| Message ID | 20230620161103.2964705-1-hongtao.liu@intel.com |
|---|---|
| State | New |
| Headers | show |
| Series | [vect] Use intermiediate integer type for float_expr/fix_trunc_expr when direct optab is not existed. | expand |
On Tue, Jun 20, 2023 at 6:11 PM liuhongt via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > I notice there's some refactor in vectorizable_conversion > for code_helper,so I've adjusted my patch to that. > Here's the patch I'm going to commit. > > We have already use intermidate type in case WIDEN, but not for NONE, > this patch extended that. > > gcc/ChangeLog: > > PR target/110018 > * tree-vect-stmts.cc (vectorizable_conversion): Use > intermiediate integer type for float_expr/fix_trunc_expr when > direct optab is not existed. > > gcc/testsuite/ChangeLog: > > * gcc.target/i386/pr110018-1.c: New test. > + > + /* For conversions between float and smaller integer types try whether we > + can use intermediate signed integer types to support the > + conversion. */ I'm trying to enhance testcase coverage with explicit signed/unsigned types (patch attached), and I have noticed that zero-extension is used for unsigned types. So, the above comment that mentions only signed integer types is not entirely correct. Uros. diff --git a/gcc/testsuite/gcc.target/i386/pr110018-1.c b/gcc/testsuite/gcc.target/i386/pr110018-1.c index b1baffd7af1..b6a3be7b7a2 100644 --- a/gcc/testsuite/gcc.target/i386/pr110018-1.c +++ b/gcc/testsuite/gcc.target/i386/pr110018-1.c @@ -4,14 +4,14 @@ /* { dg-final { scan-assembler-times {(?n)vcvt[dqw]*2p[dsh]} 5 } } */ void -foo (double* __restrict a, char* b) +foo (double* __restrict a, signed char* b) { a[0] = b[0]; a[1] = b[1]; } void -foo1 (float* __restrict a, char* b) +foo1 (float* __restrict a, signed char* b) { a[0] = b[0]; a[1] = b[1]; @@ -20,7 +20,7 @@ foo1 (float* __restrict a, char* b) } void -foo2 (_Float16* __restrict a, char* b) +foo2 (_Float16* __restrict a, signed char* b) { a[0] = b[0]; a[1] = b[1]; @@ -33,14 +33,14 @@ foo2 (_Float16* __restrict a, char* b) } void -foo3 (double* __restrict a, short* b) +foo3 (double* __restrict a, signed short* b) { a[0] = b[0]; a[1] = b[1]; } void -foo4 (float* __restrict a, char* b) +foo4 (float* __restrict a, signed char* b) { a[0] = b[0]; a[1] = b[1]; @@ -49,14 +49,14 @@ foo4 (float* __restrict a, char* b) } void -foo5 (double* __restrict b, char* a) +foo5 (double* __restrict b, signed char* a) { a[0] = b[0]; a[1] = b[1]; } void -foo6 (float* __restrict b, char* a) +foo6 (float* __restrict b, signed char* a) { a[0] = b[0]; a[1] = b[1]; @@ -65,7 +65,7 @@ foo6 (float* __restrict b, char* a) } void -foo7 (_Float16* __restrict b, char* a) +foo7 (_Float16* __restrict b, signed char* a) { a[0] = b[0]; a[1] = b[1]; @@ -78,14 +78,14 @@ foo7 (_Float16* __restrict b, char* a) } void -foo8 (double* __restrict b, short* a) +foo8 (double* __restrict b, signed short* a) { a[0] = b[0]; a[1] = b[1]; } void -foo9 (float* __restrict b, char* a) +foo9 (float* __restrict b, signed char* a) { a[0] = b[0]; a[1] = b[1]; diff --git a/gcc/testsuite/gcc.target/i386/pr110018-2.c b/gcc/testsuite/gcc.target/i386/pr110018-2.c new file mode 100644 index 00000000000..a663e074698 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr110018-2.c @@ -0,0 +1,94 @@ +/* { dg-do compile } */ +/* { dg-options "-mavx512fp16 -mavx512vl -O2 -mavx512dq" } */ +/* { dg-final { scan-assembler-times {(?n)vcvttp[dsh]2[dqw]} 5 } } */ +/* { dg-final { scan-assembler-times {(?n)vcvt[dqw]*2p[dsh]} 5 } } */ + +void +foo (double* __restrict a, unsigned char* b) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo1 (float* __restrict a, unsigned char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo2 (_Float16* __restrict a, unsigned char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; + a[4] = b[4]; + a[5] = b[5]; + a[6] = b[6]; + a[7] = b[7]; +} + +void +foo3 (double* __restrict a, unsigned short* b) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo4 (float* __restrict a, unsigned char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo5 (double* __restrict b, unsigned char* a) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo6 (float* __restrict b, unsigned char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo7 (_Float16* __restrict b, unsigned char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; + a[4] = b[4]; + a[5] = b[5]; + a[6] = b[6]; + a[7] = b[7]; +} + +void +foo8 (double* __restrict b, unsigned short* a) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo9 (float* __restrict b, unsigned char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +}
