https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102441
--- Comment #7 from CVS Commits ---
The releases/gcc-10 branch has been updated by Jakub Jelinek
:
https://gcc.gnu.org/g:a34be854e83c79229a7ece3682df7167f745dbfd
commit r10-10648-ga34be854e83c79229a7ece3682df7167f745dbfd
Author: Jakub Jelinek
Date: Sun Oct 10 12:13:22 2021 +0200
var-tracking: Fix a wrong-debug issue caused by my r10-7665 var-tracking
change [PR102441]
Since my r10-7665-g33c45e51b4914008064d9b77f2c1fc0eea1ad060 change, we get
wrong-debug on e.g. the following testcase at -O2 -g on x86_64-linux for
the
x parameter:
void bar (int *r);
int
foo (int x)
{
int r = 0;
bar ();
return r;
}
At the start of function, we have
subq$24, %rsp
leaq12(%rsp), %rdi
instructions. The x parameter is passed in %rdi, but isn't used in the
function and so the leaq instruction overwrites %rdi without remembering
%rdi anywhere. Before the r10-7665 change (which was trying to fix a large
(3% for 32-bit, 1% for 64-bit x86-64) debug info/loc growth introduced with
r10-7515), the leaq insn above resulted in a MO_VAL_SET micro-operation
that
said that the value of sp + 12, a cselib_sp_derived_value_p, is stored into
the %rdi register. The r10-7665 change added a change to add_stores that
added no micro-operation for the leaq store, with the rationale that the sp
based values can be and will be always computable some other more compact
and primarily more stable way (cfa based expression like DW_OP_fbreg, that
is the same in the whole function). That is true. But by throwing the
micro-operation on the floor, we miss another important part of the
MO_VAL_SET, in particular that the destination of the store, %rdi in this
case, now has a different value from what it had before, so the vt_*
dataflow code thinks that even after the leaq instruction %rdi still holds
the x argument value (and changes it to DW_OP_entry_value (%rdi) only in
the
middle of the call to bar). Previously and with the patches below,
the location for x changes already at the end of leaq instruction to
DW_OP_entry_value (%rdi).
My first attempt to fix this was instead of dropping the MO_VAL_SET add
a MO_CLOBBER operation:
--- gcc/var-tracking.c.jj 2021-05-04 21:02:24.196799586 +0200
+++ gcc/var-tracking.c 2021-09-24 19:23:16.420154828 +0200
@@ -6133,7 +6133,9 @@ add_stores (rtx loc, const_rtx expr, voi
{
if (preserve)
preserve_value (v);
- return;
+ mo.type = MO_CLOBBER;
+ mo.u.loc = loc;
+ goto log_and_return;
}
nloc = replace_expr_with_values (oloc);
so don't track that the value lives in the loc destination, but track
that the previous value doesn't live there anymore. That failed bootstrap
miserably, the vt_* code isn't prepared to see MO_CLOBBER of a MEM that
isn't tracked (e.g. has MEM_EXPR on it that the var-tracking code wants
to track, i.e. track_p in add_stores). On the other side, thinking about
it more, in the most common case where a cselib_sp_derived_value_p value
is stored into the sp register (and which is the reason why PR94495
testcase got larger), dropping the micro-operation on the floor is the
right thing, because we have that cselib_sp_derived_value_p tracking, any
reads from the sp hard register will be treated as
cselib_sp_derived_value_p.
Then I've tried 3 different patches described below and in the end
what is committed is patch2.
Additionally, I've gathered statistics from cc1plus by always reverting the
var-tracking.c change after finished bootstrap/regtest and rebuilding the
stage3 var-tracking.o and cc1plus, such that it would be comparable.
dwlocstat and .debug_{info,loclists} section sizes detailed below.
patch3 uses MO_VAL_SET (i.e. essentially reversion of the r10-7665
change) when destination is not a REG_P and !track_p, otherwise if
destination is sp drops the micro-operation on the floor (i.e. no change),
otherwise adds a MO_CLOBBER.
patch1 is similar, except it checks for destination not equal to sp and
!track_p, i.e. for !track_p REG_P destinations other than sp it will use
MO_VAL_SET rather than MO_CLOBBER.
Finally, patch2, the shortest patch, uses MO_VAL_SET whenever destination
is not sp and otherwise drops the micro-operation on the floor.
All the 3 patches don't affect the PR94495 testcase, all the changes
there were caused by stores of sp based values into %rsp.
While the patch2 (and patch1 which results in exactly the same sizes)
causes the largest debug loclists/info growth from the 3, it is still quite
minor (0.651% on 64-bit and 0.114% on 32-bit) compared
to the 1% and 3% PR94495 was trying to solve, and I actually think it is
the
best
