On Thu, Jan 10, 2019 at 11:20 PM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> The following testcase ICEs in dwarf2out.c, because a few sse.md patterns
> contain invalid RTL, in particular
> (const_vector:V2SF [(const_int 0) (const_int 0)])
> Elements of a V2SF const_vector should be (const_double:SF 0), not
> (const_int 0). Unfortunately, we can't add explicitly const_double 0
> constants the way one can write (const_int 0), so this patch uses
> separate define_expand to add those CONST0_RTX args and match_operand
> with "const0_rtx" "C" to match that.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2019-01-10 Jakub Jelinek <ja...@redhat.com>
>
> PR target/88785
> * config/i386/sse.md (float<floatunssuffix>v2div2sf2): Turn into
> define_expand.
> (*float<floatunssuffix>v2div2sf2): New define_insn.
> (float<floatunssuffix>v2div2sf2_mask): Turn into define_expand.
> (*float<floatunssuffix>v2div2sf2_mask): New define_insn.
> (*float<floatunssuffix>v2div2sf2_mask_1): Replace
> subrtxes (const_vector:V2SF [(const_int 0) (const_int 0)]) with
> match_operands with "const0_operand" "C".
>
> * g++.target/i386/pr88785.C: New test.
OK.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2019-01-10 11:43:17.520326228 +0100
> +++ gcc/config/i386/sse.md 2019-01-10 12:57:52.946797987 +0100
> @@ -5222,11 +5222,19 @@ (define_insn "float<floatunssuffix><ssel
> (set_attr "prefix" "evex")
> (set_attr "mode" "<MODE>")])
>
> -(define_insn "float<floatunssuffix>v2div2sf2"
> +(define_expand "float<floatunssuffix>v2div2sf2"
> [(set (match_operand:V4SF 0 "register_operand" "=v")
> (vec_concat:V4SF
> (any_float:V2SF (match_operand:V2DI 1 "nonimmediate_operand"
> "vm"))
> - (const_vector:V2SF [(const_int 0) (const_int 0)])))]
> + (match_dup 2)))]
> + "TARGET_AVX512DQ && TARGET_AVX512VL"
> + "operands[2] = CONST0_RTX (V2SFmode);")
> +
> +(define_insn "*float<floatunssuffix>v2div2sf2"
> + [(set (match_operand:V4SF 0 "register_operand" "=v")
> + (vec_concat:V4SF
> + (any_float:V2SF (match_operand:V2DI 1 "nonimmediate_operand"
> "vm"))
> + (match_operand:V2SF 2 "const0_operand" "C")))]
> "TARGET_AVX512DQ && TARGET_AVX512VL"
> "vcvt<floatsuffix>qq2ps{x}\t{%1, %0|%0, %1}"
> [(set_attr "type" "ssecvt")
> @@ -5260,16 +5268,29 @@ (define_expand "vec_pack<floatprefix>_fl
> DONE;
> })
>
> -(define_insn "float<floatunssuffix>v2div2sf2_mask"
> +(define_expand "float<floatunssuffix>v2div2sf2_mask"
> + [(set (match_operand:V4SF 0 "register_operand" "=v")
> + (vec_concat:V4SF
> + (vec_merge:V2SF
> + (any_float:V2SF (match_operand:V2DI 1 "nonimmediate_operand"
> "vm"))
> + (vec_select:V2SF
> + (match_operand:V4SF 2 "nonimm_or_0_operand" "0C")
> + (parallel [(const_int 0) (const_int 1)]))
> + (match_operand:QI 3 "register_operand" "Yk"))
> + (match_dup 4)))]
> + "TARGET_AVX512DQ && TARGET_AVX512VL"
> + "operands[4] = CONST0_RTX (V2SFmode);")
> +
> +(define_insn "*float<floatunssuffix>v2div2sf2_mask"
> [(set (match_operand:V4SF 0 "register_operand" "=v")
> (vec_concat:V4SF
> (vec_merge:V2SF
> - (any_float:V2SF (match_operand:V2DI 1 "nonimmediate_operand"
> "vm"))
> + (any_float:V2SF (match_operand:V2DI 1 "nonimmediate_operand"
> "vm"))
> (vec_select:V2SF
> (match_operand:V4SF 2 "nonimm_or_0_operand" "0C")
> (parallel [(const_int 0) (const_int 1)]))
> (match_operand:QI 3 "register_operand" "Yk"))
> - (const_vector:V2SF [(const_int 0) (const_int 0)])))]
> + (match_operand:V2SF 4 "const0_operand" "C")))]
> "TARGET_AVX512DQ && TARGET_AVX512VL"
