Hi David.
Thanks for the patch.
See a couple of comments regarding error handling below.
> BPF programs are not typically linked, which means we cannot fall back
> on library calls to implement __builtin_{memmove,memcpy} and should
> always expand them inline if possible.
>
> GCC already successfully expands these builtins inline in many cases,
> but failed to do so for a few for simple cases involving overlapping
> memmove in the kernel BPF selftests and was instead emitting a libcall.
>
> This patch implments a simple inline expansion of memcpy and memmove in
s/implments/implements/.
> the BPF backend in a verifier-friendly way, with the caveat that the
> size must be an integer constant, which is also required by clang.
>
> Tested for bpf-unknown-none on x86_64-linux-gnu host.
>
> Also tested against the BPF verifier by compiling and loading a test
> program with overlapping memmove (essentially the memmove-1.c test)
> which failed before due to a libcall, and now successfully loads and
> passes the verifier.
>
> gcc/
>
> * config/bpf/bpf-protos.h (bpf_expand_cpymem): New.
> * config/bpf/bpf.cc: (emit_move_loop, CPYMEM_EXPAND_ERR)
> (bpf_expand_cpymem): New.
> * config/bpf/bpf.md: (cpymemdi, movmemdi): New define_expands.
>
> gcc/testsuite/
>
> * gcc.target/bpf/memcpy-1.c: New test.
> * gcc.target/bpf/memmove-1.c: New test.
> * gcc.target/bpf/memmove-2.c: New test.
> ---
> gcc/config/bpf/bpf-protos.h | 2 +
> gcc/config/bpf/bpf.cc | 129 +++++++++++++++++++++++
> gcc/config/bpf/bpf.md | 36 +++++++
> gcc/testsuite/gcc.target/bpf/memcpy-1.c | 26 +++++
> gcc/testsuite/gcc.target/bpf/memmove-1.c | 46 ++++++++
> gcc/testsuite/gcc.target/bpf/memmove-2.c | 23 ++++
> 6 files changed, 262 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/bpf/memcpy-1.c
> create mode 100644 gcc/testsuite/gcc.target/bpf/memmove-1.c
> create mode 100644 gcc/testsuite/gcc.target/bpf/memmove-2.c
>
> diff --git a/gcc/config/bpf/bpf-protos.h b/gcc/config/bpf/bpf-protos.h
> index 46d950bd990..366acb87ae4 100644
> --- a/gcc/config/bpf/bpf-protos.h
> +++ b/gcc/config/bpf/bpf-protos.h
> @@ -35,4 +35,6 @@ const char *bpf_add_core_reloc (rtx *operands, const char
> *templ);
> class gimple_opt_pass;
> gimple_opt_pass *make_pass_lower_bpf_core (gcc::context *ctxt);
>
> +bool bpf_expand_cpymem (rtx *, bool);
> +
> #endif /* ! GCC_BPF_PROTOS_H */
> diff --git a/gcc/config/bpf/bpf.cc b/gcc/config/bpf/bpf.cc
> index d6ca47eeecb..c90d29d12ff 100644
> --- a/gcc/config/bpf/bpf.cc
> +++ b/gcc/config/bpf/bpf.cc
> @@ -1184,6 +1184,135 @@ bpf_use_by_pieces_infrastructure_p (unsigned
> HOST_WIDE_INT size,
> #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
> bpf_use_by_pieces_infrastructure_p
>
> +/* Helper for bpf_expand_cpymem. Emit an unrolled loop moving the bytes
> + from SRC to DST. */
> +
> +static void
> +emit_move_loop (rtx src, rtx dst, machine_mode mode, int offset, int inc,
> + unsigned iters, unsigned remainder)
> +{
> + rtx reg = gen_reg_rtx (mode);
> +
> + /* First copy in chunks as large as alignment permits. */
> + for (unsigned int i = 0; i < iters; i++)
> + {
> + emit_move_insn (reg, adjust_address (src, mode, offset));
> + emit_move_insn (adjust_address (dst, mode, offset), reg);
> + offset += inc;
> + }
> +
> + /* Handle remaining bytes which might be smaller than the chunks
> + used above. */
> + if (remainder & 4)
> + {
> + emit_move_insn (reg, adjust_address (src, SImode, offset));
> + emit_move_insn (adjust_address (dst, SImode, offset), reg);
> + offset += (inc < 0 ? -4 : 4);
> + remainder -= 4;
> + }
> + if (remainder & 2)
> + {
> + emit_move_insn (reg, adjust_address (src, HImode, offset));
> + emit_move_insn (adjust_address (dst, HImode, offset), reg);
> + offset += (inc < 0 ? -2 : 2);
> + remainder -= 2;
> + }
> + if (remainder & 1)
> + {
> + emit_move_insn (reg, adjust_address (src, QImode, offset));
> + emit_move_insn (adjust_address (dst, QImode, offset), reg);
> + }
> +}
> +
> +/* Error if we cannot completely expand the memcpy/memmove inline,
> + unless we are building libgcc. */
> +#define CPYMEM_EXPAND_ERR(LOC, MSG) \
> + do \
> + {
> \
> + if (flag_building_libgcc)
> \
> + warning (LOC, MSG " for BPF memcpy/memmove expansion; " \
> + "a libcall will be emitted"); \
> + else \
> + error (MSG " for BPF memcpy/memmove expansion"); \
> + } while (0);
For this error I would suggest to use a less "internal" jargon to avoid
confusing the user. What about something like:
"couldn't inline call to __builtin_memcpy: " MSG
"couldn't inline call to __builtin_memmove: " MSG
> +/* Expand cpymem/movmem, as from __builtin_memcpy/memmove.
