kbuild: stop adding arch/x86/pci and power unconditionally

[ulfjack]

x86's ARCH_DRIVERS was hardcoded to "arch/x86/pci/ arch/x86/power/" in both architectures.bzl and linux_image_repository.bzl. Upstream those directories are config-gated (arch/x86/Makefile:285-291):

  drivers-$(CONFIG_MATH_EMULATION) += arch/x86/math-emu/
  drivers-$(CONFIG_PCI)            += arch/x86/pci/
  drivers-$(CONFIG_PM)             += arch/x86/power/

and Makefile:1220 exports the already-resolved result:

  export ARCH_DRIVERS := $(drivers-y) $(drivers-m)

which Kbuild:110 then appends with obj-y += $(ARCH_DRIVERS). Naming the directories directly hands the model drivers-y's value for one particular config, as a constant.

The gates are not lost, though -- the Kbuild tree parser reads arch/x86/Makefile as a root makefile and already descends into drivers-$(CONFIG_PCI) with the condition attached. So ARCH_DRIVERS was not supplying missing information, it was adding a second, ungated copy of information the parser already had. Two consequences:

1. With CONFIG_PCI=n the arch/x86/pci objects were emitted anyway, and the build failed rather than dropping them:

   arch/x86/pci/bus_numa.c:52:3: error: call to undeclared function
   'pci_add_resource'
   

   (pci.h only declares it under CONFIG_PCI.) Same shape for CONFIG_PM=n and arch/x86/power.

2. Even with CONFIG_PCI=y the duplicate perturbed the result. Link order became power before pci, where upstream's drivers-y order is pci then power, and arch/x86/power/cpu.o got its per-object flags twice (flags: [-fno-stack-protector, -fno-stack-protector], remove_flags: [$(CC_FLAGS_LTO), $(CC_FLAGS_LTO)]), which changes its content ID and so its cache identity.

Emptying ARCH_DRIVERS, as the arm64 descriptor already does, leaves the gated descent as the single source of truth.

Measured with the standalone generator:

• CONFIG_PCI=y baseline: 1195 object_variants before and after, same set of objects; the arch/x86/pci block moves ahead of arch/x86/power, and arch/x86/power/cpu.o loses its duplicated flags.
• CONFIG_PCI=n: 1118 object_variants with 9 stale arch/x86/pci objects before, 1109 with none after.
• A workspace with CONFIG_PCI unset failed to build before and now builds @min_asn1//:vmlinux.

Worth noting how this was found: it was not on the list from the config work, it surfaced only when a full vmlinux was built from a config that happened to leave CONFIG_PCI unset. Like the ASN.1 defect, the model looked fine and the compiler disagreed.

[ulfjack]

I think the test failure is a false positive. It's better to use aquery to count the actions, see #82.