This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Patch: [tree-ssa] fix REG_DEAD notes

On Darwin with -mpowerpc64, the call in the following example

extern unsigned long long lseek (int, unsigned long long, int);
unsigned long long x;
int foo(int f, int z) {
   return lseek(f,x,z);

both uses and resets R3. gcc is not consistent about whether a REG_DEAD
note is applied in this case; life_analysis does not apply one, but combine
does, under some conditions. This leads to an ICE in the scheduler,
which deletes all the REG_DEAD notes coming from combine, then
recomputes them later using life analysis, and expects the totals to
match. This patch changes combine to be consistent with life_analysis.
Bootstrapped and tested on Darwin.

(The ICE occurs only with tree-ssa, because on mainline the merge of
the two-word return value into one register goes through memory, and
nobody is smart enough to use a register instead, and the bug is not
tickled. The tree-ssa optimizations cause the merge to be done in a
register, and combine eventually realizes all it needs to do is return R4
and R3 is dead. A great improvement.)

There is already a testcase, gcc.dg/macho-lo-sum.c, which demonstrates
the bug; it ICEd before and passes with my patch.  I won't be adding
another testcase.

2004-02-24 Dale Johannesen <>

        * combine.c (distribute_notes): Do not place a REG_DEAD note
        when value is both set and used.

Attachment: combinepatch.txt
Description: Text document

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]