> + OPERANDS are the same as the cpymem/movmem patterns.
> + IS_MOVE is true if this is a memmove, false for memcpy.
> + Return true if we successfully expanded, or false if we cannot
> + and must punt to a libcall. */
> +
> +bool
> +bpf_expand_cpymem (rtx *operands, bool is_move)
> +{
> + /* Size and alignment must be constant. */
> + if (!CONST_INT_P (operands[2]))
> + {
> + CPYMEM_EXPAND_ERR (RTL_LOCATION (operands[2]),
> + "unsupported variable size");
> + return false;
> + }
I find this error message a little confusing. I am not a native English
speaker, but when I read "unsupported variable size" I get the
impression other variable sizes may be supported.
Would perhaps be more clear to say something like "buffer size must be
constant"?
> + if (!CONST_INT_P (operands[3]))
> + {
> + CPYMEM_EXPAND_ERR (RTL_LOCATION (operands[3]),
> + "unsupported variable alignment");
> + return false;
> + }
The documentation says (gccint):
The fourth operand is the known shared alignment of the source and
destination, in the form of a 'const_int' rtx.
For both memcpyN and memmovN patterns.
So I think we can assert on the fact operands[3] is constant?
> + rtx dst = operands[0];
> + rtx src = operands[1];
> + rtx size = operands[2];
> + unsigned HOST_WIDE_INT size_bytes = UINTVAL (size);
> + unsigned align = UINTVAL (operands[3]);
> + enum machine_mode mode;
> + switch (align)
> + {
> + case 1: mode = QImode; break;
> + case 2: mode = HImode; break;
> + case 4: mode = SImode; break;
> + case 8: mode = DImode; break;
> + default:
> + CPYMEM_EXPAND_ERR (RTL_LOCATION (operands[3]),
> + "unsupported alignment");
> + return false;
> + }
Should this really be an error? In what normal circumstances we may
expect an odd alignment != 1, or bigger than 8? Shouldn't we assert/ICE
here instead? (The fact you are not including a test triggering this
error may be hinting at this.)
> + unsigned iters = size_bytes >> ceil_log2 (align);
> + unsigned remainder = size_bytes & (align - 1);
> +
> + int inc = GET_MODE_SIZE (mode);
> + rtx_code_label *fwd_label, *done_label;
> + if (is_move)
> + {
> + /* For memmove, be careful of overlap. It is not a concern for memcpy.
