (val@_10 = digitRes)
│ %37 = (val@_10::Union{Nothing, Some{UInt8}} !== Base.nothing)::Bool
└─── goto #13 if not %37
12 ─ (@_16 = val@_10::Some{UInt8})
└─── goto #20
13 ─ (val@_10 = may_b)
│ %42 = (val@_10::Union{Nothing, UInt8} !== Base.nothing)::Bool
└─── goto #15 if not %42
14 ─ (@_17 = val@_10::UInt8)
└─── goto #19
15 ─ (val@_10 = Main.Some(Main.nothing))
│ %47 = (val@_10::Core.Const(Some(nothing)) !== Base.nothing)::Core.Const(true)
└─── goto #17 if not %47
16 ─ (@_18 = val@_10::Core.Const(Some(nothing)))
└─── goto #18
17 ─ Core.Const(:(@_18 = Base.nothing))
18 ┄ (@_17 = @_18)
19 ┄ (@_16 = @_17)
20 ┄ %54 = @_16::Union{Some{Nothing}, Some{UInt8}, UInt8}
│ (may_b = Base.something(%54))
│ %56 = c::UInt8
│ %57 = Main.UInt8('\n')::Core.Const(0x0a)
│ %58 = (%56 == %57)::Bool
└─── goto #22 if not %58
21 ─ (digit_a = Core.typeassert(may_a, Main.UInt8))
│ (digit_b = Core.typeassert(may_b, Main.UInt8))
│ %62 = total::Int64
│ %63 = (digit_a * 0x0a)::UInt8
│ %64 = (%63 + digit_b)::UInt8
│ (total = %62 + %64)
│ (may_a = Main.nothing)
└─── (may_b = Main.nothing)
22 ┄ (@_3 = Base.iterate(%4, %14))
│ %69 = (@_3 === nothing)::Bool
│ %70 = Base.not_int(%69)::Bool
└─── goto #24 if not %70
23 ─ goto #2
24 ┄ return total
```
</details>
<details>
<summary>`@code_native debuginfo=:none` Before </summary>
```julia
julia> @code_native debuginfo=:none part1(data)
.text
.file "part1"
.globl julia_part1_418 # -- Begin function julia_part1_418
.p2align 4, 0x90
.type julia_part1_418,@function
julia_part1_418: # @julia_part1_418
# %bb.0: # %top
push rbp
mov rbp, rsp
push r15
push r14
push r13
push r12
push rbx
sub rsp, 40
mov rax, qword ptr [rdi + 8]
test rax, rax
je .LBB0_1
# %bb.2: # %L17
mov rcx, qword ptr [rdi]
dec rax
mov r10b, 1
xor r14d, r14d
# implicit-def: $r12b
# implicit-def: $r13b
# implicit-def: $r9b
# implicit-def: $sil
mov qword ptr [rbp - 64], rax # 8-byte Spill
mov al, 1
mov dword ptr [rbp - 48], eax # 4-byte Spill
# implicit-def: $al
# kill: killed $al
xor eax, eax
mov qword ptr [rbp - 56], rax # 8-byte Spill
mov qword ptr [rbp - 72], rcx # 8-byte Spill
# implicit-def: $cl
jmp .LBB0_3
.p2align 4, 0x90
.LBB0_8: # in Loop: Header=BB0_3 Depth=1
mov dword ptr [rbp - 48], 0 # 4-byte Folded Spill
.LBB0_24: # %post_union_move
# in Loop: Header=BB0_3 Depth=1
movzx r13d, byte ptr [rbp - 41] # 1-byte Folded Reload
mov r12d, r8d
cmp qword ptr [rbp - 64], r14 # 8-byte Folded Reload
je .LBB0_13
.LBB0_25: # %guard_exit113
# in Loop: Header=BB0_3 Depth=1
inc r14
mov r10d, ebx
.LBB0_3: # %L19
# =>This Inner Loop Header: Depth=1
mov rax, qword ptr [rbp - 72] # 8-byte Reload
xor ebx, ebx
xor edi, edi
movzx r15d, r9b
movzx ecx, cl
movzx esi, sil
mov r11b, 1
# implicit-def: $r9b
movzx edx, byte ptr [rax + r14]
lea eax, [rdx - 58]
lea r8d, [rdx - 48]
cmp al, -10
setae bl
setb dil
test r10b, 1