On Wed, Jun 21, 2023 at 9:50 AM Uros Bizjak via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > On Tue, Jun 20, 2023 at 6:11 PM liuhongt via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: > > > > I notice there's some refactor in vectorizable_conversion > > for code_helper,so I've adjusted my patch to that. > > Here's the patch I'm going to commit. > > > > We have already use intermidate type in case WIDEN, but not for NONE, > > this patch extended that. > > > > gcc/ChangeLog: > > > > PR target/110018 > > * tree-vect-stmts.cc (vectorizable_conversion): Use > > intermiediate integer type for float_expr/fix_trunc_expr when > > direct optab is not existed. > > > > gcc/testsuite/ChangeLog: > > > > * gcc.target/i386/pr110018-1.c: New test. > > > + > > + /* For conversions between float and smaller integer types try whether we > > + can use intermediate signed integer types to support the > > + conversion. */ > > I'm trying to enhance testcase coverage with explicit signed/unsigned > types (patch attached), and I have noticed that zero-extension is used > for unsigned types. So, the above comment that mentions only signed > integer types is not entirely correct. The comment says the intermediate sized vector types are always signed (because float conversions to/from unsigned are always somewhat awkward), but yes, if the original type was unsigned zero-extension is used and if it was signed sign-extension. The testcase adjustments / additions look good to me btw. Thanks, Richard. > > Uros.
On Wed, Jun 21, 2023 at 10:22 AM Richard Biener <richard.guenther@gmail.com> wrote: > > > + /* For conversions between float and smaller integer types try whether we > > > + can use intermediate signed integer types to support the > > > + conversion. */ > > > > I'm trying to enhance testcase coverage with explicit signed/unsigned > > types (patch attached), and I have noticed that zero-extension is used > > for unsigned types. So, the above comment that mentions only signed > > integer types is not entirely correct. > > The comment says the intermediate sized vector types are always > signed (because float conversions to/from unsigned are always somewhat > awkward), but yes, if the original type was unsigned zero-extension is > used and if it was signed sign-extension. > > The testcase adjustments / additions look good to me btw. Thanks, pushed with the following ChangeLog: vect: Add testcases for unsigned conversions [PR110018] Also test conversions with unsigned types. PR target/110018 gcc/testsuite/ChangeLog: * gcc.target/i386/pr110018-1.c: Use explicit signed types. * gcc.target/i386/pr110018-2.c: New test. Tested on x86_64-linux-gnu {,-m32}. Uros.
Hello, liuhongt via Gcc-patches <gcc-patches@gcc.gnu.org> writes: > I notice there's some refactor in vectorizable_conversion > for code_helper,so I've adjusted my patch to that. > Here's the patch I'm going to commit. > > We have already use intermidate type in case WIDEN, but not for NONE, > this patch extended that. > > gcc/ChangeLog: > > PR target/110018 > * tree-vect-stmts.cc (vectorizable_conversion): Use > intermiediate integer type for float_expr/fix_trunc_expr when > direct optab is not existed. > > gcc/testsuite/ChangeLog: > > * gcc.target/i386/pr110018-1.c: New test. > --- > gcc/testsuite/gcc.target/i386/pr110018-1.c | 94 ++++++++++++++++++++++ > gcc/tree-vect-stmts.cc | 66 ++++++++++++++- > 2 files changed, 158 insertions(+), 2 deletions(-) > create mode 100644 gcc/testsuite/gcc.target/i386/pr110018-1.c Our CI detected regressions on aarch64-linux-gnu with this commit, in some aarch64-sve and gfortran tests. I filed the following bug report with the details: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110371 Could you please check?
