http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53804
Bug #: 53804 Summary: branch reordering missed optimization Classification: Unclassified Product: gcc Version: unknown Status: UNCONFIRMED Severity: enhancement Priority: P3 Component: tree-optimization AssignedTo: unassig...@gcc.gnu.org ReportedBy: ste...@gcc.gnu.org Consider this test case: int foo1 (int a, int b) { if (a > 0) return 1; else if (b > 0 && a < 0) return -3; return 9; } int foo2 (int a, int b) { if (a > 0) return 1; else if (a < 0 && b > 0) return -3; return 9; } Ideally these two functions would be optimized to the same code, because they are semantically equivalent. The ideal form is foo2 because the result of the first comparison against "a" can be re-used for the second test, but GCC does not perform this optimization. The .227r.final dump looks like this on powerpc64-unknown-linux-gnu (all notes removed for readability): ;; Function foo1 (foo1, funcdef_no=0, decl_uid=1997, cgraph_uid=0) 11 %7:CC=cmp(%3:SI,0) # r7 = cmp(a,0) 5 %9:DI=0x1 # r9 = 1 12 pc={(%7:CC<=0)?L74:pc} # if (r7 <= 0) goto L74 L20: 26 %3:DI=%9:DI # r3 = r9 29 use %3:DI # .. 64 return # return r3 i 63: barrier L74: 14 %7:CC=cmp(%4:SI,0) # r7 = cmp (b,0) 8 %9:DI=0x9 # r9 = 9 15 pc={(%7:CC<=0)?L20:pc} # if (r7 <= 0) goto L20 53 %9:DI=-%3:DI==0 # r9 = -(r3==0) 54 {%9:DI=%9:DI&0xc;clobber scratch;} # r9 = r9 & 12 55 %9:DI=%9:DI-0x3 # r9 = r9 - 3 68 %3:DI=%9:DI # r3 = r9 69 use %3:DI # .. 70 return # return r3 i 73: barrier ;; Function foo2 (foo2, funcdef_no=1, decl_uid=2001, cgraph_uid=1) 11 %7:CC=cmp(%3:SI,0) # r7 = cmp(a,0) 12 pc={(%7:CC<=0)?L57:pc} # if (r7 <= 0) goto L57 5 %3:DI=0x1 # r3 = 1 29 use %3:DI # .. 56 return # return r3 i 55: barrier L57: 14 %7:CC=cmp(%3:DI,0) # r7 = cmp(a,0) // ??? redundant 8 %3:DI=0x9 # r3 = 9 51 use %3:DI # .. 15 pc={(%7:CC==0)?return:pc} # if (r7 == 0) return r3 17 %7:CC=cmp(%4:SI,0) # r7 = cmp(b,0) 52 use %3:DI # .. 18 pc={(%7:CC<=0)?return:pc} # if (r7 <= 0) return r3 6 %3:DI=0xfffffffffffffffd # r3 = -3 53 use %3:DI # .. 54 return # return r3 i 47: barrier Note how foo1 needs two branches whereas foo2 only needs 1. (I'm not sure why there is the redundant compare in foo2:insn 14)