On Thu, Jun 20, 2024 at 09:06:11AM +0200, Juergen Christ wrote:
> Some casts were missing leading to missed of bad vectorizations where
> casting was done scalar followed by a vector creation from the
> individual elements.
>
> gcc/ChangeLog:
>
> * config/s390/vector.md (VEC_HALF_NARROWED): New mode iterator.
> (vec_half_narrowed): ditto.
> (trunc<VI_TRUNC:mode><vec_half_narrowed>2): New pattern.
> (vec_pack_ufix_trunc_v2df): ditto.
> (vec_pack_sfix_trunc_v2df): ditto.
> (vec_unpack_sfix_trunc_lo_v4sf): ditto.
> (vec_unpack_sfix_trunc_hi_v4sf): ditto.
> (vec_unpack_ufix_trunc_lo_v4sf): ditto.
> (vec_unpack_ufix_trunc_hi_v4sf): ditto.
> (floatv2siv2sf2): ditto.
> (floatunsv2siv2sf2): ditto.
> (vec_unpacks_float_hi_v4si): ditto.
> (vec_unpacks_float_lo_v4si): ditto.
> (vec_unpacku_float_hi_v4si): ditto.
> (vec_unpacku_float_lo_v4si): ditto.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/s390/vector/vec-cast-single.c: New test.
> * gcc.target/s390/vector/vec_pack_ufix_trunc_v2df.c: New test.
>
> Bootstrapped and regtested on s390x. Ok for trunk?
>
> Signed-off-by: Juergen Christ <jchr...@linux.ibm.com>
> ---
> gcc/config/s390/vector.md | 170 ++++++++++-
> .../gcc.target/s390/vector/vec-cast-single.c | 271 ++++++++++++++++++
> .../s390/vector/vec_pack_ufix_trunc_v2df.c | 30 ++
> 3 files changed, 463 insertions(+), 8 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/s390/vector/vec-cast-single.c
> create mode 100644
> gcc/testsuite/gcc.target/s390/vector/vec_pack_ufix_trunc_v2df.c
>
> diff --git a/gcc/config/s390/vector.md b/gcc/config/s390/vector.md
> index 40de0c75a7cf..356f25d26deb 100644
> --- a/gcc/config/s390/vector.md
> +++ b/gcc/config/s390/vector.md
> @@ -89,6 +89,8 @@
>
> (define_mode_iterator VI_EXTEND [V2QI V2HI V2SI V4QI V4HI])
>
> +(define_mode_iterator VI_TRUNC [V2HI V2SI V2DI V4HI V4SI])
> +
> ; Empty string for all but TImode. This is used to hide the TImode
> ; expander name in case it is defined already. See addti3 for an
> ; example.
> @@ -211,6 +213,14 @@
> (V1SF "v1df") (V2SF "v2df") (V4SF "v4df")
> (V1DF "v1tf") (V2DF "v2tf")])
>
> +; Vector with narrowed element size and the same number of elements.
> +(define_mode_attr VEC_HALF_NARROWED [(V1HI "V1QI") (V2HI "V2QI") (V4HI
> "V4QI") (V8HI "V8QI")
> + (V1SI "V1HI") (V2SI "V2HI") (V4SI "V4HI")
> + (V1DI "V1DI") (V2DI "V2SI")])
> +(define_mode_attr vec_half_narrowed [(V1HI "v1qi") (V2HI "v2qi") (V4HI
> "v4qi") (V8HI "v8qi")
> + (V1SI "v1hi") (V2SI "v2hi") (V4SI "v4hi")
> + (V1DI "v1di") (V2DI "v2si")])
> +
> ; Vector with half the element size AND half the number of elements.