> + To handle overlap, we check (at runtime) if SRC < DST, and if so do
> + the move "backwards" starting from SRC + SIZE. */
> + fwd_label = gen_label_rtx ();
> + done_label = gen_label_rtx ();
> +
> + rtx dst_addr = copy_to_mode_reg (Pmode, XEXP (dst, 0));
> + rtx src_addr = copy_to_mode_reg (Pmode, XEXP (src, 0));
> + emit_cmp_and_jump_insns (src_addr, dst_addr, GEU, NULL_RTX, Pmode,
> + true, fwd_label, profile_probability::even ());
> +
> + /* Emit the "backwards" unrolled loop. */
> + emit_move_loop (src, dst, mode, size_bytes, -inc, iters, remainder);
> + emit_jump_insn (gen_jump (done_label));
> + emit_barrier ();
> +
> + emit_label (fwd_label);
> + }
> +
> + emit_move_loop (src, dst, mode, 0, inc, iters, remainder);
> +
> + if (is_move)
> + emit_label (done_label);
> +
> + return true;
> +}
> +
> +#undef CPYMEM_EXPAND_ERR
> +
> /* Finally, build the GCC target. */
>
> struct gcc_target targetm = TARGET_INITIALIZER;
> diff --git a/gcc/config/bpf/bpf.md b/gcc/config/bpf/bpf.md
> index 50df1aaa3e2..ca677bc6b50 100644
> --- a/gcc/config/bpf/bpf.md
> +++ b/gcc/config/bpf/bpf.md
> @@ -627,4 +627,40 @@ (define_insn "ldabs<ldop>"
> "{ldabs<ldop>\t%0|r0 = *(<pldop> *) skb[%0]}"
> [(set_attr "type" "ld")])
>
> +;;; memmove and memcopy
> +
> +;; 0 is dst
> +;; 1 is src
> +;; 2 is size of copy in bytes
> +;; 3 is alignment
> +
> +(define_expand "cpymemdi"
> + [(match_operand:BLK 0 "memory_operand")
> + (match_operand:BLK 1 "memory_operand")
> + (match_operand:DI 2 "general_operand")
> + (match_operand:DI 3 "immediate_operand")]
> + ""
> +{
> + if (bpf_expand_cpymem (operands, false))
> + DONE;
> + FAIL;
> +})
> +
> +;; 0 is dst
> +;; 1 is src
> +;; 2 is size of copy in bytes
> +;; 3 is alignment
> +
> +(define_expand "movmemdi"
> + [(match_operand:BLK 0 "memory_operand")
> + (match_operand:BLK 1 "memory_operand")
> + (match_operand:DI 2 "general_operand")
> + (match_operand:DI 3 "immediate_operand")]
> + ""
> +{
> + if (bpf_expand_cpymem (operands, true))
> + DONE;
> + FAIL;
> +})
> +
> (include "atomic.md")
> diff --git a/gcc/testsuite/gcc.target/bpf/memcpy-1.c
> b/gcc/testsuite/gcc.target/bpf/memcpy-1.c
> new file mode 100644
> index 00000000000..6c9707f24e8
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/bpf/memcpy-1.c
> @@ -0,0 +1,26 @@
> +/* Ensure memcpy is expanded inline rather than emitting a libcall. */
> +
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +
> +struct context {
> + unsigned int data;
> + unsigned int data_end;
> + unsigned int data_meta;
> + unsigned int ingress;
> + unsigned int queue_index;
> + unsigned int egress;
> +};
> +
> +void
> +cpy_1(struct context *ctx)
> +{
> + void *data = (void *)(long)ctx->data;
> + char *dest;
> + dest = data;
> + dest += 16;
> +
> + __builtin_memcpy (dest, data, 8);
> +}
> +
> +/* { dg-final { scan-assembler-times "call" 0 } } */
> diff --git a/gcc/testsuite/gcc.target/bpf/memmove-1.c
> b/gcc/testsuite/gcc.target/bpf/memmove-1.c
> new file mode 100644
> index 00000000000..3b8ba82639e
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/bpf/memmove-1.c
> @@ -0,0 +1,46 @@
> +/* Ensure memmove is expanded inline rather than emitting a libcall. */
> +
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +
> +struct context {
> + unsigned int data;
> + unsigned int data_end;
> + unsigned int data_meta;
> + unsigned int ingress;
> + unsigned int queue_index;
> + unsigned int egress;
> +};
> +
> +void
> +mov_1_nooverlap (struct context *ctx)
> +{
> + void *data = (void *)(long)ctx->data;
> + char *dest;
> + dest = data;
> + dest += 16;
> +
> + __builtin_memmove (dest, data, 12);
> +}
> +
> +void
> +mov_1_overlap (struct context *ctx)
> +{
> + void *data = (void *)(long)ctx->data;
> + char *dest;
> + dest = data;
> + dest += 4;
> +
> + __builtin_memmove (dest, data, 12);
> +}
> +
> +void
> +mov_1_arbitrary (struct context *ctx_a, struct context *ctx_b)
> +{
> + void *src = (void *)(long)ctx_a->data;
> + void *dst = (void *)(long)ctx_b->data;
> +
> + __builtin_memmove (dst, src, 12);
> +}
> +
> +/* { dg-final { scan-assembler-times "call" 0 } } */
> diff --git a/gcc/testsuite/gcc.target/bpf/memmove-2.c
> b/gcc/testsuite/gcc.target/bpf/memmove-2.c
> new file mode 100644
> index 00000000000..0ef34131080
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/bpf/memmove-2.c
> @@ -0,0 +1,23 @@
> +/* Test that we error if memmove cannot be expanded inline. */
> +
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +
> +typedef unsigned int __u32;
> +
> +struct context {
> + unsigned int data;
> + unsigned int data_end;
> + unsigned int data_meta;
> +};
> +
> +void
> +mov_2_unsupported (struct context *ctx)
> +{
> + void *data = (void *)(long)ctx->data;
> + char *dest;
> + dest = data;
> + dest += 4;
> +
> + __builtin_memmove (dest, data, ctx->data_meta); /* { dg-error "unsupported
> variable size" } */
> +}