cmovne r15d, edi
mov edi, 0
cmovne ecx, ebx
mov bl, 1
cmovne esi, edi
test r15b, 1
jne .LBB0_7
# %bb.4: # %L76
# in Loop: Header=BB0_3 Depth=1
mov r11b, 2
test cl, 1
jne .LBB0_5
# %bb.6: # %L78
# in Loop: Header=BB0_3 Depth=1
mov ebx, r10d
mov r9d, r15d
mov byte ptr [rbp - 41], r13b # 1-byte Spill
test sil, 1
je .LBB0_26
.LBB0_7: # %L82
# in Loop: Header=BB0_3 Depth=1
cmp al, -11
jbe .LBB0_9
jmp .LBB0_8
.p2align 4, 0x90
.LBB0_5: # in Loop: Header=BB0_3 Depth=1
mov ecx, r8d
mov sil, 1
xor ebx, ebx
mov byte ptr [rbp - 41], r8b # 1-byte Spill
xor r9d, r9d
xor ecx, ecx
cmp al, -11
ja .LBB0_8
.LBB0_9: # %L90
# in Loop: Header=BB0_3 Depth=1
test byte ptr [rbp - 48], 1 # 1-byte Folded Reload
jne .LBB0_23
# %bb.10: # %L115
# in Loop: Header=BB0_3 Depth=1
cmp dl, 10
jne .LBB0_11
# %bb.14: # %L122
# in Loop: Header=BB0_3 Depth=1
test r15b, 1
jne .LBB0_15
# %bb.12: # %L130.thread
# in Loop: Header=BB0_3 Depth=1
movzx eax, byte ptr [rbp - 41] # 1-byte Folded Reload
mov bl, 1
add eax, eax
lea eax, [rax + 4*rax]
add al, r12b
movzx eax, al
add qword ptr [rbp - 56], rax # 8-byte Folded Spill
mov al, 1
mov dword ptr [rbp - 48], eax # 4-byte Spill
cmp qword ptr [rbp - 64], r14 # 8-byte Folded Reload
jne .LBB0_25
jmp .LBB0_13
.p2align 4, 0x90
.LBB0_23: # %L115.thread
# in Loop: Header=BB0_3 Depth=1
mov al, 1
# implicit-def: $r8b
mov dword ptr [rbp - 48], eax # 4-byte Spill
cmp dl, 10
jne .LBB0_24
jmp .LBB0_21
.LBB0_11: # in Loop: Header=BB0_3 Depth=1
mov r8d, r12d
jmp .LBB0_24
.LBB0_1:
xor eax, eax
mov qword ptr [rbp - 56], rax # 8-byte Spill
.LBB0_13: # %L159
mov rax, qword ptr [rbp - 56] # 8-byte Reload
add rsp, 40
pop rbx
pop r12
pop r13
pop r14
pop r15
pop rbp
ret
.LBB0_21: # %L122.thread
test r15b, 1
jne .LBB0_15
# %bb.22: # %post_box_union58
movabs rdi, offset .L_j_str1
movabs rax, offset ijl_type_error
movabs rsi, 140008511215408
movabs rdx, 140008667209736
call rax
.LBB0_15: # %fail
cmp r11b, 1
je .LBB0_19
# %bb.16: # %fail
movzx eax, r11b
cmp eax, 2
jne .LBB0_17
# %bb.20: # %box_union54
movzx eax, byte ptr [rbp - 41] # 1-byte Folded Reload
movabs rcx, offset jl_boxed_uint8_cache
mov rdx, qword ptr [rcx + 8*rax]
jmp .LBB0_18
.LBB0_26: # %L80
movabs rax, offset ijl_throw
movabs rdi, 140008495049392
call rax
.LBB0_19: # %box_union
movabs rdx, 140008667209736
jmp .LBB0_18
.LBB0_17:
xor edx, edx
.LBB0_18: # %post_box_union
movabs rdi, offset .L_j_str1
movabs rax, offset ijl_type_error
movabs rsi, 140008511215408
call rax
.Lfunc_end0:
.size julia_part1_418, .Lfunc_end0-julia_part1_418
# -- End function
.type .L_j_str1,@object # @_j_str1
.section .rodata.str1.1,"aMS",@progbits,1
.L_j_str1:
.asciz "typeassert"
.size .L_j_str1, 11
.section ".note.GNU-stack","",@progbits
```
</details>
<details>
<summary>`@code_warntype` After</summary>
```julia
[sukera@tower 01]$ julia -q --project=. -L 01.jl
