From: Juzhe-Zhong <juzhe.zh...@rivai.ai> This patch is the prerequiste patch for more RVV auto-vectorization support.
Since when we enable a very simple binary operations, we will end up with such following ICE: during RTL pass: expand add_run-1.c: In function 'main': add_run-1.c:28:1: internal compiler error: Segmentation fault 0x1618ea3 crash_signal ../../../riscv-gcc/gcc/toplev.cc:314 0xe76cd9 single_set(rtx_insn const*) ../../../riscv-gcc/gcc/rtl.h:3602 0x1080f8a emit_move_insn(rtx_def*, rtx_def*) ../../../riscv-gcc/gcc/expr.cc:4342 0x170c458 insert_value_copy_on_edge ../../../riscv-gcc/gcc/tree-outof-ssa.cc:352 0x170d58e eliminate_phi ../../../riscv-gcc/gcc/tree-outof-ssa.cc:785 0x170df17 expand_phi_nodes(ssaexpand*) ../../../riscv-gcc/gcc/tree-outof-ssa.cc:1024 0xef27e2 execute ../../../riscv-gcc/gcc/cfgexpand.cc:6818 This is because LoopVectorizer assume target is able to handle series const vector when we enable binary operations. Then it will be easily causing ICE like that. gcc/ChangeLog: * config/riscv/autovec.md (@vec_series<mode>): New pattern * config/riscv/riscv-protos.h (expand_vec_series): New function. * config/riscv/riscv-v.cc (emit_binop): Ditto. (emit_indexop): Ditto. (expand_vec_series): Ditto. (expand_const_vector): Add series vector handling. * config/riscv/riscv.cc (riscv_const_insns): Enable series vector for testing. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/series-1.c: New test. * gcc.target/riscv/rvv/autovec/series_run-1.c: New test. --- gcc/config/riscv/autovec.md | 24 ++++ gcc/config/riscv/riscv-protos.h | 1 + gcc/config/riscv/riscv-v.cc | 118 +++++++++++++++++- gcc/config/riscv/riscv.cc | 27 +++- .../gcc.target/riscv/rvv/autovec/series-1.c | 50 ++++++++ .../riscv/rvv/autovec/series_run-1.c | 20 +++ 6 files changed, 236 insertions(+), 4 deletions(-) create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/series-1.c create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/series_run-1.c diff --git a/gcc/config/riscv/autovec.md b/gcc/config/riscv/autovec.md index f1c5ff5951b..99dc4f046b0 100644 --- a/gcc/config/riscv/autovec.md +++ b/gcc/config/riscv/autovec.md @@ -58,3 +58,27 @@ DONE; } ) + +;; ========================================================================= +;; == Vector creation +;; ========================================================================= + +;; ------------------------------------------------------------------------- +;; ---- [INT] Linear series +;; ------------------------------------------------------------------------- +;; Includes: +;; - vid.v +;; - vmul.vx +;; - vadd.vx/vadd.vi +;; ------------------------------------------------------------------------- + +(define_expand "@vec_series<mode>" + [(match_operand:VI 0 "register_operand") + (match_operand:<VEL> 1 "reg_or_int_operand") + (match_operand:<VEL> 2 "reg_or_int_operand")] + "TARGET_VECTOR" + { + riscv_vector::expand_vec_series (operands[0], operands[1], operands[2]); + DONE; + } +) diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h index c0293a306f9..e8a728ae226 100644 --- a/gcc/config/riscv/riscv-protos.h +++ b/gcc/config/riscv/riscv-protos.h @@ -219,6 +219,7 @@ rtx gen_avl_for_scalar_move (rtx); void expand_tuple_move (machine_mode, rtx *); machine_mode preferred_simd_mode (scalar_mode); opt_machine_mode get_mask_mode (machine_mode); +void expand_vec_series (rtx, rtx, rtx); } /* We classify builtin types into two classes: diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc index 7ca49ca67c1..0c3b1b4c40b 100644 --- a/gcc/config/riscv/riscv-v.cc +++ b/gcc/config/riscv/riscv-v.cc @@ -248,6 +248,111 @@ emit_nonvlmax_op (unsigned icode, rtx dest, rtx src, rtx len, emit_pred_op (icode, NULL_RTX, dest, src, len, mask_mode, false); } +/* Emit binary operations. */ + +static void +emit_binop (unsigned icode, rtx *ops, machine_mode mask_mode, + machine_mode scalar_mode) +{ + insn_expander<9> e; + machine_mode mode = GET_MODE (ops[0]); + e.add_output_operand (ops[0], mode); + e.add_all_one_mask_operand (mask_mode); + e.add_vundef_operand (mode); + if (VECTOR_MODE_P (GET_MODE (ops[1]))) + e.add_input_operand (ops[1], GET_MODE (ops[1])); + else + e.add_input_operand (ops[1], scalar_mode); + if (VECTOR_MODE_P (GET_MODE (ops[2]))) + e.add_input_operand (ops[2], GET_MODE (ops[2])); + else + e.add_input_operand (ops[2], scalar_mode); + rtx vlmax = gen_reg_rtx (Pmode); + emit_vlmax_vsetvl (mode, vlmax); + e.add_input_operand (vlmax, Pmode); + e.add_policy_operand (get_prefer_tail_policy (), get_prefer_mask_policy ()); + e.add_avl_type_operand (avl_type::VLMAX); + e.expand ((enum insn_code) icode, false); +} + +/* Emit vid.v instruction. */ + +static void +emit_indexop (rtx target, machine_mode mask_mode) +{ + insn_expander<7> e; + machine_mode mode = GET_MODE (target); + e.add_output_operand (target, mode); + e.add_all_one_mask_operand (mask_mode); + e.add_vundef_operand (mode); + rtx vlmax = gen_reg_rtx (Pmode); + emit_vlmax_vsetvl (mode, vlmax); + e.add_input_operand (vlmax, Pmode); + e.add_policy_operand (get_prefer_tail_policy (), get_prefer_mask_policy ()); + e.add_avl_type_operand (avl_type::VLMAX); + e.expand (code_for_pred_series (mode), false); +} + +/* Expand series const vector. */ + +void +expand_vec_series (rtx dest, rtx base, rtx step) +{ + machine_mode mode = GET_MODE (dest); + machine_mode inner_mode = GET_MODE_INNER (mode); + machine_mode mask_mode; + gcc_assert (get_mask_mode (mode).exists (&mask_mode)); + + /* VECT_IV = BASE + I * STEP. */ + + /* Step 1: Generate I = { 0, 1, 2, ... } by vid.v. */ + rtx tmp = gen_reg_rtx (mode); + emit_indexop (tmp, mask_mode); + if (rtx_equal_p (step, const1_rtx) && rtx_equal_p (base, const0_rtx)) + { + emit_move_insn (dest, tmp); + return; + } + + /* Step 2: Generate I * STEP. + - STEP is 1, we don't emit any instructions. + - STEP is power of 2, we use vsll.vi/vsll.vx. + - STEP is non-power of 2, we use vmul.vx. */ + rtx tmp2 = gen_reg_rtx (mode); + if (!rtx_equal_p (step, const1_rtx)) + { + if (CONST_INT_P (step) && pow2p_hwi (INTVAL (step))) + { + /* Emit logical left shift operation. */ + int shift = exact_log2 (INTVAL (step)); + rtx shift_amount = gen_int_mode (shift, Pmode); + rtx ops[3] = {tmp2, tmp, shift_amount}; + insn_code icode = code_for_pred_scalar (ASHIFT, mode); + emit_binop (icode, ops, mask_mode, Pmode); + } + else + { + rtx ops[3] = {tmp2, tmp, step}; + insn_code icode = code_for_pred_scalar (MULT, mode); + emit_binop (icode, ops, mask_mode, inner_mode); + } + if (rtx_equal_p (base, const0_rtx)) + { + emit_move_insn (dest, tmp2); + return; + } + } + + /* Step 3: Generate BASE + I * STEP. + - BASE is 0, we don't emit any instructions. + - BASE is not 0, we use vadd.vx/vadd.vi. */ + rtx tmp3 = gen_reg_rtx (mode); + rtx ops[3] = {tmp3, rtx_equal_p (step, const1_rtx) ? tmp : tmp2, base}; + insn_code icode = code_for_pred_scalar (PLUS, mode); + emit_binop (icode, ops, mask_mode, inner_mode); + emit_move_insn (dest, tmp3); +} + static void expand_const_vector (rtx target, rtx src, machine_mode mask_mode) { @@ -280,12 +385,19 @@ expand_const_vector (rtx target, rtx src, machine_mode mask_mode) return; } + /* Support scalable const series vector. */ + rtx base, step; + if (const_vec_series_p (src, &base, &step)) + { + emit_insn (gen_vec_series (mode, target, base, step)); + return; + } + /* TODO: We only support const duplicate vector for now. More cases will be supported when we support auto-vectorization: - 1. series vector. - 2. multiple elts duplicate vector. - 3. multiple patterns with multiple elts. */ + 1. multiple elts duplicate vector. + 2. multiple patterns with multiple elts. */ } /* Expand a pre-RA RVV data move from SRC to DEST. diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index ff90c44d811..84e9267bcb2 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -1266,9 +1266,34 @@ riscv_const_insns (rtx x) } case CONST_DOUBLE: - case CONST_VECTOR: /* We can use x0 to load floating-point zero. */ return x == CONST0_RTX (GET_MODE (x)) ? 