With baseline x64 instruction set, shift count can only be an immediate or in %cl. The implicit dependency on %cl makes it necessary to shuffle registers around and/or add push/pop operations. BMI2 provides shift instructions that can use any general register as the shift count, saving us instructions and a few bytes in most cases. Suboptimal codegen when %ecx is source and/or destination is also addressed and unnecessary instructions are removed. test_progs: Summary: 267/1340 PASSED, 25 SKIPPED, 0 FAILED test_progs-no_alu32: Summary: 267/1333 PASSED, 26 SKIPPED, 0 FAILED test_verifier: Summary: 1367 PASSED, 636 SKIPPED, 0 FAILED (same result with or without BMI2) test_maps: OK, 0 SKIPPED lib/test_bpf: test_bpf: Summary: 1026 PASSED, 0 FAILED, [1014/1014 JIT'ed] test_bpf: test_tail_calls: Summary: 10 PASSED, 0 FAILED, [10/10 JIT'ed] test_bpf: test_skb_segment: Summary: 2 PASSED, 0 FAILED --- v3 -> v4: - Fixed a regression when BMI2 isn't available Jie Meng (3): bpf,x64: avoid unnecessary instructions when shift dest is ecx bpf,x64: use shrx/sarx/shlx when available bpf: add selftests for lsh, rsh, arsh with reg operand arch/x86/net/bpf_jit_comp.c | 89 +++++++++++++++++++--- tools/testing/selftests/bpf/verifier/jit.c | 24 ++++++ 2 files changed, 101 insertions(+), 12 deletions(-) -- 2.30.2