diff --git a/gcc/testsuite/gcc.target/i386/pr110018-1.c b/gcc/testsuite/gcc.target/i386/pr110018-1.c new file mode 100644 index 00000000000..b1baffd7af1 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr110018-1.c @@ -0,0 +1,94 @@ +/* { dg-do compile } */ +/* { dg-options "-mavx512fp16 -mavx512vl -O2 -mavx512dq" } */ +/* { dg-final { scan-assembler-times {(?n)vcvttp[dsh]2[dqw]} 5 } } */ +/* { dg-final { scan-assembler-times {(?n)vcvt[dqw]*2p[dsh]} 5 } } */ + +void +foo (double* __restrict a, char* b) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo1 (float* __restrict a, char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo2 (_Float16* __restrict a, char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; + a[4] = b[4]; + a[5] = b[5]; + a[6] = b[6]; + a[7] = b[7]; +} + +void +foo3 (double* __restrict a, short* b) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo4 (float* __restrict a, char* b) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo5 (double* __restrict b, char* a) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo6 (float* __restrict b, char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} + +void +foo7 (_Float16* __restrict b, char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; + a[4] = b[4]; + a[5] = b[5]; + a[6] = b[6]; + a[7] = b[7]; +} + +void +foo8 (double* __restrict b, short* a) +{ + a[0] = b[0]; + a[1] = b[1]; +} + +void +foo9 (float* __restrict b, char* a) +{ + a[0] = b[0]; + a[1] = b[1]; + a[2] = b[2]; + a[3] = b[3]; +} diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc index 056a0ecb2be..ae24f3e66e6 100644 --- a/gcc/tree-vect-stmts.cc +++ b/gcc/tree-vect-stmts.cc @@ -5041,7 +5041,7 @@ vectorizable_conversion (vec_info *vinfo, tree scalar_dest; tree op0, op1 = NULL_TREE; loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo); - tree_code tc1; + tree_code tc1, tc2; code_helper code, code1, code2; code_helper codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; tree new_temp; @@ -5249,6 +5249,57 @@ vectorizable_conversion (vec_info *vinfo, code1 = tc1; break; } + + /* For conversions between float and smaller integer types try whether we + can use intermediate signed integer types to support the + conversion. */ + if ((code == FLOAT_EXPR + && GET_MODE_SIZE (lhs_mode) > GET_MODE_SIZE (rhs_mode)) + || (code == FIX_TRUNC_EXPR + && GET_MODE_SIZE (rhs_mode) > GET_MODE_SIZE (lhs_mode))) + { + bool float_expr_p = code == FLOAT_EXPR; + scalar_mode imode = float_expr_p ? rhs_mode : lhs_mode; + fltsz = GET_MODE_SIZE (float_expr_p ? lhs_mode : rhs_mode); + code1 = float_expr_p ? code : NOP_EXPR; + codecvt1 = float_expr_p ? NOP_EXPR : code; + FOR_EACH_2XWIDER_MODE (rhs_mode_iter, imode) + { + imode = rhs_mode_iter.require (); + if (GET_MODE_SIZE (imode) > fltsz) + break; + + cvt_type + = build_nonstandard_integer_type (GET_MODE_BITSIZE (imode), + 0); + cvt_type = get_vectype_for_scalar_type (vinfo, cvt_type, + slp_node); + /* This should only happened for SLP as long as loop vectorizer + only supports same-sized vector. */ + if (cvt_type == NULL_TREE + || maybe_ne (TYPE_VECTOR_SUBPARTS (cvt_type), nunits_in) + || !supportable_convert_operation ((tree_code) code1, + vectype_out, + cvt_type, &tc1) + || !supportable_convert_operation ((tree_code) codecvt1, + cvt_type, + vectype_in, &tc2)) + continue; + + found_mode = true; + break; + } + + if (found_mode) + { + multi_step_cvt++; + interm_types.safe_push (cvt_type); + cvt_type = NULL_TREE; + code1 = tc1; + codecvt1 = tc2; + break; + } + } /* FALLTHRU */ unsupported: if (dump_enabled_p ()) @@ -5513,7 +5564,18 @@ vectorizable_conversion (vec_info *vinfo, FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) { /* Arguments are ready, create the new vector stmt. */ - gimple *new_stmt = vect_gimple_build (vec_dest, code1, vop0); + gimple* new_stmt; + if (multi_step_cvt) + { + gcc_assert (multi_step_cvt == 1); + new_stmt = vect_gimple_build (vec_dest, codecvt1, vop0); + new_temp = make_ssa_name (vec_dest, new_stmt); + gimple_assign_set_lhs (new_stmt, new_temp); + vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); + vop0 = new_temp; + vec_dest = vec_dsts[0]; + } + new_stmt = vect_gimple_build (vec_dest, code1, vop0); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_set_lhs (new_stmt, new_temp); vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);