julia> data = read("input.txt");
julia> @code_warntype part1(data)
MethodInstance for part1(::Vector{UInt8})
from part1(data) @ Main ~/Documents/projects/AOC/2023/01/01.jl:7
Arguments
#self#::Core.Const(part1)
data::Vector{UInt8}
Locals
@_3::Union{Nothing, Tuple{UInt8, Int64}}
may_b::Union{Nothing, UInt8}
may_a::Union{Nothing, UInt8}
total::Int64
val@_7::Union{}
val@_8::Union{}
c::UInt8
digit_b::UInt8
digit_a::UInt8
##215::Some{Nothing}
##216::Union{Nothing, UInt8}
##217::Union{Nothing, Some{UInt8}}
##212::Some{Nothing}
##213::Union{Nothing, Some{UInt8}}
##214::Union{Nothing, UInt8}
digitRes::Union{Nothing, Some{UInt8}}
@_19::Union{Nothing, UInt8}
@_20::Union{Nothing, UInt8}
@_21::Nothing
@_22::Union{Nothing, UInt8}
@_23::Union{Nothing, UInt8}
@_24::Nothing
Body::Int64
1 ── (total = 0)
│ (may_a = Main.nothing)
│ (may_b = Main.nothing)
│ %4 = data::Vector{UInt8}
│ (@_3 = Base.iterate(%4))
│ %6 = @_3::Union{Nothing, Tuple{UInt8, Int64}}
│ %7 = (%6 === nothing)::Bool
│ %8 = Base.not_int(%7)::Bool
└─── goto #24 if not %8
2 ┄─ Core.NewvarNode(:(val@_7))
│ Core.NewvarNode(:(val@_8))
│ Core.NewvarNode(:(digit_b))
│ Core.NewvarNode(:(digit_a))
│ Core.NewvarNode(:(##215))
│ Core.NewvarNode(:(##216))
│ Core.NewvarNode(:(##217))
│ Core.NewvarNode(:(##212))
│ Core.NewvarNode(:(##213))
│ %19 = @_3::Tuple{UInt8, Int64}
│ (c = Core.getfield(%19, 1))
│ %21 = Core.getfield(%19, 2)::Int64
│ %22 = c::UInt8
│ (digitRes = Main.someDigit(%22))
│ %24 = may_a::Union{Nothing, UInt8}
│ (##214 = %24)
│ %26 = Base.:!::Core.Const(!)
│ %27 = ##214::Union{Nothing, UInt8}
│ %28 = Base.isnothing(%27)::Bool
│ %29 = (%26)(%28)::Bool
└─── goto #4 if not %29
3 ── %31 = ##214::UInt8
│ (@_19 = Base.something(%31))
└─── goto #11
4 ── %34 = digitRes::Union{Nothing, Some{UInt8}}
│ (##213 = %34)
│ %36 = Base.:!::Core.Const(!)
│ %37 = ##213::Union{Nothing, Some{UInt8}}
│ %38 = Base.isnothing(%37)::Bool
│ %39 = (%36)(%38)::Bool
└─── goto #6 if not %39
5 ── %41 = ##213::Some{UInt8}
│ (@_20 = Base.something(%41))
└─── goto #10
6 ── %44 = Main.Some::Core.Const(Some)
│ %45 = Main.nothing::Core.Const(nothing)
│ (##212 = (%44)(%45))
│ %47 = Base.:!::Core.Const(!)
│ %48 = ##212::Core.Const(Some(nothing))
│ %49 = Base.isnothing(%48)::Core.Const(false)
│ %50 = (%47)(%49)::Core.Const(true)
└─── goto #8 if not %50
7 ── %52 = ##212::Core.Const(Some(nothing))
│ (@_21 = Base.something(%52))
└─── goto #9
8 ── Core.Const(nothing)
│ Core.Const(:(val@_8 = Base.something(Base.nothing)))
│ Core.Const(nothing)
│ Core.Const(:(val@_8))
└─── Core.Const(:(@_21 = %58))
9 ┄─ %60 = @_21::Core.Const(nothing)
└─── (@_20 = %60)
10 ┄ %62 = @_20::Union{Nothing, UInt8}
└─── (@_19 = %62)
11 ┄ %64 = @_19::Union{Nothing, UInt8}
│ (may_a = %64)
│ %66 = digitRes::Union{Nothing, Some{UInt8}}
│ (##217 = %66)
│ %68 = Base.:!::Core.Const(!)
