This part LGTM. juzhe.zhong@rivai.ai From: Robin Dapp Date: 2023-05-10 23:24 To: gcc-patches; juzhe.zhong@rivai.ai; Kito Cheng; Michael Collison; palmer; jeffreyalaw CC: rdapp.gcc Subject: [PATCH] riscv: Clarify vlmax and length handling. Hi, this patch tries to improve the wrappers that emit either vlmax or non-vlmax operations. Now, emit_len_op can be used to emit a regular operation. Depending on whether a length != NULL is passed either no VLMAX flags are set or we emit a vsetvli and set VLMAX flags. The patch also adds some comments that describes some of the rationale of the current handling of vlmax/nonvlmax operations. Bootstrapped and regtested. Regards Robin -- gcc/ChangeLog: * config/riscv/autovec.md: Use renamed functions. * config/riscv/riscv-protos.h (emit_vlmax_op): Rename. (emit_vlmax_reg_op): To this. (emit_nonvlmax_op): Rename. (emit_len_op): To this. (emit_nonvlmax_binop): Rename. (emit_len_binop): To this. * config/riscv/riscv-v.cc (emit_pred_op): Add default parameter. (emit_pred_binop): Remove vlmax_p. (emit_vlmax_op): Rename. (emit_vlmax_reg_op): To this. (emit_nonvlmax_op): Rename. (emit_len_op): To this. (emit_nonvlmax_binop): Rename. (emit_len_binop): To this. (sew64_scalar_helper): Use renamed functions. (expand_tuple_move): Use renamed functions. * config/riscv/riscv.cc (vector_zero_call_used_regs): Use renamed functions. * config/riscv/vector.md: Use renamed functions. --- gcc/config/riscv/autovec.md | 24 +++++----- gcc/config/riscv/riscv-protos.h | 8 ++-- gcc/config/riscv/riscv-v.cc | 82 ++++++++++++++++++++------------- gcc/config/riscv/riscv.cc | 4 +- gcc/config/riscv/vector.md | 12 +++-- 5 files changed, 75 insertions(+), 55 deletions(-) diff --git a/gcc/config/riscv/autovec.md b/gcc/config/riscv/autovec.md index 15f8d007e07..8347e42bb9c 100644 --- a/gcc/config/riscv/autovec.md +++ b/gcc/config/riscv/autovec.md @@ -31,8 +31,8 @@ (define_expand "len_load_" (match_operand 3 "const_0_operand")] "TARGET_VECTOR" { - riscv_vector::emit_nonvlmax_op (code_for_pred_mov (mode), operands[0], - operands[1], operands[2], mode); + riscv_vector::emit_len_op (code_for_pred_mov (mode), operands[0], + operands[1], operands[2], mode); DONE; }) @@ -43,8 +43,8 @@ (define_expand "len_store_" (match_operand 3 "const_0_operand")] "TARGET_VECTOR" { - riscv_vector::emit_nonvlmax_op (code_for_pred_mov (mode), operands[0], - operands[1], operands[2], mode); + riscv_vector::emit_len_op (code_for_pred_mov (mode), operands[0], + operands[1], operands[2], mode); DONE; }) @@ -79,15 +79,15 @@ (define_expand "3" if (inner == E_QImode || inner == E_HImode || inner == E_SImode) op2mode = inner; - riscv_vector::emit_nonvlmax_binop (code_for_pred_scalar - (, mode), - operands[0], operands[1], cst, - NULL_RTX, mode, op2mode); + riscv_vector::emit_len_binop (code_for_pred_scalar + (, mode), + operands[0], operands[1], cst, + NULL_RTX, mode, op2mode); } else - riscv_vector::emit_nonvlmax_binop (code_for_pred - (, mode), - operands[0], operands[1], operands[2], - NULL_RTX, mode); + riscv_vector::emit_len_binop (code_for_pred + (, mode), + operands[0], operands[1], operands[2], + NULL_RTX, mode); DONE; }) diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h index 75cdb90b9c9..bfdf09b17ee 100644 --- a/gcc/config/riscv/riscv-protos.h +++ b/gcc/config/riscv/riscv-protos.h @@ -167,10 +167,10 @@ bool legitimize_move (rtx, rtx, machine_mode); void emit_vlmax_vsetvl (machine_mode, rtx); void emit_hard_vlmax_vsetvl (machine_mode, rtx); void emit_vlmax_op (unsigned, rtx, rtx, machine_mode); -void emit_vlmax_op (unsigned, rtx, rtx, rtx, machine_mode); -void emit_nonvlmax_op (unsigned, rtx, rtx, rtx, machine_mode); -void emit_nonvlmax_binop (unsigned, rtx, rtx, rtx, rtx, machine_mode, - machine_mode = VOIDmode); +void emit_vlmax_reg_op (unsigned, rtx, rtx, rtx, machine_mode); +void emit_len_op (unsigned, rtx, rtx, rtx, machine_mode); +void