> "vcvt<floatsuffix>qq2ps{x}\t{%1, %0%{%3%}%N2|%0%{%3%}%N2, %1}"
> [(set_attr "type" "ssecvt")
> @@ -5282,9 +5303,9 @@ (define_insn "*float<floatunssuffix>v2di
> (vec_merge:V2SF
> (any_float:V2SF (match_operand:V2DI 1
> "nonimmediate_operand" "vm"))
> - (const_vector:V2SF [(const_int 0) (const_int 0)])
> + (match_operand:V2SF 3 "const0_operand" "C")
> (match_operand:QI 2 "register_operand" "Yk"))
> - (const_vector:V2SF [(const_int 0) (const_int 0)])))]
> + (match_operand:V2SF 4 "const0_operand" "C")))]
> "TARGET_AVX512DQ && TARGET_AVX512VL"
> "vcvt<floatsuffix>qq2ps{x}\t{%1, %0%{%2%}%{z%}|%0%{%2%}%{z%}, %1}"
> [(set_attr "type" "ssecvt")
> --- gcc/testsuite/g++.target/i386/pr88785.C.jj 2019-01-10 13:08:22.987439456
> +0100
> +++ gcc/testsuite/g++.target/i386/pr88785.C 2019-01-10 13:08:17.396531359
> +0100
> @@ -0,0 +1,197 @@
> +// PR target/88785
> +// { dg-do compile }
> +// { dg-options "-O2 -g -std=c++17 -mavx512vl -mavx512dq" }
> +
> +namespace a {
> +template <class> class b;
> +template <class> class d;
> +}
> +template <int f> struct g { static constexpr int e = f; };
> +template <typename> struct aa;
> +template <typename...> struct o;
> +template <typename h, typename ab> struct o<h, ab> : aa<h>::ac {};
> +template <typename...> struct j;
> +template <typename h, typename ab> struct j<h, ab> : aa<ab>::ac {};
> +template <typename... k> constexpr bool l = o<k...>::e;
> +template <typename, typename> struct r : g<false> {};
> +template <typename m> struct r<m, m> : g<true> {};
> +template <typename ad> struct aa { typedef ad ac; };
> +template <bool, typename ad, typename> using ae = ad;
> +template <typename...> using af = void;
> +typedef float ag __attribute__((__vector_size__(16)));
> +ag ah;
> +ag ai(__attribute__((__vector_size__(2 * sizeof(long long)))) long long z) {
> + ah = ag{};
> + __attribute__((__vector_size__(4 * sizeof(float)))) float aj = ah;
> + return __builtin_ia32_cvtqq2ps128_mask(z, aj, 1);
> +}
> +namespace a {
> +int ak;
> +int al;
> +template <long> int am;
> +template <class> struct an;
> +template <class> struct ao;
> +template <class m> using ap = typename ao<m>::ac;
> +template <class, unsigned long, class = af<>> struct aq;
> +template <class m> using ar = aq<m, sizeof(m)>;
> +template <class as, class at> as au(at);
> +template <class> struct av;
> +template <class m> using aw = typename av<m>::ac;
> +template <class m, int ax> struct ay {
> + static constexpr long az = ax * sizeof(int);
> + using ac [[gnu::__vector_size__(az)]] = m;
> +};
> +template <class m, long ba> using bb = typename ay<m, ba>::ac;
> +template <class m> struct p {
> + using bc = decltype(m()[0]);
> + static constexpr int bd = sizeof(bc);
> +};
> +template <class as, class m, class = p<m>> as be(m);
> +template <class as, class... at> as bf(at... z) { return be<as>(z...); }
> +template <class m, unsigned long ax, class bg = aw<m>> bg bh(aq<m, ax>) {
> + return bg();
> +}
> +template <> struct av<float> {
> + using ac [[gnu::__vector_size__(16)]] = float;
> +};
> +template <class> struct av {
> + using ac [[gnu::__vector_size__(16)]] = long long;
> +};
> +template <unsigned long bi> struct aq<bool, bi> {};
> +template <class m, unsigned long bi, class bj = bb<m, bi>,
> + bool = l<r<bb<m, bi>, aw<m>>, r<bj, m>>>
> +struct bk;
> +template <class m, unsigned long bi, class bj> struct bk<m, bi, bj, true> {
> + bj bl;
> + bk(bb<m, bi> z) : bl(z) {}
> +};
> +template <class m, unsigned long bi, class bj> struct bk<m, bi, bj, false>