1 : 0; + case CONST_VECTOR: + { + /* TODO: This is not accurate, we will need to + adapt the COST of CONST_VECTOR in the future + for the following cases: + + - 1. const duplicate vector with element value + in range of [-16, 15]. + - 2. const duplicate vector with element value + out range of [-16, 15]. + - 3. const series vector. + ...etc. */ + if (riscv_v_ext_vector_mode_p (GET_MODE (x))) + { + /* const series vector. */ + rtx base, step; + if (const_vec_series_p (x, &base, &step)) + { + /* This is not accurate, we will need to adapt the COST + * accurately according to BASE && STEP. */ + return 1; + } + } + /* TODO: We may support more const vector in the future. */ + return x == CONST0_RTX (GET_MODE (x)) ? 1 : 0; + } case CONST: /* See if we can refer to X directly. */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/series-1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/series-1.c new file mode 100644 index 00000000000..a01f6ce7411 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/series-1.c @@ -0,0 +1,50 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv32gcv -mabi=ilp32d --param riscv-autovec-preference=fixed-vlmax --param riscv-autovec-lmul=m4" } */ + +#include <stdint.h> + +#define NUM_ELEMS(TYPE) (64 / sizeof (TYPE)) + +#define DEF_LOOP(TYPE, BASE, STEP, SUFFIX) \ + void __attribute__ ((noinline, noclone)) \ + loop_##TYPE##_##SUFFIX (TYPE *restrict a) \ + { \ + for (int i = 0; i < NUM_ELEMS (TYPE); ++i) \ + a[i] = (BASE) + i * (STEP); \ + } + +#define TEST_SEW32_TYPES(T, BASE, STEP, SUFFIX) \ + T (uint32_t, BASE, STEP, SUFFIX) \ + T (int32_t, BASE, STEP, SUFFIX) + +#define TEST_ALL(T) \ + TEST_SEW32_TYPES (T, 0, 1, b0s1) \ + TEST_SEW32_TYPES (T, 0, 2, b0s2) \ + TEST_SEW32_TYPES (T, 0, 3, b0s3) \ + TEST_SEW32_TYPES (T, 0, 8, b0s8) \ + TEST_SEW32_TYPES (T, 0, 9, b0s9) \ + TEST_SEW32_TYPES (T, 0, 16, b0s16) \ + TEST_SEW32_TYPES (T, 0, 17, b0s17) \ + TEST_SEW32_TYPES (T, 0, 32, b0s32) \ + TEST_SEW32_TYPES (T, 0, 33, b0s33) \ + TEST_SEW32_TYPES (T, -16, 1, bm16s1) \ + TEST_SEW32_TYPES (T, 15, 1, b15s1) \ + TEST_SEW32_TYPES (T, -17, 1, bm17s1) \ + TEST_SEW32_TYPES (T, 16, 1, b16s1) \ + TEST_SEW32_TYPES (T, -16, 128, bm16s128) \ + TEST_SEW32_TYPES (T, 15, 128, b15s128) \ + TEST_SEW32_TYPES (T, -17, 128, bm17s128) \ + TEST_SEW32_TYPES (T, 16, 128, b16s128) \ + TEST_SEW32_TYPES (T, -16, 179, bm16s179) \ + TEST_SEW32_TYPES (T, 15, 179, b15s179) \ + TEST_SEW32_TYPES (T, -17, 179, bm17s179) \ + TEST_SEW32_TYPES (T, 16, 179, b16s179) \ + TEST_SEW32_TYPES (T, -16, 65536, bm16s65536) \ + TEST_SEW32_TYPES (T, 15, 65536, b15s65536) \ + TEST_SEW32_TYPES (T, -17, 65536, bm17s65536) \ + TEST_SEW32_TYPES (T, 16, 65536, b16s65536) + +TEST_ALL (DEF_LOOP) + +/* { dg-final { scan-assembler-times {vid\.v\s+v[0-9]+} 50 } } */ +/* { dg-final { scan-assembler-times {vsll\.vi\s+v[0-9]+} 24 } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/series_run-1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/series_run-1.c new file mode 100644 index 00000000000..09a20809c65 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/series_run-1.c @@ -0,0 +1,20 @@ +/* { dg-do run { target { riscv_vector } } } */ +/* { dg-options "--param riscv-autovec-preference=fixed-vlmax --param riscv-autovec-lmul=m4" } */ + +#include "series-1.c" + +#define TEST_LOOP(TYPE, BASE, STEP, SUFFIX) \ + { \ + TYPE array[NUM_ELEMS (TYPE)] = {}; \ + loop_##TYPE##_##SUFFIX (array); \ + for (int i = 0; i < NUM_ELEMS (TYPE); i++) \ + if (array[i] != (TYPE) (BASE + i * STEP)) \ + __builtin_abort (); \ + } + +int __attribute__ ((optimize (1))) +main () +{ + TEST_ALL (TEST_LOOP) + return 0; +} -- 2.36.3