🚀 go-pugleaf

There is a bug in gcc 2.2.2 -fschedule-insns2 on Sparcstations which
prevents GNU diff 1.15 from compiling correctly with -O2.

Specifically, the function "diag" in file "analyze.c" ends up using a
corrupted index.  What happens is that when compiled with:

	gcc -O2 -fno-schedule-insns2 -c analyze.c

the code generates the instructions:

	! ... calculates initial value for "bmid"; result in %l6
	st %l6,[%fp-60]
	ld [%fp-60],%o0
	! uses %o0 as "bmid" value ...

which stores the "bmid" value in the stack frame, and then loads the
value from the stack frame into %o0.  This isn't great code (a "mov
%l6,%o0" would be an improvement on the "ld"), but it is correct.
However, when compiled with "-fschedule-insns2", the "st" instruction
is moved to below the "ld" instruction, like this:

	! ... calculates initial value for "bmid"; result in %l6
	ld [%fp-60],%o0		! *** Loads random garbage from stack ***
	! uses %o0 as "bmid" value...
	...
	st %l6,[%fp-60]		! Heres where the store instr ended up
	! still thinks "bmid" value is in %o0 ...

So a random value ends up being used for "bmid", resulting in core
dumps and generally bad behaviour.

The details on my configuration:

	SparcStation ELC (Sun 4/25)
	Sun OS 4.1.1
	gcc 2.2.2

I assume the GNU diff 1.15 sources are widely enough available that I
can skip enclosing them.  Sorry, but I don't have a simpler example
reproducing the bug.

More detailed (although still truncated) assembler output follows.  I
did add some comments to it.

--Lon Willett		willett@math.utah.edu

------------------------------------------------------------------------

! Lines starting with "<" are from: gcc -O2 -fno-schedule-insns2 -S analyze.c
! Lines starting with ">" are from: gcc -O2 -S analyze.c

gcc2_compiled.:
.text
	.align 4
	.proc	04
_diag:
	!#PROLOGUE# 0
	save %sp,-200,%sp
	!#PROLOGUE# 1
	mov %i0,%l3
	mov %i1,%o4
	mov %i2,%l4
	mov %i3,%l5
	st %i4,[%fp-20]
	sethi %hi(_xvec),%i0
	ld [%i0+%lo(_xvec)],%l6
<	st %l6,[%fp-28]
	sethi %hi(_yvec),%g3
	ld [%g3+%lo(_yvec)],%l7
	sub %l3,%l5,%o5
	st %o5,[%fp-36]
<	sub %o4,%l4,%l6
<	st %l6,[%fp-44]
	sub %l3,%l4,%o5
	st %o5,[%fp-52]
	ld [%fp-52],%o1
>	mov 1,%l0
>	st %l0,[%fp-100]
>	mov 2,%o5
>	st %o5,[%fp-84]
>	st %l6,[%fp-28]
>	sub %o4,%l4,%l6
>	st %l6,[%fp-44]
<	mov %o1,%o3
	sub %o4,%l5,%l6
<	st %l6,[%fp-60]		! Store "bmid" value (-fno-schedule-insns2)
	ld [%fp-60],%o0		! *** Loads bad %o0 ("bmid")
				! Replace with "mov %l6,%o0" and the
				! problem goes away.
<	mov %o0,%o2
>	mov %o1,%o3
	sub %o1,%o0,%g2		! *** First use of bad %o0 ("bmid")
	and %g2,1,%g2
	st %g2,[%fp-68]
<	mov 1,%l0
<	st %l0,[%fp-100]
<	mov 2,%o5
<	st %o5,[%fp-84]
	sethi %hi(_fdiag),%g2
	ld [%g2+%lo(_fdiag)],%l2
>	st %l6,[%fp-60]		! *** Finally stores "bmid" value, but
				! register %o0 (register "bmid") is still
				! corrupted, and will be used below.
	sll %o1,2,%g2
	st %l3,[%l2+%g2]
	sethi %hi(_bdiag),%g2
	ld [%g2+%lo(_bdiag)],%l1
>	mov %o0,%o2
	sll %o0,2,%g2
	st %o4,[%l1+%g2]
L2:
	...
------------------------------------------------------------------------