Oh, I see.  It's not that the targets are rebuilt in the wrong order, it's that the expansion of the secondary prerequisites isn't happening in the order in which targets are rebuilt.

That's probably OK.  I don't think we need to guarantee that the order in which secondary expansion happens across targets is any particular way.

In theory it should be the same way they are listed in the makefile since that's how non-secondary expansion prerequisites are expanded, but I don't think we need to guarantee that.

Certain tests assume that make builds prerequisites in a specific order.

E.g. test features/se_explicit.diff.13 assumes that makes builds prerequisite 2.x before 1.x.

features/se_explicit .................................... FAILED (14/18 passed)

4 Tests in 1 Category Failed (See .diff* files in work dir for details) :-(

$ cat work/features/se_explicit.diff
se_explicit.diff.13  se_explicit.diff.14  se_explicit.diff.15  se_explicit.diff.17
$ cat work/features/se_explicit.diff.13
*** work/features/se_explicit.base.13   Sat Mar 12 16:37:19 2022
--- work/features/se_explicit.log.13    Sat Mar 12 16:37:19 2022
***************
*** 1,3 ****
- @=2.x,<=5.z,^=5.z 6.z,+=5.z 6.z 5.z,|=7.z 8.z,?=,*=,%=
  @=1.x,<=1.z,^=1.z 2.z,+=1.z 2.z 2.z,|=3.z 4.z,?=,*=,%=
  make: Nothing to be done for 'all'.
--- 1,3 ----
  @=1.x,<=1.z,^=1.z 2.z,+=1.z 2.z 2.z,|=3.z 4.z,?=,*=,%=
+ @=2.x,<=5.z,^=5.z 6.z,+=5.z 6.z 5.z,|=7.z 8.z,?=,*=,%=
  make: Nothing to be done for 'all'.
$
$ cat work/features/se_explicit.mk.12

.SECONDEXPANSION:
all: 2.x 1.x
2.x: 5.z 6.z 5.z | 7.z 7.z 8.z
1.x: 1.z 2.z 2.z | 3.z 4.z
2.x 1.x: 9.z $$(info @=$$@,<=$$<,^=$$^,+=$$+,|=$$|,?=$$?,*=$$*,%=$$%) ;
%.z: ;
$

features/se_implicit .................................... FAILED (24/25 passed)

1 Test in 1 Category Failed (See .diff* files in work dir for details) :-(

features/se_implicit .................................... FAILED (24/25 passed)

1 Test in 1 Category Failed (See .diff* files in work dir for details) :-(

$ cat work/features/se_implicit.diff.21
*** work/features/se_implicit.base.21   Sat Mar 12 16:41:59 2022
--- work/features/se_implicit.log.21    Sat Mar 12 16:41:59 2022
***************
*** 1,3 ****
- @=2.x,<=5.z,^=5.z 6.z,+=5.z 6.z 5.z,|=7.z 8.z,?=,*=2
  @=1.x,<=1.z,^=1.z 2.z,+=1.z 2.z 2.z,|=3.z 4.z,?=,*=1
  make: Nothing to be done for 'all'.
--- 1,3 ----
  @=1.x,<=1.z,^=1.z 2.z,+=1.z 2.z 2.z,|=3.z 4.z,?=,*=1
+ @=2.x,<=5.z,^=5.z 6.z,+=5.z 6.z 5.z,|=7.z 8.z,?=,*=2
  make: Nothing to be done for 'all'.
$

This happens, because when second expansion is enabled, snap_deps obtains the list of prerequisites from hash_dump.
Those filenames hash to different buckets on different architectures and this causes make to process prerequisites in the order determined by hashing.
i observed these failures on big endian sparc-sun and powerpc-aix


This patch uses prerequisite names intentionally chosen to hash in the same order on both little and big endian.
Tested on big endian sparc-sun and powerpc-aix and little endian intel-linux, all in 64 and 32 bit modes.

Some other tests assume that make removes intermediate files in a specific order.

features/patternrules ................................... FAILED (26/31 passed)

5 Tests in 1 Category Failed (See .diff* files in work dir for details) :-(

$ cat work/features/patternrules.diff.21
*** work/features/patternrules.base.21  Sat Mar 12 16:36:25 2022
--- work/features/patternrules.log.21   Sat Mar 12 16:36:25 2022
***************
*** 1,3 ****
  touch a.1 a.2
  cat a.1 a.2 >a.4
! rm a.1 a.2
--- 1,3 ----
  touch a.1 a.2
  cat a.1 a.2 >a.4
! rm a.2 a.1