│ %69 = ##217::Union{Nothing, Some{UInt8}}
│ %70 = Base.isnothing(%69)::Bool
│ %71 = (%68)(%70)::Bool
└─── goto #13 if not %71
12 ─ %73 = ##217::Some{UInt8}
│ (@_22 = Base.something(%73))
└─── goto #20
13 ─ %76 = may_b::Union{Nothing, UInt8}
│ (##216 = %76)
│ %78 = Base.:!::Core.Const(!)
│ %79 = ##216::Union{Nothing, UInt8}
│ %80 = Base.isnothing(%79)::Bool
│ %81 = (%78)(%80)::Bool
└─── goto #15 if not %81
14 ─ %83 = ##216::UInt8
│ (@_23 = Base.something(%83))
└─── goto #19
15 ─ %86 = Main.Some::Core.Const(Some)
│ %87 = Main.nothing::Core.Const(nothing)
│ (##215 = (%86)(%87))
│ %89 = Base.:!::Core.Const(!)
│ %90 = ##215::Core.Const(Some(nothing))
│ %91 = Base.isnothing(%90)::Core.Const(false)
│ %92 = (%89)(%91)::Core.Const(true)
└─── goto #17 if not %92
16 ─ %94 = ##215::Core.Const(Some(nothing))
│ (@_24 = Base.something(%94))
└─── goto #18
17 ─ Core.Const(nothing)
│ Core.Const(:(val@_7 = Base.something(Base.nothing)))
│ Core.Const(nothing)
│ Core.Const(:(val@_7))
└─── Core.Const(:(@_24 = %100))
18 ┄ %102 = @_24::Core.Const(nothing)
└─── (@_23 = %102)
19 ┄ %104 = @_23::Union{Nothing, UInt8}
└─── (@_22 = %104)
20 ┄ %106 = @_22::Union{Nothing, UInt8}
│ (may_b = %106)
│ %108 = Main.:(==)::Core.Const(==)
│ %109 = c::UInt8
│ %110 = Main.UInt8('\n')::Core.Const(0x0a)
│ %111 = (%108)(%109, %110)::Bool
└─── goto #22 if not %111
21 ─ %113 = may_a::Union{Nothing, UInt8}
│ (digit_a = Core.typeassert(%113, Main.UInt8))
│ %115 = may_b::Union{Nothing, UInt8}
│ (digit_b = Core.typeassert(%115, Main.UInt8))
│ %117 = Main.:+::Core.Const(+)
│ %118 = total::Int64
│ %119 = Main.:+::Core.Const(+)
│ %120 = Main.:*::Core.Const(*)
│ %121 = digit_a::UInt8
│ %122 = (%120)(%121, 0x0a)::UInt8
│ %123 = digit_b::UInt8
│ %124 = (%119)(%122, %123)::UInt8
│ (total = (%117)(%118, %124))
│ (may_a = Main.nothing)
└─── (may_b = Main.nothing)
22 ┄ (@_3 = Base.iterate(%4, %21))
│ %129 = @_3::Union{Nothing, Tuple{UInt8, Int64}}
│ %130 = (%129 === nothing)::Bool
│ %131 = Base.not_int(%130)::Bool
└─── goto #24 if not %131
23 ─ goto #2
24 ┄ %134 = total::Int64
└─── return %134
```
</details>
<details>
<summary>`@code_native debuginfo=:none` After </summary>
```julia
julia> @code_native debuginfo=:none part1(data)
.text
.file "part1"
.globl julia_part1_1203 # -- Begin function julia_part1_1203
.p2align 4, 0x90
.type julia_part1_1203,@function
julia_part1_1203: # @julia_part1_1203
; Function Signature: part1(Array{UInt8, 1})
# %bb.0: # %top
#DEBUG_VALUE: part1:data <- [DW_OP_deref] $rdi
push rbp
mov rbp, rsp
push r15
push r14
push r13
push r12
push rbx
sub rsp, 40
vxorps xmm0, xmm0, xmm0
#APP
mov rax, qword ptr fs:[0]
#NO_APP
lea rdx, [rbp - 64]
vmovaps xmmword ptr [rbp - 64], xmm0
mov qword ptr [rbp - 48], 0
mov rcx, qword ptr [rax - 8]
mov qword ptr [rbp - 64], 4
mov rax, qword ptr [rcx]
mov qword ptr [rbp - 72], rcx # 8-byte Spill
mov qword ptr [rbp - 56], rax
