Inline StepRange construction (#49270)

This can improve performance when iterating over a range with a step as currently steprange_last is not inlined. About a 12ns improvement

I don't think this should slow anything down as at worst you're making a call which is what you'd have to do without inlining. Also, it should be unlikely that branch is taken.
This should allow for better automatic inlining.

base/multidimensional.jl

@@ -526,7 +526,7 @@ module IteratorsMD
  dimrev = ntuple(i -> sum(==(i), dims; init = 0) == 1, Val(length(indices)))
  length(dims) == sum(dimrev) || throw(ArgumentError(Base.LazyString("invalid dimensions ", dims, " in reverse")))
  length(dims) == length(indices) && return Base._reverse(iter, :)
- indices′ = map((i, f) -> f ? reverse(i) : i, indices, dimrev)
+ indices′ = map((i, f) -> f ? (@noinline reverse(i)) : i, indices, dimrev)
  return CartesianIndices(indices′)
  end
 

base/range.jl

@@ -350,7 +350,8 @@ function steprange_last(start, step, stop)::typeof(stop)
  # Compute remainder as a nonnegative number:
  if absdiff isa Signed && absdiff < zero(absdiff)
  # unlikely, but handle the signed overflow case with unsigned rem
- remain = convert(typeof(absdiff), unsigned(absdiff) % absstep)
+ overflow_case(absdiff, absstep) = (@noinline; convert(typeof(absdiff), unsigned(absdiff) % absstep))
+ remain = overflow_case(absdiff, absstep)
  else
  remain = convert(typeof(absdiff), absdiff % absstep)
  end

test/ranges.jl

@@ -1,6 +1,7 @@
 # This file is a part of Julia. License is MIT: https://julialang.org/license
 
 using Base.Checked: checked_length
+using InteractiveUtils: code_llvm
 
 @testset "range construction" begin
  @test_throws ArgumentError range(start=1, step=1, stop=2, length=10)
@@ -2400,3 +2401,27 @@ end
  @test test_firstindex(StepRange{Union{Int64,Int128},Int}(Int64(1), 1, Int128(1)))
  @test test_firstindex(StepRange{Union{Int64,Int128},Int}(Int64(1), 1, Int128(0)))
 end
+
+@testset "Inline StepRange Construction #49270" begin
+ x = rand(Float32, 80)
+ a = rand(round(Int, length(x) / 2):length(x), 10^6)
+
+ function test(x, a)
+ c = zero(Float32)
+
+ @inbounds for j in a
+ for i in 1:8:j
+ c += x[i]
+ end
+ end
+
+ return c
+ end
+
+ llvm_ir(f, args) = sprint((io, args...) -> code_llvm(io, args...; debuginfo=:none), f, Base.typesof(args...))
+
+ ir = llvm_ir(test, (x, a))
+ @test !occursin("steprange_last", ir)
+ @test !occursin("_colon", ir)
+ @test !occursin("StepRange", ir)
+end