> {};
> +template <class m, unsigned long bi> struct aq<m, bi> : bk<m, bi> {
> + using bm = bb<m, bi>;
> + static constexpr long bd = bi;
> + aq();
> + template <class bn> aq(bn z) : bk<m, bi>(z) {}
> + m operator[](long);
> +};
> +template <class> constexpr long bo = g<0>::e;
> +template <class m> struct ao { using ac = typename an<m>::br; };
> +template <class bp, class m> class bq {
> + using bu = m;
> + bp bs;
> + m bt;
> +
> +public:
> + template <class bn, class bv> void q(bn z, bv bw) {
> + auto s = bx(bt), by = bx(bs);
> + ap<bu>::bz(s, z, bw, by);
> + }
> +};
> +class ca {
> +public:
> + template <class bn, class bv> void cb(bn, bv);
> +};
> +template <class m> ca cc(typename b<m>::cd, m &);
> +template <class m> bq<d<m>, b<m>> cc(typename b<m>::cd, const m &);
> +struct ce;
> +template <class t, int ax> struct cf {
> + using br = typename t::br;
> + using cg = aq<long, ax>;
> + using ch = aq<long, ax>;
> +};
> +struct ci {
> + template <class m> static constexpr long cj = sizeof(m);
> + struct ck : j<int, g<sizeof(int)>> {};
> + template <class> static constexpr bool cl = ck::e;
> + using br = ce;
> + template <class m> using cn = ae<cl<m>, cf<ci, cj<m>>, int>;
> +};
> +template <class> struct an : ci::cn<long> {};
> +template <class> class d : an<int> {
> + using cm = ch;
> +
> +public:
> + cm bl;
> +};
> +template <class m> auto bx(m z) { return z.bl; }
> +template <class> class b : an<int> {
> + using cm = cg;
> +
> +public:
> + using cd = d<long>;
> + static long cu();
> + b();
> + template <class bn, class bv> b(bn, bv) {}
> + cm bl;
> +};
> +template <class m> auto bx(b<m> z) { return z.bl; }
> +template <class m, class co> void cq(ar<m>, co, aq<bool, ar<m>::bd>);
> +template <class as, class bu, class cp> as be(bu z) {
> + using cs = typename cp::bc;
> + constexpr long ax = cp::bd;
> + auto cr = bh(z);
> + aq<cs, ax> f;
> + using bn = typename p<as>::bc;
> + constexpr bool cy = sizeof(f), ct = sizeof(bn);
> + if (ct)
> + if (cy) {
> + ag cw = ai(cr);
> + return cw;
> + }
> +}
> +template <class as, class at> auto cv(at z) { return bf<as>(z); }
> +struct G {
> + template <class m> using ch = typename ci::cn<m>::ch;
> + template <class m, unsigned long ax, class bn, class co>
> + static void bz(aq<m, ax> z, bn *, co, ch<m> cx) {
> + using da = aq<bn, sizeof(bn)>;
> + using bu = typename da::bm;
> + using bp = aq<ae<1, bool, bn>, da::bd>;
> + auto cz = cv<bu>(z);
> + cq(da(cz), ae<0, int, co>(), au<bp>(cx));
> + }
> +};
> +struct ce : G {};
> +}
> +class D {
> +public:
> + D(...);
> + template <typename db> void operator<<(db);
> +};
> +template <class dc> dc dd;
> +struct de {
> + long cu();
> +};
> +template <typename dc> void dg() {
> + using db = long;
> + auto df = dd<a::b<long>>;
> + using a::ak;
> + using a::al;
> + constexpr long di = 1, alignment = a::bo<a::b<long>>;
> + using dh = ae<di, decltype(al), int>;
> + dh dk, am = a::am<alignment>;
> + const a::b<long> dj;
> + de u;
> + auto dm = 0 ? u.cu() : 0;
> + float dl[dm];
> + db reference;
> + a::b<long> x;
> + auto compare = [&](long) {
> + int n;
> + a::b<long>(reference, ak);
> + for (auto i = 0; 0; ++i)
> + [] {
> + auto v = 0, w = 0;
> + return D(w, v);
> + }() << n;
> + };
> + compare(0);
> + using c = a::b<long>::cd;
> + c dn;
> + a::b y = df;
> + auto v(y);
> + cc(dn, x).cb(dl, am);
> + long i;
> + cc(dn, dj).q(&dl[dc::cu()], dk);
> + ++i;
> +}
> +void test() { dg<a::b<long>>(); }
>
> Jakub