mov qword ptr [rcx], rdx
#DEBUG_VALUE: part1:data <- [DW_OP_deref] 0
mov r15, qword ptr [rdi + 16]
test r15, r15
je .LBB0_1
# %bb.2: # %L34
mov r14, qword ptr [rdi]
dec r15
mov r11b, 1
mov r13b, 1
# implicit-def: $r12b
# implicit-def: $r10b
xor eax, eax
jmp .LBB0_3
.p2align 4, 0x90
.LBB0_4: # in Loop: Header=BB0_3 Depth=1
xor r11d, r11d
mov ebx, edi
mov r10d, r8d
.LBB0_9: # %L114
# in Loop: Header=BB0_3 Depth=1
mov r12d, esi
test r15, r15
je .LBB0_12
.LBB0_10: # %guard_exit126
# in Loop: Header=BB0_3 Depth=1
inc r14
dec r15
mov r13d, ebx
.LBB0_3: # %L36
# =>This Inner Loop Header: Depth=1
movzx edx, byte ptr [r14]
test r13b, 1
movzx edi, r13b
mov ebx, 1
mov ecx, 0
cmove ebx, edi
cmovne edi, ecx
movzx ecx, r10b
lea esi, [rdx - 48]
lea r9d, [rdx - 58]
movzx r8d, sil
cmove r8d, ecx
cmp r9b, -11
ja .LBB0_4
# %bb.5: # %L89
# in Loop: Header=BB0_3 Depth=1
test r11b, 1
jne .LBB0_8
# %bb.6: # %L102
# in Loop: Header=BB0_3 Depth=1
cmp dl, 10
jne .LBB0_7
# %bb.13: # %L106
# in Loop: Header=BB0_3 Depth=1
test r13b, 1
jne .LBB0_14
# %bb.11: # %L114.thread
# in Loop: Header=BB0_3 Depth=1
add ecx, ecx
mov bl, 1
mov r11b, 1
lea ecx, [rcx + 4*rcx]
add cl, r12b
movzx ecx, cl
add rax, rcx
test r15, r15
jne .LBB0_10
jmp .LBB0_12
.p2align 4, 0x90
.LBB0_8: # %L102.thread
# in Loop: Header=BB0_3 Depth=1
mov r11b, 1
# implicit-def: $sil
cmp dl, 10
jne .LBB0_9
jmp .LBB0_15
.LBB0_7: # in Loop: Header=BB0_3 Depth=1
mov esi, r12d
jmp .LBB0_9
.LBB0_1:
xor eax, eax
.LBB0_12: # %L154
mov rcx, qword ptr [rbp - 56]
mov rdx, qword ptr [rbp - 72] # 8-byte Reload
mov qword ptr [rdx], rcx
add rsp, 40
pop rbx
pop r12
pop r13
pop r14
pop r15
pop rbp
ret
.LBB0_15: # %L106.thread
test r13b, 1
jne .LBB0_14
# %bb.16: # %post_box_union47
movabs rax, offset jl_nothing
movabs rcx, offset jl_small_typeof
movabs rdi, offset ".L_j_str_typeassert#1"
mov rdx, qword ptr [rax]
mov rsi, qword ptr [rcx + 336]
movabs rax, offset ijl_type_error
mov qword ptr [rbp - 48], rsi
call rax
.LBB0_14: # %post_box_union
movabs rax, offset jl_nothing
movabs rcx, offset jl_small_typeof
movabs rdi, offset ".L_j_str_typeassert#1"
mov rdx, qword ptr [rax]
mov rsi, qword ptr [rcx + 336]
movabs rax, offset ijl_type_error
mov qword ptr [rbp - 48], rsi
call rax
.Lfunc_end0:
.size julia_part1_1203, .Lfunc_end0-julia_part1_1203
# -- End function
.type ".L_j_str_typeassert#1",@object # @"_j_str_typeassert#1"
.section .rodata.str1.1,"aMS",@progbits,1
".L_j_str_typeassert#1":
.asciz "typeassert"
.size ".L_j_str_typeassert#1", 11
.section ".note.GNU-stack","",@progbits
```
</details>
Co-authored-by: Sukera <[email protected]>
Merged
aviatesk
added a commit
that referenced
this issue
Oct 1, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 1, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
vtjnash
pushed a commit
that referenced
this issue
Oct 1, 2024
`memoryref(mem, i)` will otherwise emit a boundscheck.
