Committed, thanks Juzhe. Pan
From: 钟居哲 <juzhe.zh...@rivai.ai> Sent: Sunday, September 24, 2023 2:06 PM To: Li, Pan2 <pan2...@intel.com>; gcc-patches <gcc-patches@gcc.gnu.org> Cc: Li, Pan2 <pan2...@intel.com>; Wang, Yanzhang <yanzhang.w...@intel.com>; kito.cheng <kito.ch...@gmail.com>; patrick <patr...@rivosinc.com> Subject: Re: [PATCH v2] RISC-V: Fix fortran ICE/PR111546 when RV32 vec_init LGTM ________________________________ juzhe.zh...@rivai.ai<mailto:juzhe.zh...@rivai.ai> From: pan2.li<mailto:pan2...@intel.com> Date: 2023-09-24 13:50 To: gcc-patches<mailto:gcc-patches@gcc.gnu.org> CC: juzhe.zhong<mailto:juzhe.zh...@rivai.ai>; pan2.li<mailto:pan2...@intel.com>; yanzhang.wang<mailto:yanzhang.w...@intel.com>; kito.cheng<mailto:kito.ch...@gmail.com>; patrick<mailto:patr...@rivosinc.com> Subject: [PATCH v2] RISC-V: Fix fortran ICE/PR111546 when RV32 vec_init From: Pan Li <pan2...@intel.com<mailto:pan2...@intel.com>> When broadcast the reperated element, we take the mask_int_mode by mistake. This patch would like to fix it by leveraging the machine mode of the element. The below test case in RV32 will be fixed. * gcc/testsuite/gfortran.dg/overload_5.f90 PR target/111546 gcc/ChangeLog: * config/riscv/riscv-v.cc (expand_vector_init_merge_repeating_sequence): Bugfix Signed-off-by: Pan Li <pan2...@intel.com<mailto:pan2...@intel.com>> --- gcc/config/riscv/riscv-v.cc | 22 ++++++++++++++-------- 1 file changed, 14 insertions(+), 8 deletions(-) diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc index c2466b1354f..a1ffefb23f3 100644 --- a/gcc/config/riscv/riscv-v.cc +++ b/gcc/config/riscv/riscv-v.cc @@ -2054,12 +2054,17 @@ static void expand_vector_init_merge_repeating_sequence (rtx target, const rvv_builder &builder) { - machine_mode dup_mode = get_repeating_sequence_dup_machine_mode (builder); - machine_mode mask_mode = get_mask_mode (builder.mode ()); + /* We can't use BIT mode (BI) directly to generate mask = 0b01010... + since we don't have such instruction in RVV. + Instead, we should use INT mode (QI/HI/SI/DI) with integer move + instruction to generate the mask data we want. */ + machine_mode mask_int_mode + = get_repeating_sequence_dup_machine_mode (builder); + machine_mode mask_bit_mode = get_mask_mode (builder.mode ()); uint64_t full_nelts = builder.full_nelts ().to_constant (); /* Step 1: Broadcast the first pattern. */ - rtx ops[] = {target, force_reg (GET_MODE_INNER (dup_mode), builder.elt (0))}; + rtx ops[] = {target, force_reg (builder.inner_mode (), builder.elt (0))}; emit_vlmax_insn (code_for_pred_broadcast (builder.mode ()), UNARY_OP, ops); /* Step 2: Merge the rest iteration of pattern. */ @@ -2067,8 +2072,8 @@ expand_vector_init_merge_repeating_sequence (rtx target, { /* Step 2-1: Generate mask register v0 for each merge. */ rtx merge_mask = builder.get_merge_scalar_mask (i); - rtx mask = gen_reg_rtx (mask_mode); - rtx dup = gen_reg_rtx (dup_mode); + rtx mask = gen_reg_rtx (mask_bit_mode); + rtx dup = gen_reg_rtx (mask_int_mode); if (full_nelts <= builder.inner_bits_size ()) /* vmv.s.x. */ { @@ -2078,14 +2083,15 @@ expand_vector_init_merge_repeating_sequence (rtx target, } else /* vmv.v.x. */ { - rtx ops[] = {dup, force_reg (GET_MODE_INNER (dup_mode), merge_mask)}; + rtx ops[] = {dup, + force_reg (GET_MODE_INNER (mask_int_mode), merge_mask)}; rtx vl = gen_int_mode (CEIL (full_nelts, builder.inner_bits_size ()), Pmode); - emit_nonvlmax_insn (code_for_pred_broadcast (dup_mode), UNARY_OP, + emit_nonvlmax_insn (code_for_pred_broadcast (mask_int_mode), UNARY_OP, ops, vl); } - emit_move_insn (mask, gen_lowpart (mask_mode, dup)); + emit_move_insn (mask, gen_lowpart (mask_bit_mode, dup)); /* Step 2-2: Merge pattern according to the mask. */ rtx ops[] = {target, target, builder.elt (i), mask}; -- 2.34.1