> (define_mode_attr vec_halfhalf
> [(V2HI "V2QI") (V4HI "V4QI") (V8HI "V8QI")
> @@ -2422,6 +2432,17 @@
> operands[2] = gen_reg_rtx (V4SFmode);
> })
>
> +;; vector truncate
> +
> +; downcasts
> +
> +(define_insn "trunc<VI_TRUNC:mode><vec_half_narrowed>2"
> + [(set (match_operand:<VEC_HALF_NARROWED> 0 "register_operand" "=v")
> + (truncate:<VEC_HALF_NARROWED> (match_operand:VI_TRUNC 1
> "register_operand" "v")))]
> + "TARGET_VX"
> + "vpk<bhfgq>\t %0,%1,%1"
~~~~^~~~~
whitespace
> + [(set_attr "op_type" "VRR")])
> +
> ;; vector unpack v16qi
>
> ; signed
> @@ -3177,17 +3198,150 @@
> emit_move_insn (len, gen_rtx_ZERO_EXTEND (SImode, operands[2]));
> emit_insn (gen_vstlv16qi (operands[1], len, mem));
> DONE;
> -});;
> +})
> +
> +(define_expand "vec_pack_ufix_trunc_v2df"
> + [(match_operand:V4SI 0 "register_operand")
> + (match_operand:V2DF 1 "register_operand")
> + (match_operand:V2DF 2 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r1 = gen_reg_rtx (V2DImode);
> + rtx r2 = gen_reg_rtx (V2DImode);
> +
> + emit_insn (gen_fixuns_truncv2dfv2di2 (r1, operands[1]));
> + emit_insn (gen_fixuns_truncv2dfv2di2 (r2, operands[2]));
> + emit_insn (gen_vec_pack_trunc_v2di (operands[0], r1, r2));
> + DONE;
> +})
I haven't really wrapped my head around this, however, this two step
conversion could miss an IEEE-inexact-exception if a double fits into a
64-bit integer but not in a 32-bit integer. What does the IL/vectorizer
say about exceptions? Ok to miss some or do we have to guard this by
no-trapping-math et al.?
> +
> +(define_expand "vec_pack_sfix_trunc_v2df"
> + [(match_operand:V4SI 0 "register_operand")
> + (match_operand:V2DF 1 "register_operand")
> + (match_operand:V2DF 2 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r1 = gen_reg_rtx (V2DImode);
> + rtx r2 = gen_reg_rtx (V2DImode);
> +
> + emit_insn (gen_fix_truncv2dfv2di2 (r1, operands[1]));
> + emit_insn (gen_fix_truncv2dfv2di2 (r2, operands[2]));
> + emit_insn (gen_vec_pack_trunc_v2di (operands[0], r1, r2));
> + DONE;
> +})
same as above
> +
> +; v4sf -> v2di
> +(define_expand "vec_unpack_sfix_trunc_lo_v4sf"
> + [(match_operand:V2DI 0 "register_operand")
> + (match_operand:V4SF 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V4SImode);
~~~^~~~
whitespace
> +
> + emit_insn (gen_fix_truncv4sfv4si2 (r, operands[1]));
> + emit_insn (gen_vec_unpacks_lo_v4si (operands[0], r));
> + DONE;
> +})
The wording of the internals document leaves some room for
interpretation. When is widening supposed to happen? The point in time
when widening happens makes a difference if a rounded float does not fit
in a 32-bit integer but in a 64-bit integer. My gut feeling is that the
current implementation is correct, i.e., first converting a 32-bit float
into a 32-bit integer and then extending it to a 64-bit integer.
