http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55875
Bug #: 55875 Summary: [4.8 Regression] IVopts caused miscompilation Classification: Unclassified Product: gcc Version: 4.8.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: tree-optimization AssignedTo: unassig...@gcc.gnu.org ReportedBy: ja...@gcc.gnu.org The following testcase is miscompiled at -O3 on x86_64-linux starting with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192989 (ok, that revision doesn't compile, needs r192900 follow-up). struct A { short int a1; unsigned char a2; unsigned int a3; }; struct B { unsigned short b1; const A *b2; }; B b; __attribute__((noinline, noclone)) int foo (unsigned x) { __asm volatile ("" : "+r" (x) : : "memory"); return x; } inline void bar (const int &) { } __attribute__((noinline)) void baz () { const A *a = b.b2; unsigned int i; unsigned short n = b.b1; for (i = 0; i < n; ++i) if (a[i].a1 == 11) { if (i > 0 && (a[i - 1].a2 & 1)) continue; bar (foo (2)); return; } } int main () { A a[4] = { { 10, 0, 0 }, { 11, 1, 0 }, { 11, 1, 0 }, { 11, 1, 0 } }; b.b1 = 4; b.b2 = a; baz (); return 0; } In *.slp we have: <bb 5>: if (i_21 != 0) goto <bb 6>; else goto <bb 7>; <bb 6>: _11 = i_21 + 4294967295; _12 = (long unsigned int) _11; _13 = _12 * 8; _14 = a_4 + _13; _15 = _14->a2; but ivopts turns that into: <bb 5>: i_25 = (unsigned int) ivtmp.9_31; if (i_25 != 0) goto <bb 6>; else goto <bb 7>; <bb 6>: _28 = ivtmp.9_31 * 8; _27 = a_4 + _28; _26 = _27 + 34359738362; _15 = MEM[base: _26, offset: 0B]; which is wrong, i_21 + -1U wrapped around (and wasn't executed for i_21 being 0), while 34359738362 is 0xffffffffULL * 8 + 2, thus it ignores the wrapping and does what the original code would do for _12 = (long unsigned int) i_21; _77 = _12 + 4294967295; _13 = _77 * 8; i.e. as if the -1U addition was done in the wider precision.