Add build_target global to opt-in to generic CPU at precompile-time

We already support this dynamically, but for some reason it wasn't
possible to get this configuration at precompilation time. Also adds
a "build_target" preference for (optional) override.

Author: topolarity

src/HostCPUFeatures.jl

@@ -14,6 +14,17 @@ using BitTwiddlingConvenienceFunctions: prevpow2, nextpow2, intlog2
 export has_feature, fma_fast, pick_vector_width, pick_vector_width_shift, register_count,
   register_size, simd_integer_register_size
 
+# if not 'nothing', this build target will opt-in to "freeze" an under-
+# approximation of the CPU features at precompile-time. Currently we only
+# do this if "native" was excluded from the CPU target
+const build_target = let _cpu_target = Base.unsafe_string(Base.JLOptions().cpu_target)
+  if @has_preference("build_target")
+    @load_preference("build_target", nothing)
+  else
+    occursin("native", _cpu_target) ? nothing : split(_cpu_target, ";")[1]
+  end
+end
+
 function get_cpu_name()::String
   if isdefined(Sys, :CPU_NAME)
     Sys.CPU_NAME
@@ -46,14 +57,26 @@ end
 const BASELINE_CPU_NAME = get_cpu_name()
 const allow_eval = @load_preference("allow_runtime_invalidation", false)
 
+function make_generic(target)
+  if Sys.ARCH === :x86_64 || Sys.ARCH === :i686
+    make_generic_x86(target)
+    return true
+  else
+    return false
+  end
+end
+
+if build_target !== nothing
+  make_generic(build_target)
+end
+
 function __init__()
   ccall(:jl_generating_output, Cint, ()) == 1 && return
-  if Sys.ARCH === :x86_64 || Sys.ARCH === :i686
-    target = Base.unsafe_string(Base.JLOptions().cpu_target)
-    if !occursin("native",  target)
-      make_generic(target)
-      return nothing
-    end
+  build_target !== nothing && return # CPU info fixed at precompile-time
+
+  runtime_target = Base.unsafe_string(Base.JLOptions().cpu_target)
+  if !occursin("native", runtime_target) && make_generic(runtime_target)
+    return nothing
   end
   if BASELINE_CPU_NAME != Sys.CPU_NAME::String
     redefine()

src/cpu_info.jl

@@ -39,9 +39,10 @@ function set_features!()
   end
   Libc.free(features_cstring)
 end
-set_features!()
-
 
+if build_target === nothing
+    set_features!()
+end
 
 function reset_features!()
     features, features_cstring = feature_string()
@@ -70,5 +71,7 @@ function redefine_cpu_name()
       @warn "Runtime invalidation was disabled, but the CPU info is out-of-date.\nWill continue with incorrect CPU name (from build time)."
     end
 end
-cpu = QuoteNode(Symbol(get_cpu_name()))
-@eval cpu_name() = Val{$cpu}()
+
+let _cpu = QuoteNode(Symbol(build_target !== nothing ? build_target : get_cpu_name()))
+   @eval cpu_name() = Val{$_cpu}()
+end

src/cpu_info_aarch64.jl

@@ -28,7 +28,7 @@ function _set_sve_vector_width!(bytes = _dynamic_register_size())
 end
 
 
-if _has_aarch64_sve()# && !(Bool(has_feature(Val(:aarch64_sve))))
+if build_target === nothing && _has_aarch64_sve()# && !(Bool(has_feature(Val(:aarch64_sve))))
     has_feature(::Val{:aarch64_sve_cpuid}) = True()
     _set_sve_vector_width!()
 else

src/cpu_info_x86.jl

@@ -51,7 +51,7 @@ end
   end
 end
 
-function make_generic(target)
+function make_generic_x86(target)
   if occursin("tigerlake", target) || occursin("znver4", target) || occursin("sapphirerapids", target)
     # most feature-complete architectures we use
     setfeaturetrue(:x86_64_avx512ifma)