LLVM commit [1] changed loop pragma behavior such that
full loop unroll is always honored with user pragma.
Previously, unroll count also depends on the unrolled
code size. For pyperf600, without [1], the loop unroll
count is 150. With [1], the loop unroll count is 600.
The unroll count of 600 caused the program size close to
298k and this caused the following code is generated:
0: 7b 1a 00 ff 00 00 00 00 *(u64 *)(r10 - 256) = r1
; uint64_t pid_tgid = bpf_get_current_pid_tgid();
1: 85 00 00 00 0e 00 00 00 call 14
2: bf 06 00 00 00 00 00 00 r6 = r0
; pid_t pid = (pid_t)(pid_tgid >> 32);
3: bf 61 00 00 00 00 00 00 r1 = r6
4: 77 01 00 00 20 00 00 00 r1 >>= 32
5: 63 1a fc ff 00 00 00 00 *(u32 *)(r10 - 4) = r1
6: bf a2 00 00 00 00 00 00 r2 = r10
7: 07 02 00 00 fc ff ff ff r2 += -4
; PidData* pidData = bpf_map_lookup_elem(&pidmap, &pid);
8: 18 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r1 = 0 ll
10: 85 00 00 00 01 00 00 00 call 1
11: bf 08 00 00 00 00 00 00 r8 = r0
; if (!pidData)
12: 15 08 15 e8 00 00 00 00 if r8 == 0 goto -6123 <LBB0_27588+0xffffffffffdae100>
Note that insn 12 has a branch offset -6123 which is clearly illegal
and will be rejected by the verifier. The negative offset is due to
the branch range is greater than INT16_MAX.
This patch changed the unroll count to be 150 to avoid above
branch target insn out-of-range issue. Also the llvm is enhanced ([2])
to assert if the branch target insn is out of INT16 range.
[1] https://reviews.llvm.org/D119148
[2] https://reviews.llvm.org/D123877
Signed-off-by: Yonghong Song <yhs@xxxxxx>
---
tools/testing/selftests/bpf/progs/pyperf.h | 4 ++++
tools/testing/selftests/bpf/progs/pyperf600.c | 11 +++++++----
2 files changed, 11 insertions(+), 4 deletions(-)
diff --git a/tools/testing/selftests/bpf/progs/pyperf.h b/tools/testing/selftests/bpf/progs/pyperf.h
index 1ed28882daf3..5d3dc4d66d47 100644
--- a/tools/testing/selftests/bpf/progs/pyperf.h
+++ b/tools/testing/selftests/bpf/progs/pyperf.h
@@ -299,7 +299,11 @@ int __on_event(struct bpf_raw_tracepoint_args *ctx)
#ifdef NO_UNROLL
#pragma clang loop unroll(disable)
#else
+#ifdef UNROLL_COUNT
+#pragma clang loop unroll_count(UNROLL_COUNT)
+#else
#pragma clang loop unroll(full)
+#endif
#endif /* NO_UNROLL */
/* Unwind python stack */
for (int i = 0; i < STACK_MAX_LEN; ++i) {
diff --git a/tools/testing/selftests/bpf/progs/pyperf600.c b/tools/testing/selftests/bpf/progs/pyperf600.c
index cb49b89e37cd..ce1aa5189cc4 100644
--- a/tools/testing/selftests/bpf/progs/pyperf600.c
+++ b/tools/testing/selftests/bpf/progs/pyperf600.c
@@ -1,9 +1,12 @@
// SPDX-License-Identifier: GPL-2.0
// Copyright (c) 2019 Facebook
#define STACK_MAX_LEN 600
-/* clang will not unroll the loop 600 times.
- * Instead it will unroll it to the amount it deemed
- * appropriate, but the loop will still execute 600 times.
- * Total program size is around 90k insns
+/* Full unroll of 600 iterations will have total
+ * program size close to 298k insns and this may
+ * cause BPF_JMP insn out of 16-bit integer range.
+ * So limit the unroll size to 150 so the
+ * total program size is around 80k insns but
+ * the loop will still execute 600 times.
*/
+#define UNROLL_COUNT 150
#include "pyperf.h"
--
2.30.2
