http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48571
Summary: Missed data-dependence for (bogus?) reconstructed array-refs Product: gcc Version: 4.7.0 Status: UNCONFIRMED Keywords: wrong-code Severity: normal Priority: P3 Component: tree-optimization AssignedTo: unassig...@gcc.gnu.org ReportedBy: rgue...@gcc.gnu.org Target: i?86-*-* We vectorize the following loop, ignoring dependences between c[i] and c[i-1]. unsigned int c[624]; void bar (void) { unsigned int i; for (i = 1; i < 624; ++i) /* Obfuscated c[i] = c[i-1] * 2. */ *(unsigned int *)((void *)c + (__SIZE_TYPE__)i * 4) = *(unsigned int *)((void *)c + ((__SIZE_TYPE__)i + ((__SIZE_TYPE__)-4)/4) * 4) * 2; } This is because we re-construct array-refs <bb 3>: # i_17 = PHI <i_12(4), 1(2)> # ivtmp.8_21 = PHI <ivtmp.8_20(4), 623(2)> D.2689_3 = (long unsigned int) i_17; D.2692_7 = D.2689_3 + 1073741823; D.2695_10 = MEM[(unsigned int *)&c][D.2692_7]{lb: 0 sz: 4}; D.2696_11 = D.2695_10 * 2; MEM[(unsigned int *)&c][D.2689_3]{lb: 0 sz: 4} = D.2696_11; and the array index D.2692_7 is out of bounds (but of course the address computation will make the access wrap into the correct place). The middle-end happily performs such obfuscation via fold_plusminus_mult_expr from (i*4 + -4U). The array-refs are re-built by forwprop. I'm not sure who is wrong here - bogus constrained interpretation of array-refs by data dependence analysis or bogus construction of constrained array-refs from pointer arithmetic. Executable testcase, x86_64, -O3 -msse2 -m32: unsigned int c[624]; void __attribute__((noinline)) bar (void) { unsigned int i; /* Obfuscated c[i] = c[i-1] * 2. */ for (i = 1; i < 624; ++i) *(unsigned int *)((void *)c + (__SIZE_TYPE__)i * 4) = 2 * *(unsigned int *)((void *)c + ((__SIZE_TYPE__)i + ((__SIZE_TYPE__)-4)/4) * 4); } extern void abort (void); int main() { unsigned int i, j; for (i = 0; i < 624; ++i) c[i] = 1; bar(); j = 1; for (i = 0; i < 624; ++i) { if (c[i] != j) abort (); j = j * 2; } return 0; }