On Mon, Nov 30, 2020 at 09:11:10PM +0800, Hongtao Liu wrote:
> +;; PR96906 - optimize vpsubusw compared to 0 into vpcmpleuw or vpcmpnltuw.
> +(define_split
> + [(set (match_operand:<avx512fmaskmode> 0 "register_operand")
> + (unspec:<avx512fmaskmode>
> + [(us_minus:VI12_AVX512VL
> + (match_operand:VI12_AVX512VL 1 "vector_operand")
> + (match_operand:VI12_AVX512VL 2 "vector_operand"))
> + (match_operand:VI12_AVX512VL 3 "const0_operand")
> + (match_operand:SI 4 "const0_operand")]
> + UNSPEC_PCMP))]
> + "TARGET_AVX512BW && ix86_binary_operator_ok (US_MINUS, <MODE>mode,
> operands)"
Too long line, please wrap it.
Also, INTVAL (operands[4]) == 0 is EQ comparison, can't we handle also
NE (i.e. INTVAL (operands[4]) == 4?
I.e. replace the "const0_operand" in there with "const_0_to_7_operand"
and check in conditions that (INTVAL (operands[4]) & 3) == 0.
> + [(const_int 0)]
> + {
> + /* LE: 2, NLT: 5. */
> + rtx cmp_predicate = GEN_INT (2);
> + if (MEM_P (operands[1]))
> + {
> + std::swap (operands[1], operands[2]);
> + cmp_predicate = GEN_INT (5);
For INTVAL (operands[4]) == 4 it would then be cmp_predictate NLE: 4 resp.
LT: 3 I think.
Also, this handles only UNSPEC_PCMP, can't we handle UNSPEC_UNSIGNED_PCMP
too? I mean, for equality comparisons it doesn't really matter if we have
signed or unsigned == or !=. And for unsigned
x == 0U is equivalent to x <= 0U, and x != 0U equivalent to x > 0U.
Jakub