emit_len_binop (unsigned, rtx, rtx, rtx, rtx, machine_mode, machine_mode = + VOIDmode); enum vlmul_type get_vlmul (machine_mode); unsigned int get_ratio (machine_mode); unsigned int get_nf (machine_mode); diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc index 3c43dfc5eea..07b7783282f 100644 --- a/gcc/config/riscv/riscv-v.cc +++ b/gcc/config/riscv/riscv-v.cc @@ -99,27 +99,24 @@ public: add_vundef_operand (dest_mode); } - void set_len_and_policy (rtx len, bool vlmax_p) + void set_len_and_policy (rtx len, bool force_vlmax = false) { + bool vlmax_p = force_vlmax; gcc_assert (has_dest); - gcc_assert (len || vlmax_p); - if (len) - add_input_operand (len, Pmode); - else + if (!len) { - rtx vlmax = gen_reg_rtx (Pmode); - emit_vlmax_vsetvl (dest_mode, vlmax); - add_input_operand (vlmax, Pmode); + vlmax_p = true; + len = gen_reg_rtx (Pmode); + emit_vlmax_vsetvl (dest_mode, len); } + add_input_operand (len, Pmode); + if (GET_MODE_CLASS (dest_mode) != MODE_VECTOR_BOOL) add_policy_operand (get_prefer_tail_policy (), get_prefer_mask_policy ()); - if (vlmax_p) - add_avl_type_operand (avl_type::VLMAX); - else - add_avl_type_operand (avl_type::NONVLMAX); + add_avl_type_operand (vlmax_p ? avl_type::VLMAX : avl_type::NONVLMAX); } void expand (enum insn_code icode, bool temporary_volatile_p = false) @@ -251,14 +248,14 @@ set_expander_dest_and_mask (insn_expander &e, rtx mask, rtx dest, /* Emit an RVV unmask && vl mov from SRC to DEST. */ static void emit_pred_op (unsigned icode, rtx mask, rtx dest, rtx src, rtx len, - machine_mode mask_mode, bool vlmax_p) + machine_mode mask_mode, bool force_vlmax = false) { insn_expander<8> e; e.set_dest_and_mask (mask, dest, mask_mode); e.add_input_operand (src, GET_MODE (src)); - e.set_len_and_policy (len, vlmax_p); + e.set_len_and_policy (len, force_vlmax); e.expand ((enum insn_code) icode, MEM_P (dest) || MEM_P (src)); } @@ -266,8 +263,7 @@ emit_pred_op (unsigned icode, rtx mask, rtx dest, rtx src, rtx len, /* Emit an RVV unmask && vl mov from SRC to DEST. */ static void emit_pred_binop (unsigned icode, rtx mask, rtx dest, rtx src1, rtx src2, - rtx len, machine_mode mask_mode, machine_mode op2mode, - bool vlmax_p) + rtx len, machine_mode mask_mode, machine_mode op2mode) { insn_expander<9> e; e.set_dest_and_mask (mask, dest, mask_mode); @@ -275,37 +271,59 @@ emit_pred_binop (unsigned icode, rtx mask, rtx dest, rtx src1, rtx src2, e.add_input_operand (src1, GET_MODE (src1)); e.add_input_operand (src2, op2mode == VOIDmode ? GET_MODE (src2) : op2mode); - e.set_len_and_policy (len, vlmax_p); + e.set_len_and_policy (len); e.expand ((enum insn_code) icode, MEM_P (dest) || MEM_P (src1) || MEM_P (src2)); } +/* The RISC-V vsetvli pass uses vlmax operation for optimization. Whether or + not an instruction actually is a vlmax operation is not recognizable from + the length operand alone but the avl_type operand is used instead. + In general, there are two cases: + + - Emit a vlmax operation by passing a NULL length. Here we emit + a vsetvli with vlmax configuration and set the avl_type to VLMAX. + - Emit an operation that uses the existing (last-set) length and + set the avl_type to NONVLMAX. + + Sometimes we also need to set the VLMAX avl_type to an operation that + already uses a given length register. This can happen during or after + register allocation when we are not allowed to create a new register. + For that case we also allow to set the avl_type to VLMAX. +*/ + +/* This function emits a VLMAX vsetvli followed by the actual operation. */ void emit_vlmax_op (unsigned icode, rtx dest, rtx src, machine_mode mask_mode) { - emit_pred_op (icode, NULL_RTX, dest, src, NULL_RTX, mask_mode, true); + emit_pred_op (icode, NULL_RTX, dest, src, NULL_RTX, mask_mode); } +/* This function emits an operation with a given LEN that is determined + by a previously emitted VLMAX vsetvli. */ void -emit_vlmax_op (unsigned icode, rtx dest, rtx src, rtx len, - machine_mode mask_mode) +emit_len_op (unsigned icode, rtx dest, rtx src, rtx len, + machine_mode mask_mode) { - emit_pred_op (icode, NULL_RTX, dest, src, len, mask_mode, true); + emit_pred_op (icode, NULL_RTX, dest, src, len, mask_mode); } +/* This function emits an operation with a given LEN that is known to be + a preceding VLMAX. It also sets the VLMAX flag which allows further + optimization in the vsetvli pass. */ void -emit_nonvlmax_op (unsigned icode, rtx dest, rtx src, rtx len, - machine_mode mask_mode) +emit_vlmax_reg_op (unsigned icode, rtx dest, rtx src, rtx len, + machine_mode mask_mode) { - emit_pred_op (icode, NULL_RTX, dest, src, len, mask_mode, false); + emit_pred_op (icode, NULL_RTX, dest, src, len, mask_mode, + /* Force VLMAX */ true); } void -emit_nonvlmax_binop (unsigned icode, rtx dest, rtx src1, rtx src2, rtx len, - machine_mode mask_mode, machine_mode op2mode) +emit_len_binop (unsigned icode, rtx dest, rtx src1, rtx src2, rtx len, + machine_mode mask_mode, machine_mode op2mode) { - emit_pred_binop (icode, NULL_RTX, dest, src1, src2, len, mask_mode, op2mode, - false); + emit_pred_binop (icode, NULL_RTX, dest, src1, src2, len, mask_mode, op2mode); } static void @@ -667,8 +685,8 @@ sew64_scalar_helper (rtx *operands, rtx *scalar_op, rtx vl, *scalar_op = force_reg (scalar_mode, *scalar_op); rtx tmp = gen_reg_rtx (vector_mode); - riscv_vector::emit_nonvlmax_op (code_for_pred_broadcast (vector_mode), tmp, - *scalar_op, vl, mask_mode); + riscv_vector::emit_len_op (code_for_pred_broadcast (vector_mode), tmp, + *scalar_op, vl, mask_mode); emit_vector_func (operands, tmp); return true; @@ -974,7 +992,7 @@ expand_tuple_move (machine_mode mask_mode, rtx *ops) rtx mem = gen_rtx_MEM (subpart_mode, ops[3]); if (fractional_p) - emit_vlmax_op (code_for_pred_mov (subpart_mode), subreg, mem, + emit_vlmax_reg_op (code_for_pred_mov (subpart_mode), subreg, mem, ops[4], mask_mode); else emit_move_insn (subreg, mem); @@ -996,7 +1014,7 @@ expand_tuple_move (machine_mode mask_mode, rtx *ops) rtx mem = gen_rtx_MEM (subpart_mode, ops[3]); if (fractional_p) - emit_vlmax_op (code_for_pred_mov (subpart_mode), mem, subreg, + emit_vlmax_reg_op (code_for_pred_mov (subpart_mode), mem, subreg, ops[4], mask_mode); else emit_move_insn (mem, subreg); diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index 9e08a1bfe35..9a493295a38 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -7373,8 +7373,8 @@ vector_zero_call_used_regs (HARD_REG_SET need_zeroed_hardregs) emitted_vlmax_vsetvl = true; } - riscv_vector::emit_vlmax_op (code_for_pred_mov (mode), target, - CONST0_RTX (mode), vl, mask_mode); + riscv_vector::emit_vlmax_reg_op (code_for_pred_mov (mode), target, + CONST0_RTX (mode), vl, mask_mode); SET_HARD_REG_BIT (zeroed_hardregs, regno); } diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md index 9ccc0d6a513..75479f27bcd 100644 --- a/gcc/config/riscv/vector.md +++ b/gcc/config/riscv/vector.md @@ -721,8 +721,9 @@ (define_insn_and_split "*mov_lra" else { riscv_vector::emit_vlmax_vsetvl (mode, operands[2]); - riscv_vector::emit_vlmax_op (code_for_pred_mov (mode), - operands[0], operands[1], operands[2], mode); + riscv_vector::emit_vlmax_reg_op (code_for_pred_mov (mode), + operands[0], operands[1], operands[2], + mode); } DONE; }) @@ -741,8 +742,9 @@ (define_insn_and_split "*mov_lra" else { riscv_vector::emit_vlmax_vsetvl (mode, operands[2]); - riscv_vector::emit_vlmax_op (code_for_pred_mov (mode), - operands[0], operands[1], operands[2], mode); + riscv_vector::emit_vlmax_reg_op (code_for_pred_mov (mode), + operands[0], operands[1], operands[2], + mode); } DONE; }) @@ -806,7 +808,7 @@ (define_expand "vec_duplicate" "TARGET_VECTOR" { riscv_vector::emit_vlmax_op (code_for_pred_broadcast (mode), - operands[0], operands[1], mode); + operands[0], operands[1], mode); DONE; } ) -- 2.40.0