[j-modernc-org]

A struct passed to a copied reference parameter (AST_COPYREFTYPE) is
copied into memory AstFuncTempMemory takes on the stack, an AST_ALLOCA at
the top of the function's body. AstFuncTempMemory does not set the
function's uses_alloca, which NeedFramePointer reads to give the function
a frame and NeverInline reads to keep it from being inlined; only an
alloca the front ends see in the source sets it. So a function that needs
no frame of its own, one that only passes a struct on, takes the copy with
add ptra, #N (add sp, #N on P1) and returns without giving it back,
and so does a copy of it inlined into a caller. Every call leaves the
stack higher by the size of the struct, and a loop of such calls runs the
stack past the end of memory, with no error.

A struct goes this way when it is larger than 16 bytes or has a member
that is not a long, so this is common: a "getter" like

typedef struct { long long sec; int nsec; void *loc; } Time;
int Minute(Time t) { return (int)(timeAbs(t) % 3600) / 60; }

leaks 16 bytes a call. It showed as a scheduler on a P2-EDGE that stopped
printing, with no error, after about 10800 passes of a search loop.

This sets uses_alloca in AstFuncTempMemory. Test/Expect/stest184.pasm, a
struct of five ints passed on, shows it: fetchv moved sp up by 20 and never
back, and now saves and restores a frame. Test/cexec10.c calls two such
functions 100 times each and prints how far the stack moved: on master
1200 and 2000 bytes on a P2, and on a P1 it hangs; with this change 0.
The bytecode backends pass such a struct their own way and got the sums
in cexec10 wrong before and after this change, so the test prints the
expected text for them, as cexec09 does.

Testing:

• make test_offline: 588 of 588, with stest184's golden updated; that is
  the only golden the change touches.
• cexec10 on a P2-EDGE built with -2 -O2 and with -2nu, and on P1 under
  spinsim as ASM and as bytecode (--interp=rom): all print the expected
  text with this change. On master the -2 -O2 build prints 1200 and 2000,
  the P1 ASM build hangs, and -2nu prints 0 already.
• Builds with OPT="-O2 -Werror".
• 1546 C programs (the test corpus and fuzzer output of a compiler that
  emits C for flexcc, and its reproducers), compiled with -2 at b5993a8
  with our other patches applied, with and without this change: 5 build
  differently, and on a P2-EDGE each prints what it printed before. A scan
  of their listings for a function that moves ptra up without a frame
  finds 15 before and none after.