Most floating point tests fail when GDB is built with GCC trunk for mips-linux. I assume it shows up on all targets. The miscompiled function is doublest.c:put_field. Here is a reduced test case:
int put_field (unsigned int start, unsigned int len) { int cur_bitshift = ((start + len) % 8) - 8; if (cur_bitshift > -8) not_removed (); } When compiled with -O2, the resulting assembly file does not include a call to not_removed. But in fact it should be called in seven out of eight cases. ((start + len) % 8) is an unsigned int in [0, 7]. "((start + len) % 8) - 8" is an unsigned int in [UINT_MAX - 8, UINT_MAX]. cur_bitshift is an int, which can not represent UINT_MAX; this is an implementation defined conversion, and the GCC manual says that the value is reduced modulo (UINT_MAX+1) until it fits in the range of int. So the defined result of the conversion is in [-8, -1]. But VRP says: Value ranges after VRP: start_1(D): VARYING len_2(D): VARYING D.1559_3: [0, +INF] EQUIVALENCES: { } (0 elements) D.1560_4: [0, 7] EQUIVALENCES: { } (0 elements) D.1561_5: [4294967288, +INF] EQUIVALENCES: { } (0 elements) cur_bitshift_6: [-INF(OVF), -2147483648] EQUIVALENCES: { } (0 elements) Folding predicate cur_bitshift_6 >= -7 to 0 Removing basic block 3 Merging blocks 2 and 4 put_field (start, len) { int cur_bitshift; unsigned int D.1561; unsigned int D.1560; unsigned int D.1559; <bb 2>: D.1559_3 = start_1(D) + len_2(D); D.1560_4 = D.1559_3 & 7; D.1561_5 = D.1560_4 + 4294967288; cur_bitshift_6 = (int) D.1561_5; return; } -- Summary: VRP eliminates a useful test due to overflow Product: gcc Version: 4.3.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: tree-optimization AssignedTo: unassigned at gcc dot gnu dot org ReportedBy: drow at gcc dot gnu dot org GCC target triplet: mips-linux-gnu http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31605