[Bug middle-end/115916] New: [15 Regression] wrong code on highway-1.2.0

[slyfox at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115916

            Bug ID: 115916
           Summary: [15 Regression] wrong code on highway-1.2.0
           Product: gcc
           Version: 15.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: slyfox at gcc dot gnu.org
  Target Milestone: ---

Seemingly a recent regression.

On gcc-master r15-2026-g44c9403ed1833a highway-1.2.0 testsuite fails as:

The following tests FAILED:
        1190 -
HwyShuffle4TestGroup/HwyShuffle4Test.TestAllPer4LaneBlockShuffle/EMU128  #
GetParam() = 2305843009213693952 (Subprocess aborted)

Here is the minimized version extracted out of original example:

// $ cat shuffle4_test.cc
#include <stddef.h>
#include <stdint.h>

struct ve {
    ve() = default;
    ve(const ve&) = default;
    ve& operator=(const ve&) = default;

    // note that the code usually uses the first half of this array
    uint8_t raw[16] = {};
};

static ve First8_(void) {
    ve m;
    __builtin_memset(m.raw, 0xff, 8);
    return m;
}

static ve And_(ve a, ve b) {
    ve au;
    __builtin_memcpy(au.raw, a.raw, 16);
    for (size_t i = 0; i < 8; ++i) {
        au.raw[i] &= b.raw[i];
    }
    return au;
}

__attribute__((noipa, optimize(0)))
static void vec_assert(ve a) {
    if (a.raw[6] != 0x06 && a.raw[6] != 0x07)
        __builtin_trap();
}

static ve Reverse4_(ve v) {
    ve ret;
    for (size_t i = 0; i < 8; i += 4) {
        ret.raw[i + 0] = v.raw[i + 3];
        ret.raw[i + 1] = v.raw[i + 2];
        ret.raw[i + 2] = v.raw[i + 1];
        ret.raw[i + 3] = v.raw[i + 0];
    }
    return ret;
}

static ve DupEven_(ve v) {
    for (size_t i = 0; i < 8; i += 2) {
        v.raw[i + 1] = v.raw[i];
    }
    return v;
}

template <bool b>
ve Per4LaneBlockShuffle_(ve v) {
    if (b) {
        return Reverse4_(v);
    } else {
        return DupEven_(v);
    }
}

template <bool b>
static inline __attribute__((always_inline)) void
DoTestPer4LaneBlkShuffle(const ve v) {
    ve actual = Per4LaneBlockShuffle_<b>(v);
    const auto valid_lanes_mask = First8_();
    ve actual_masked = And_(valid_lanes_mask, actual);
    vec_assert(actual_masked);
}

static void DoTestPer4LaneBlkShuffles(const ve v) {
    alignas(128) uint8_t src_lanes[8];
    __builtin_memcpy(src_lanes, v.raw, 8);
    // need both, hm
    DoTestPer4LaneBlkShuffle<true >(v);
    DoTestPer4LaneBlkShuffle<false>(v);
}

__attribute__((noipa, optimize(0)))
static void bug(void) {
   uint8_t iv[8] = {1,2,3,4,5,6,7,8};
   ve v;
   __builtin_memcpy(v.raw, iv, 8);
   DoTestPer4LaneBlkShuffles(v);
}

int main(void) {
    bug();
}

Bad:

$ gcc/xg++ -Bgcc shuffle4_test.cc -O3 -o bug && ./bug
Illegal instruction (core dumped)

Ok:

$ gcc/xg++ -Bgcc shuffle4_test.cc -O2 -o bug && ./bug

$ gcc/xg++ -Bgcc -v
Reading specs from gcc/specs
COLLECT_GCC=gcc/xg++
COLLECT_LTO_WRAPPER=gcc/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: /home/slyfox/dev/git/gcc/configure --disable-multilib
--disable-bootstrap --disable-lto --disable-libsanitizer
--disable-libstdcxx-pch --enable-languages=c,c++ --disable-libgomp
--disable-libquadmath --disable-libvtv CFLAGS='-O1 -g0' CXXFLAGS='-O1 -g0'
LDFLAGS='-O1 -g0'
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 15.0.0 20240714 (experimental) (GCC)