depends: Fully determine path for darwin_{CC,CXX}

Instead of doing the awkward /bin path prepending at config.site
creation time, set darwin_{CC,CXX} in a way that fully determines the
program's path (clang/clang++) similar to how AC_PATH_{TOOL,PROG} would
do.

Also see the added comment block in depends/Makefile for more context on
determining $PATH for our config.site.

1 files changed, 27 insertions, 7 deletions

diff --git a/depends/Makefile b/depends/Makefile
index 0de98f8e3c..e0265973cb 100644
--- a/depends/Makefile
+++ b/depends/Makefile
@@ -167,11 +167,7 @@ $(host_arch)_$(host_os)_native_toolchain?=$($(host_os)_native_toolchain)
 include funcs.mk
 
 binutils_path=$($($(host_arch)_$(host_os)_native_binutils)_prefixbin)
-ifeq ($(strip $(FORCE_USE_SYSTEM_CLANG)),)
-toolchain_path=$($($(host_arch)_$(host_os)_native_toolchain)_prefixbin)
-else
-toolchain_path=
-endif
+
 final_build_id_long+=$(shell $(build_SHA256SUM) config.site.in)
 final_build_id+=$(shell echo -n "$(final_build_id_long)" | $(build_SHA256SUM) | cut -c-$(HASH_LENGTH))
 $(host_prefix)/.stamp_$(final_build_id): $(native_packages) $(packages)
@@ -182,11 +178,35 @@ $(host_prefix)/.stamp_$(final_build_id): $(native_packages) $(packages)
 	$(AT)cd $(@D); $(foreach package,$^, tar xf $($(package)_cached); )
 	$(AT)touch $@
 
+# $PATH is not preserved between ./configure and make by convention. Its
+# modification and overriding at ./configure time is (as I understand it)
+# supposed to be captured by the AC_{PROG_{,OBJ}CXX,PATH_{PROG,TOOL}} macros,
+# which will expand the program names to their full absolute paths. The notable
+# exception is command line overriding: ./configure CC=clang, which skips the
+# program name expansion step, and works because the user implicitly indicates
+# with CC=clang that clang will be available in $PATH at all times, and is most
+# likely part of the user's system.
+#
+# Therefore, when we "seed the autoconf cache"/"override well-known program
+# vars" by setting AR=<blah> in our config.site, either one of two things needs
+# to be true for the build system to work correctly:
+#
+#   1. If we refer to the program by name (e.g. AR=riscv64-gnu-linux-ar), the
+#      tool needs to be available in $PATH at all times.
+#
+#   2. If the tool is _**not**_ expected to be available in $PATH at all times
+#      (such as is the case for our native_cctools binutils tools), it needs to
+#      be referred to by its absolute path, such as would be output by the
+#      AC_PATH_{PROG,TOOL} macros.
+#
+# Minor note: it is also okay to refer to tools by their absolute path even if
+# we expect them to be available in $PATH at all times, more specificity does
+# not hurt.
 $(host_prefix)/share/config.site : config.site.in $(host_prefix)/.stamp_$(final_build_id)
 	$(AT)@mkdir -p $(@D)
 	$(AT)sed -e 's|@HOST@|$(host)|' \
-            -e 's|@CC@|$(toolchain_path)$(host_CC)|' \
-            -e 's|@CXX@|$(toolchain_path)$(host_CXX)|' \
+            -e 's|@CC@|$(host_CC)|' \
+            -e 's|@CXX@|$(host_CXX)|' \
             -e 's|@AR@|$(binutils_path)$(host_AR)|' \
             -e 's|@RANLIB@|$(binutils_path)$(host_RANLIB)|' \
             -e 's|@NM@|$(binutils_path)$(host_NM)|' \