libexpr/meson: Rice the compiler inlining heuristics to improve perf of the bison generated parser

Turns out both GCC and Clang need a bit of hand-holding to optimize the bison generated
code well, otherwise parser performance tanks.

(Comparisons against baseline in 7e8db2e):

For GCC:

Benchmark 1 (15 runs): result/bin/nix-instantiate --parse ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           335ms ± 2.89ms     332ms …  342ms          0 ( 0%)        0%

Benchmark 2 (16 runs): result-old/bin/nix-instantiate --parse ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           330ms ± 2.87ms     326ms …  337ms          0 ( 0%)          -  1.4% ±  0.6%

For Clang:

Benchmark 1 (15 runs): result-clang/bin/nix-instantiate --parse ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           340ms ± 1.43ms     338ms …  343ms          0 ( 0%)        0%

Benchmark 2 (15 runs): result-old-clang/bin/nix-instantiate --parse ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           334ms ± 1.61ms     332ms …  338ms          0 ( 0%)        ⚡-  1.7% ±  0.3%

Author: xokdvium

src/libexpr/meson.build

@@ -183,17 +183,62 @@ subdir('primops')
 subdir('nix-meson-build-support/export-all-symbols')
 subdir('nix-meson-build-support/windows-version')
 
+# Turns out that Bison/Flex are particularly sensitive to compilers
+# failing to inline functions. For that reason we crank up the inlining
+# threshold manually for optimized builds. Yes, this can be considered 'ricing'
+# the compiler, but it does pay off.
+#
+# NOTE: missed inlining can be spotted (for Clang) using -Rpass-missed=inline
+# and -fdump-ipa-inline-missed (for GCC).
+parser_library_cpp_args = []
+
+if not get_option('debug')
+  if cxx.get_id() == 'clang'
+    # The default as of LLVM 21 is 225:
+    # llc --help-hidden | grep inline-threshold
+    parser_library_cpp_args += [
+      '-mllvm',
+      '-inline-threshold=5000',
+    ]
+  elif cxx.get_id() == 'gcc'
+    parser_library_cpp_args += [
+      '--param=max-inline-insns-single=1000',
+      '--param=max-inline-insns-auto=1000',
+      '--param=inline-unit-growth=400',
+    ]
+  endif
+endif
+
+# Working around https://github.com/mesonbuild/meson/issues/1367.
+parser_library = static_library(
+  'nixexpr-parser',
+  parser_tab,
+  lexer_tab,
+  cpp_args : parser_library_cpp_args,
+  dependencies : deps_public + deps_private + deps_other,
+  include_directories : include_dirs,
+  # 1. Stdlib and regular assertions regress parser performance significantly, so build without
+  # them for this one library when building in a release configuration.
+  # 2. Disable LTO for GCC because then inlining flags won't apply, since LTO in GCC is done
+  # by plonking down GIMPLE in the archive.
+  override_options : [
+    'b_ndebug=@0@'.format(not get_option('debug')),
+    'b_lto=@0@'.format(cxx.get_id() != 'gcc'),
+  ],
+)
+
 this_library = library(
   'nixexpr',
   sources,
   config_priv_h,
-  parser_tab,
-  lexer_tab,
+  parser_tab[1],
+  lexer_tab[1],
   generated_headers,
   soversion : nix_soversion,
   dependencies : deps_public + deps_private + deps_other,
   include_directories : include_dirs,
   link_args : linker_export_flags,
+  link_whole : [ parser_library ],
   prelink : true, # For C++ static initializers
   install : true,
   cpp_pch : do_pch ? [ 'pch/precompiled-headers.hh' ] : [],