On 7/10/26 12:12 PM, Vineet Gupta wrote:
The default cost of MULT/DIV/MOD in rtx_cost () is high: 5,7 insns respectively. This causes even a trivial mpy by 7 to be synthesized. Given these have direct equivalents in BPF ISA, fix the cost to generate native BPF insns.Note the existing divmod-licall-2.c test was a bit fragile as it forced cast signed an actual unsigned int which is provably non-negative. In the new cost model compiler would generate a native unsigned divide not available for -mcpu=v3, tripping up the test. Fix by ensuring the arg is actually signed. gcc/ChangeLog: * config/bpf/bpf.cc (bpf_rtx_costs): Assign MPY/DIV/MOD cost 1. gcc/testsuite/ChangeLog: * gcc.target/bpf/divmod-libcall-2.c: Change args/ret to signed. * gcc.target/bpf/mult-large.c: New test. * gcc.target/bpf/mult-small.c: New test. Signed-off-by: Vineet Gupta <[email protected]>
Please ignore this - fat fingered - only difference is typo fix in target hook name in subject
Thx, -Vineet
--- gcc/config/bpf/bpf.cc | 24 ++++++++++++++++--- .../gcc.target/bpf/divmod-libcall-2.c | 15 +++++++----- gcc/testsuite/gcc.target/bpf/mult-large.c | 22 +++++++++++++++++ gcc/testsuite/gcc.target/bpf/mult-small.c | 20 ++++++++++++++++ 4 files changed, 72 insertions(+), 9 deletions(-) create mode 100644 gcc/testsuite/gcc.target/bpf/mult-large.c create mode 100644 gcc/testsuite/gcc.target/bpf/mult-small.c diff --git a/gcc/config/bpf/bpf.cc b/gcc/config/bpf/bpf.cc index 89a4917cd0a0..58d2318f972b 100644 --- a/gcc/config/bpf/bpf.cc +++ b/gcc/config/bpf/bpf.cc @@ -602,14 +602,32 @@ bpf_legitimate_address_p (machine_mode mode, `rtx_cost' should recurse. */static bool-bpf_rtx_costs (rtx x ATTRIBUTE_UNUSED, +bpf_rtx_costs (rtx x, enum machine_mode mode ATTRIBUTE_UNUSED, int outer_code ATTRIBUTE_UNUSED, int opno ATTRIBUTE_UNUSED, - int *total ATTRIBUTE_UNUSED, + int *total, bool speed ATTRIBUTE_UNUSED) { - /* To be written. */ + switch (GET_CODE (x)) + { + case MULT: + case DIV: + case UDIV: + case MOD: + case UMOD: + /* BPF implements these as a single instruction, so keep the native + operation cheaper than synthesized sequence. + Only influences choice between actually available alternatives; + if the operation has no insn (e.g. a 64-bit signed divide before + -mcpu=v4) expand_divmod () still falls back to a libcall. + Return false so caller rtx_cost keeps recursing for operands. */ + *total = COSTS_N_INSNS (1); + return false; + + default: + return false; + } return false; }diff --git a/gcc/testsuite/gcc.target/bpf/divmod-libcall-2.c b/gcc/testsuite/gcc.target/bpf/divmod-libcall-2.cindex 792d689395a2..7296579171d2 100644 --- a/gcc/testsuite/gcc.target/bpf/divmod-libcall-2.c +++ b/gcc/testsuite/gcc.target/bpf/divmod-libcall-2.c @@ -1,16 +1,19 @@ +/* Inverse of divmod-libcall-1.c. Ensure libcalls are generated for + -mcpu=v3 due to lack of signed div/mod. */ + /* { dg-do compile } */ /* { dg-options "-O2 -mcpu=v3" } */ /* { dg-final { scan-assembler "global\t__divdi3" } } */ /* { dg-final { scan-assembler "global\t__moddi3" } } */-int-foo (unsigned int len) +long +foo (long len) { - return ((long)len) * 234 / 5; + return len * 234 / 5; }-int-bar (unsigned int len) +long +bar (long len) { - return ((long)len) * 234 % 5; + return len * 234 % 5; } diff --git a/gcc/testsuite/gcc.target/bpf/mult-large.c b/gcc/testsuite/gcc.target/bpf/mult-large.c new file mode 100644 index 000000000000..4a5ef5da2b15 --- /dev/null +++ b/gcc/testsuite/gcc.target/bpf/mult-large.c @@ -0,0 +1,22 @@ +/* The multiply core of memset-4.c: a memset of a non-constant value + broadcasts a byte across the word with "x * 0x01010101". With the generic + COSTS_N_INSNS (5) multiply cost this large-constant multiply was + strength-reduced into a shift/add sequence; BPF has a single-instruction + mul, so the native mul must be used instead. + + Use a full 32-bit input (not 'unsigned char'): for a byte 'x * 0x01010101' + equals the carry-free broadcast 'x | x<<8 | x<<16 | x<<24', which the + compiler could lower without a multiply at all, weakening the test. A + 32-bit input has no such equivalence, so this is a genuine multiply. */ + +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +unsigned +bcast (unsigned x) +{ + return x * 0x01010101u; +} + +/* { dg-final { scan-assembler {\*= 16843009} } } */ +/* { dg-final { scan-assembler-not {<<=} } } */ diff --git a/gcc/testsuite/gcc.target/bpf/mult-small.c b/gcc/testsuite/gcc.target/bpf/mult-small.c new file mode 100644 index 000000000000..7c82f3276266 --- /dev/null +++ b/gcc/testsuite/gcc.target/bpf/mult-small.c @@ -0,0 +1,20 @@ +/* Verify that a multiply by a small constant uses BPF's native single + instruction mul rather than a synthesized shift/add(-sub) sequence. + + BPF has a single-instruction multiply, so "x * 7" should be "r0 *= 7", + not the strength-reduced "(x << 3) - x". synth_mult () only avoids the + native multiply when it finds a cheaper sequence, so the rtx_cost hook + must report MULT as a single instruction (rather than the generic + COSTS_N_INSNS (5) schoolbook default) for the native op to win. */ + +/* { dg-do compile } */ +/* { dg-options "-O2 -mcpu=v4" } */ + +unsigned +mul7 (unsigned x) +{ + return x * 7; +} + +/* { dg-final { scan-assembler {\*= 7} } } */ +/* { dg-final { scan-assembler-not {<<=} } } */
