[sorinirimies]

Summary

SpirvBuilder::invoke_rustc was ~270 lines mixing RUSTGPU_* env reads, flag building and running cargo, so the flag logic was only covered indirectly by difftests. Pure refactor, no behavior change.

 invoke_rustc
   validate crate path / target, find backend
-  build rustflags + llvm_args + target_features inline (~130 lines)
-  tracked_env_var_get(..) scattered through it
+  env_overrides = EnvOverrides { codegen_args, rustflags, cargoflags }  // read once, still rerun-if-env-changed
+  rustflags = self.rustflags(backend, &env_overrides)
+    ├── self.llvm_args(&env_overrides)   // -Cllvm-args
+    └── self.target_features()           // -Ctarget-feature
   resolve target dir, run cargo, parse output

Evidence

Same rustc invocation before/after — sky-shader via multibuilder with RUSTGPU_CARGOFLAGS=-v, diffing the Running \rustc --crate-name sky_shader ...`` line:

before: 89 args
after:  89 args
diff:   IDENTICAL

New tests::rustflags:

fixed_flags_start_with_the_backend                                    ok
silent_exit_has_no_llvm_args_or_target_features                       ok
builder_options_map_to_flags                                          ok
env_codegen_args_replace_extra_args_and_env_rustflags_are_appended    ok

cargo test -p spirv-builder --features watch,clap: 8 passed. Clippy -D warnings clean (watch,clap and --no-default-features).

Merge Danger

Door: two-way

Blast Radius: small — internal methods only; the only observable difference is that the three cargo:rerun-if-env-changed lines are now printed together before cargo runs.