dscpPolicy: more bpf verifier optimizations
This change avoids the use of multiple conditionals/ifs
by merging them all into one via bitwise arithmetic.
Bug: 361492282
Test: TreeHugger, atest CtsNetTestCases:android.net.cts.DscpPolicyTest
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Change-Id: I05c2f9cbddf14e0388cb8ac99c63ce0c76e46960
diff --git a/bpf/progs/bpf_net_helpers.h b/bpf/progs/bpf_net_helpers.h
index a86c3e6..a5664ba 100644
--- a/bpf/progs/bpf_net_helpers.h
+++ b/bpf/progs/bpf_net_helpers.h
@@ -139,6 +139,24 @@
if (skb->data_end - skb->data < len) bpf_skb_pull_data(skb, len);
}
+// anti-compiler-optimizer no-op: explicitly force full calculation of 'v'
+//
+// The use for this is to force full calculation of a complex arithmetic (likely binary
+// bitops) value, and then check the result only once (thus likely reducing the number
+// of required conditional jump instructions that badly affect bpf verifier runtime)
+//
+// The compiler cannot look into the assembly statement, so it doesn't know it does nothing.
+// Since the statement takes 'v' as both input and output in a register (+r),
+// the compiler must fully calculate the precise value of 'v' before this,
+// and must use the (possibly modified) value of 'v' afterwards (thus cannot
+// do funky optimizations to use partial results from before the asm).
+//
+// As this is not flagged 'volatile' this may still be moved out of a loop,
+// or even entirely optimized out if 'v' is never used afterwards.
+//
+// See: https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
+#define COMPILER_FORCE_CALCULATION(v) asm ("" : "+r" (v))
+
struct egress_bool { bool egress; };
#define INGRESS ((struct egress_bool){ .egress = false })
#define EGRESS ((struct egress_bool){ .egress = true })