```
; │ @ /home/vchuravy/WorkstealingQueues/src/CLL.jl:53 within `setindex_atomic!` @ genericmemory.jl:329
; │┌ @ boot.jl:545 within `memoryref`
%ptls_field = getelementptr inbounds i8, ptr %tls_pgcstack, i64 16
%ptls_load = load ptr, ptr %ptls_field, align 8
%"box::GenericMemoryRef" = call noalias nonnull align 8 dereferenceable(32) ptr @ijl_gc_small_alloc(ptr %ptls_load, i32 552, i32 32, i64 23456076646928) #9
%"box::GenericMemoryRef.tag_addr" = getelementptr inbounds i64, ptr %"box::GenericMemoryRef", i64 -1
store atomic i64 23456076646928, ptr %"box::GenericMemoryRef.tag_addr" unordered, align 8
store ptr %memoryref_data, ptr %"box::GenericMemoryRef", align 8
%.repack8 = getelementptr inbounds { ptr, ptr }, ptr %"box::GenericMemoryRef", i64 0, i32 1
store ptr %memoryref_mem, ptr %.repack8, align 8
call void @ijl_bounds_error_int(ptr nonnull %"box::GenericMemoryRef", i64 %7)
unreachable
```
For the Julia code:
```julia
function Base.setindex_atomic!(buf::WSBuffer{T}, order::Symbol, val::T, idx::Int64) where T
@inbounds Base.setindex_atomic!(buf.buffer, order, val,((idx - 1) & buf.mask) + 1)
end
```
from
https://github.com/gbaraldi/WorkstealingQueues.jl/blob/0ebc57237cf0c90feedf99e4338577d04b67805b/src/CLL.jl#L41
aviatesk
added a commit
that referenced
this issue
Oct 2, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 2, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 2, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 4, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 4, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 4, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 5, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 9, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 11, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 11, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 12, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 15, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 16, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked as
`broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical.
In particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible.
aviatesk
added a commit
that referenced
this issue
Oct 16, 2024
E.g. this allows `finalizer` inlining in the following case:
```julia
mutable struct ForeignBuffer{T}
const ptr::Ptr{T}
end
const foreign_buffer_finalized = Ref(false)
function foreign_alloc(::Type{T}, length) where T
ptr = Libc.malloc(sizeof(T) * length)
ptr = Base.unsafe_convert(Ptr{T}, ptr)
obj = ForeignBuffer{T}(ptr)
return finalizer(obj) do obj
Base.@assume_effects :notaskstate :nothrow
foreign_buffer_finalized[] = true
Libc.free(obj.ptr)
end
end
function f_EA_finalizer(N::Int)
workspace = foreign_alloc(Float64, N)
GC.@preserve workspace begin
(;ptr) = workspace
Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr)
end
end
```
```julia
julia> @code_typed f_EA_finalizer(42)
CodeInfo(
1 ── %1 = Base.mul_int(8, N)::Int64
│ %2 = Core.lshr_int(%1, 63)::Int64
│ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8
│ %4 = Core.eq_int(%3, 0x01)::Bool
└─── goto #3 if not %4
2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{}
└─── unreachable
3 ── goto #4
4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64
└─── goto #5
5 ── goto #6
6 ── goto #7
7 ── goto #8
8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing}
└─── goto #9
9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64}
│ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64}
└─── goto #10
10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17)))
│ %20 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing
│ $(Expr(:gc_preserve_end, :(%19)))
│ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool}
│ Base.setfield!(%23, :x, true)::Bool
│ %25 = Base.getfield(%17, :ptr)::Ptr{Float64}
│ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing}
│ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing
└─── return nothing
) => Nothing
```
However, this is still a WIP. Before merging, I want to improve EA's
precision a bit and at least fix the test case that is currently marked
as `broken`. I also need to check its impact on compiler performance.
Additionally, I believe this feature is not yet practical. In
particular, there is still significant room for improvement in the
following areas:
- EA's interprocedural capabilities: currently EA is performed ad-hoc
for limited frames because of latency reasons, which significantly
reduces its precision in the presence of interprocedural calls.
- Relaxing the `:nothrow` check for finalizer inlining: the current
algorithm requires `:nothrow`-ness on all paths from the allocation of
the mutable struct to its last use, which is not practical for
real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary
optimizations such as inserting a `finalize` call after the last use
might still be possible (#55990).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This non-default test has been failing for quite some time, I think.
The text was updated successfully, but these errors were encountered: