[Bug target/110227] gcc generates invalid AVX-512 code

[joseph.weening at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110227

--- Comment #1 from Joe Weening  ---
Sorry, forgot to include the command line:

$ gcc -march=cooperlake -O3 -c bug.c

[Bug target/110227] New: gcc generates invalid AVX-512 code

[joseph.weening at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110227

Bug ID: 110227
   Summary: gcc generates invalid AVX-512 code
   Product: gcc
   Version: 13.1.0
Status: UNCONFIRMED
  Severity: normal
  Priority: P3
 Component: target
  Assignee: unassigned at gcc dot gnu.org
  Reporter: joseph.weening at gmail dot com
  Target Milestone: ---

gcc version 13.1.0 (GCC)
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-13.1.0/configure --prefix=/usr/local/gcc/13.1.0
--disable-multilib --enable-languages=c,c++,fortran
--with-gmp=/usr/local/gmp/6.2.1 --with-mpc=/usr/local/mpc/1.3.1
--with-mpfr=/usr/local/mpfr/4.2.0 --with-isl=/usr/local/isl/0.24

The following program generates an error:
  /tmp/ccAuHFqz.s: Assembler messages:
  /tmp/ccAuHFqz.s:28: Error: unsupported instruction `vpcmpeqd'
The assembly code contains
vpcmpeqd %xmm16, %xmm16, %xmm16
which perhaps is invalid for xmm registers above 15.

#include 

__attribute__((noinline))
static void vswap(int32_t *x) {
  __m256i x0 = _mm256_loadu_si256((__m256i *) ([0]));
  __m256i x1 = _mm256_loadu_si256((__m256i *) ([1]));
  _mm256_storeu_si256((__m256i *) ([0]),(x1));
  _mm256_storeu_si256((__m256i *) ([1]),(x0));
}

void vproc(int32_t *x) {
  for (int32_t p=4; p>=1; p>>=1) {
if (p == 4) {
  __m256i mask = _mm256_set_epi32(0, 0, 0, 0, -1, -1, -1, -1);
  __m256i x0 = _mm256_loadu_si256((__m256i *) ([0]));
  x0 = _mm256_xor_si256(x0, mask);
  _mm256_storeu_si256((__m256i *) ([0]),(x0));
}
vswap(x);
  }
}

[Bug target/91652] New: -march=skylake-avx512 -mno-fma -O2 generates FMA instructions

[joseph.weening at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91652

Bug ID: 91652
   Summary: -march=skylake-avx512 -mno-fma -O2 generates FMA
instructions
   Product: gcc
   Version: 9.2.0
Status: UNCONFIRMED
  Severity: normal
  Priority: P3
 Component: target
  Assignee: unassigned at gcc dot gnu.org
  Reporter: joseph.weening at gmail dot com
  Target Milestone: ---

Created attachment 46807
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46807=edit
program to demonstrate generation of FMA instructions

On x86_64, compiling C code with -march=skylake-avx512 -mno-fma -O2 generates
FMA instructions.  (-mno-fma works correctly with -march=skylake or
-march=haswell.)

The attached program demonstrates the problem.  It prints "d is 0.00" when
not using FMA, and "d is -1.00" when using FMA.

% gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/var/tmp/gcc/libexec/gcc/x86_64-pc-linux-gnu/9.2.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-9.2.0/configure --prefix=/var/tmp/gcc
--disable-multilib --enable-languages=c --with-gmp=/usr/local/gmp/6.1.2
--with-mpfr=/usr/local/mpfr/4.0.1 --with-mpc=/usr/local/mpc/1.1.0
Thread model: posix
gcc version 9.2.0 (GCC) 

Command to demonstrate the bug:
% gcc -march=skylake-avx512 -mno-fma -O2 fmabug.i && a.out
Compiler errors, warnings:
(none)