> +
> +(define_expand "vec_unpack_sfix_trunc_hi_v4sf"
> + [(match_operand:V2DI 0 "register_operand")
> + (match_operand:V4SF 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V4SImode);
~~~^~~~
whitespace
> +
> + emit_insn (gen_fix_truncv4sfv4si2 (r, operands[1]));
> + emit_insn (gen_vec_unpacks_hi_v4si (operands[0], r));
> + DONE;
> +})
same as above
> +
> +(define_expand "vec_unpack_ufix_trunc_lo_v4sf"
> + [(match_operand:V2DI 0 "register_operand")
> + (match_operand:V4SF 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V4SImode);
~~~^~~~
whitespace
> +
> + emit_insn (gen_fixuns_truncv4sfv4si2 (r, operands[1]));
> + emit_insn (gen_vec_unpacku_lo_v4si (operands[0], r));
> + DONE;
> +})
same as above
> +
> +(define_expand "vec_unpack_ufix_trunc_hi_v4sf"
> + [(match_operand:V2DI 0 "register_operand")
> + (match_operand:V4SF 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V4SImode);
~~~^~~~
whitespace
> +
> + emit_insn (gen_fixuns_truncv4sfv4si2 (r, operands[1]));
> + emit_insn (gen_vec_unpacku_hi_v4si (operands[0], r));
> + DONE;
> +})
same as above
>
> +(define_insn "floatv2siv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand" "=v")
> + (float:V2SF (match_operand:V2SI 1 "register_operand" "v")))]
> + "TARGET_VXE2"
> + "vcefb\t%v0,%v1,0,5"
> + [(set_attr "op_type" "VRR")])
> +
> +(define_insn "floatunsv2siv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand" "=v")
> + (unsigned_float:V2SF (match_operand:V2SI 1 "register_operand" "v")))]
> + "TARGET_VXE2"
> + "vcelfb\t%v0,%v1,0,5"
> + [(set_attr "op_type" "VRR")])
> +
> +(define_expand "vec_unpacks_float_hi_v4si"
> + [(match_operand:V2DF 0 "register_operand")
> + (match_operand:V4SI 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V2DImode);
~~~^~~~
whitespace
> +
~~^~~
trailing whitespace
> + emit_insn (gen_vec_unpacks_hi_v4si (r, operands[1]));
> + emit_insn (gen_floatv2div2df2 (operands[0], r));
> + DONE;
> +})
> +
> +(define_expand "vec_unpacks_float_lo_v4si"
> + [(match_operand:V2DF 0 "register_operand")
> + (match_operand:V4SI 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V2DImode);
~~~^~~~
whitespace
> +
~~^~~
trailing whitespace
> + emit_insn (gen_vec_unpacks_lo_v4si (r, operands[1]));
> + emit_insn (gen_floatv2div2df2 (operands[0], r));
> + DONE;
> +})
> +
> +(define_expand "vec_unpacku_float_hi_v4si"
> + [(match_operand:V2DF 0 "register_operand")
> + (match_operand:V4SI 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V2DImode);
~~~^~~~
whitespace
> +
~~^~~
trailing whitespace
> + emit_insn (gen_vec_unpacku_hi_v4si (r, operands[1]));
> + emit_insn (gen_floatunsv2div2df2 (operands[0], r));
> + DONE;
> +})
> +
> +(define_expand "vec_unpacku_float_lo_v4si"
> + [(match_operand:V2DF 0 "register_operand")
> + (match_operand:V4SI 1 "register_operand")]
> + "TARGET_VX"
> +{
> + rtx r = gen_reg_rtx(V2DImode);
~~~^~~~
whitespace
> +
~~^~~
trailing whitespace
> + emit_insn (gen_vec_unpacku_lo_v4si (r, operands[1]));
> + emit_insn (gen_floatunsv2div2df2 (operands[0], r));
> + DONE;
> +})
>
> ; reduc_smin
> ; reduc_smax
> ; reduc_umin
> ; reduc_umax
> -
> -; vec_pack_sfix_trunc: convert + pack ?
