[Bug tree-optimization/128434] New: BB SLP builds address vectors at function entry that are hoisted from rarely block inside of loop.

[hyeonseongyang09 at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128434

            Bug ID: 128434
           Summary: BB SLP builds address vectors at function entry that
                    are  hoisted from rarely block inside of loop.
           Product: gcc
           Version: 16.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hyeonseongyang09 at gmail dot com
  Target Milestone: ---

When a block inside a loop writes several constants into an array, BB SLP
combines the writes into vector stores.

The vectors code always hoist into start of the SLP region. so this moves them
out of the loop to the function entry.

But this makes fill the constants on every call, even if the block
runs only once or never.

I attach some simplified code of this issue.
Expected: the addresses are loaded in the OP_INIT case, where they are
stored.
Actual: they are loaded at the start of run(), before the loop, on every
call.
Testcase:

enum { OP_INC, OP_DOUBLE, OP_DEC, OP_HALT, OP_INIT };

void f_inc(void);
void f_dec(void);
void f_double(void);
void f_square(void);
void f_neg(void);
void f_half(void);
void f_print(void);
void f_halt(void);

void (*callbacks[8])(void);

long run(int *pc, long acc)
{
    for (;;) {
        switch (*pc++) {
        case OP_INC:    acc += 1; break;
        case OP_DOUBLE: acc *= 2; break;
        case OP_DEC:    acc -= 3; break;
        case OP_HALT:   return acc;
        case OP_INIT:
            callbacks[0] = f_inc;
            callbacks[1] = f_dec;
            callbacks[2] = f_double;
            callbacks[3] = f_square;
            callbacks[4] = f_neg;
            callbacks[5] = f_half;
            callbacks[6] = f_print;
            callbacks[7] = f_halt;
            break;
        }
    }
}

Start of run() before the loop:

run:
        movq    .LC0(%rip), %xmm3
        movq    %rsi, %rax
        movq    .LC2(%rip), %xmm2
        movhps  .LC1(%rip), %xmm3
        movq    .LC4(%rip), %xmm1
        movhps  .LC3(%rip), %xmm2
        movq    .LC6(%rip), %xmm0
        movhps  .LC5(%rip), %xmm1
        movhps  .LC7(%rip), %xmm0
.L2:
        cmpl    $4, (%rdi)
        ...

(.LC0 - .LC7 hold the addresses of f_inc ... f_halt.)

For comparison, this is code from Clang 22 -O2:

run:
        movq    %rsi, %rax
.L2:
        jmpq    *-8(%rdi)

Environment:

$ gcc -v
Target: x86_64-redhat-linux
Configured with: ../configure --enable-bootstrap
  --enable-languages=c,c++,fortran,objc,obj-c++,ada,go,d,m2,cobol,algol68,lto
  --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info
  --with-bugurl=https://bugzilla.redhat.com/ --enable-shared
  --enable-threads=posix --enable-checking=release --with-system-zlib
  --enable-__cxa_atexit --disable-libunwind-exceptions
  --enable-gnu-unique-object --enable-linker-build-id
  --with-gcc-major-version-only --enable-plugin --enable-initfini-array
  --enable-libstdcxx-backtrace
  --with-libstdcxx-zoneinfo=/usr/share/zoneinfo --with-isl=...
  --enable-gnu-indirect-function --enable-multilib
  --with-linker-hash-style=gnu
--enable-offload-targets=nvptx-none,amdgcn-amdhsa
  --enable-offload-defaulted --without-cuda-driver --enable-cet
  --with-tune=generic --with-tls=gnu2 --with-arch_32=i686
  --build=x86_64-redhat-linux --with-build-config=bootstrap-lto
  --enable-link-serialization=1 --disable-libssp
Thread model: posix
gcc version 16.2.1 20261029 (Red Hat 16.2.1-3) (GCC)


Command line:

$ gcc -O2 -S toy-interp.c -Wall -Wextra
No warnings.