> -; vec_pack_ufix_trunc
> -; vec_unpacks_float_hi
> -; vec_unpacks_float_lo
> -; vec_unpacku_float_hi
> -; vec_unpacku_float_lo
> diff --git a/gcc/testsuite/gcc.target/s390/vector/vec-cast-single.c
> b/gcc/testsuite/gcc.target/s390/vector/vec-cast-single.c
> new file mode 100644
> index 000000000000..59a154594e9f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/s390/vector/vec-cast-single.c
> @@ -0,0 +1,271 @@
> +/* Check that the single-step vector conversions work. */
> +
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -mzarch -march=z15 -ftree-vectorize
> -fvect-cost-model=unlimited -fdump-tree-slp-all" } */
> +/* { dg-final { scan-tree-dump-not "conversion not supported by target"
> "slp" } } */
> +
> +void
> +extendv4hiv4si2 (short *in, int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +void
> +zero_extendv4hiv4si2 (unsigned short *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +void
> +vec_unpacks_v4si (int *in, long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +void
> +vec_unpacku_v4si (unsigned int *in, unsigned long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +extedv2siv2di2 (int *in, long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +zero_extedv2siv2di2 (unsigned int *in, unsigned long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +truncv4siv4hi2_signed (int *in, short *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +truncv4siv4hi2_unsigned (unsigned int *in, unsigned short *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +truncv2div2si2_signed (long *in, int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +truncv2div2si2_unsigned (unsigned long *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +fix_truncv4sfv4si2 (float *in, int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +fixuns_truncv4sfv4si2 (float *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_pack_trunc_v2di__signed (long *in, int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_pack_trunc_v2di__unsigned (unsigned long *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_pack_sfix_trunc_v2df (double *in, int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_pack_ufix_trunc_v2df (double *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_unpack_sfix_trunc (float *in, long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_unpack_ufix_trunc (float *in, unsigned long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +fix_truncv2dfv2di2 (double *in, long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +fixuns_truncv2dfv2di2 (double *in, unsigned long *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +floatv4hiv4sf2 (short *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +floatunsv4hiv4sf2 (unsigned short *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +floatv4siv4sf2 (int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +floatunsv4siv4sf2 (unsigned int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_packs_float_v2di (int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_packu_float_v2di (unsigned int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +floatv2div2df2 (long *in, double *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +floatunsv2div2df2 (unsigned long *in, double *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +
> +void
> +floatv2siv2sf2 (int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +
> +void
> +floatunsv2siv2sf2 (unsigned int *in, float *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> +}
> +void
> +vec_unpacks_float_hi_v4si (int *in, double *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +void
> +vec_unpacku_float_hi_v4si (unsigned int *in, double *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> diff --git a/gcc/testsuite/gcc.target/s390/vector/vec_pack_ufix_trunc_v2df.c
> b/gcc/testsuite/gcc.target/s390/vector/vec_pack_ufix_trunc_v2df.c
> new file mode 100644
> index 000000000000..4fcfbd88abe4
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/s390/vector/vec_pack_ufix_trunc_v2df.c
> @@ -0,0 +1,30 @@
> +/* Check that vec_pack_ufix_trunc_v2df pattern is correctly used. Even
> without
> + this pattern, we will vectorize this code, but produce wrong output. */
> +
> +/* { dg-do run } */
> +/* { dg-options "-O3 -mzarch -march=z13 -ftree-vectorize
> -fvect-cost-model=unlimited" } */
> +
> +__attribute__((noinline,noclone,noipa))
> +void
> +vec_pack_ufix_trunc_v2df (double *in, unsigned int *out);
> +
> +void
> +vec_pack_ufix_trunc_v2df (double *in, unsigned int *out)
> +{
> + out[0] = in[0];
> + out[1] = in[1];
> + out[2] = in[2];
> + out[3] = in[3];
> +}
> +
> +int main()
> +{
> + double in[] = {-1,-2,-3,-4};
> + unsigned int out[4];
> +
> + vec_pack_ufix_trunc_v2df (in, out);
> + for (int i = 0; i < 4; ++i)
> + if (out[i] != 0)
> + __builtin_abort();
> + return 0;
> +}
> --
> 2.43.0
>
