[PATCH v4 0/2] target/riscv: Generate strided vector ld/st with tcg

[PATCH v4 0/2] target/riscv: Generate strided vector ld/st with tcg

[Chao Liu]

Hi all,

Thanks Richard for the review.

For patch v4, the main modification is to improve the implementation of the
get_log2() function.

+static inline uint32_t get_log2(uint32_t a)
+{
-    uint32_t i = 0;
-    if (a == 0) {
-        return i;
-    }
-    for (; a > 1;) {
-        a >>= 1;
-        i++;
-    } 
+    assert(is_power_of_2(a));
+    return ctz32(a);
+}

patch v3 changes:
- Fix the get_log2() function:
  https://lore.kernel.org/qemu-riscv/cover.1755287531.git.chao.liu@yeah.net/T/#t
- Add test for vlsseg8e32 instruction.
- Rebase on top of the latest master.

patch v2 changes:
- Split the TCG node emulation of the complex strided load/store operation into
  two separate functions to simplify the implementation:
  https://lore.kernel.org/qemu-riscv/20250312155547.289642-1-paolo.savini@embecosm.com/


Best regards,

Chao

Chao Liu (2):
  Generate strided vector loads/stores with tcg nodes.
  tests/tcg/riscv64: Add test for vlsseg8e32 instruction

 target/riscv/insn_trans/trans_rvv.c.inc   | 319 ++++++++++++++++++----
 tests/tcg/riscv64/Makefile.softmmu-target |   8 +-
 tests/tcg/riscv64/test-vlsseg8e32.S       | 107 ++++++++
 3 files changed, 382 insertions(+), 52 deletions(-)
 create mode 100644 tests/tcg/riscv64/test-vlsseg8e32.S

-- 
2.50.1

[PATCH v4 1/2] target/riscv: Generate strided vector loads/stores with tcg nodes.

[Chao Liu]

This commit improves the performance of QEMU when emulating strided vector
loads and stores by substituting the call for the helper function with the
generation of equivalent TCG operations.

Signed-off-by: Paolo Savini <paolo.savini@embecosm.com>
Signed-off-by: Chao Liu <chao.liu@yeah.net>
Tested-by: Eric Biggers <ebiggers@kernel.org>
Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
---
 target/riscv/insn_trans/trans_rvv.c.inc | 319 ++++++++++++++++++++----
 1 file changed, 269 insertions(+), 50 deletions(-)

diff --git a/target/riscv/insn_trans/trans_rvv.c.inc b/target/riscv/insn_trans/trans_rvv.c.inc
index 71f98fb350..341e392064 100644
--- a/target/riscv/insn_trans/trans_rvv.c.inc
+++ b/target/riscv/insn_trans/trans_rvv.c.inc
@@ -864,32 +864,282 @@ GEN_VEXT_TRANS(vlm_v, MO_8, vlm_v, ld_us_mask_op, ld_us_mask_check)
 GEN_VEXT_TRANS(vsm_v, MO_8, vsm_v, st_us_mask_op, st_us_mask_check)
 
 /*
- *** stride load and store
+ * MAXSZ returns the maximum vector size can be operated in bytes,
+ * which is used in GVEC IR when vl_eq_vlmax flag is set to true
+ * to accelerate vector operation.
+ */
+static inline uint32_t MAXSZ(DisasContext *s)
+{
+    int max_sz = s->cfg_ptr->vlenb << 3;
+    return max_sz >> (3 - s->lmul);
+}
+
+static inline uint32_t get_log2(uint32_t a)
+{
+    assert(is_power_of_2(a));
+    return ctz32(a);
+}
+
+typedef void gen_tl_ldst(TCGv, TCGv_ptr, tcg_target_long);
+
+/*
+ * Simulate the strided load/store main loop:
+ *
+ * for (i = env->vstart; i < env->vl; env->vstart = ++i) {
+ *     k = 0;
+ *     while (k < nf) {
+ *         if (!vm && !vext_elem_mask(v0, i)) {
+ *             vext_set_elems_1s(vd, vma, (i + k * max_elems) * esz,
+ *                               (i + k * max_elems + 1) * esz);
+ *             k++;
+ *             continue;
+ *         }
+ *         target_ulong addr = base + stride * i + (k << log2_esz);
+ *         ldst(env, adjust_addr(env, addr), i + k * max_elems, vd, ra);
+ *         k++;
+ *     }
+ * }
  */
-typedef void gen_helper_ldst_stride(TCGv_ptr, TCGv_ptr, TCGv,
-                                    TCGv, TCGv_env, TCGv_i32);
+static void gen_ldst_stride_main_loop(DisasContext *s, TCGv dest, uint32_t rs1,
+                                      uint32_t rs2, uint32_t vm, uint32_t nf,
+                                      gen_tl_ldst *ld_fn, gen_tl_ldst *st_fn,
+                                      bool is_load)
+{
+    TCGv addr = tcg_temp_new();
+    TCGv base = get_gpr(s, rs1, EXT_NONE);
+    TCGv stride = get_gpr(s, rs2, EXT_NONE);
+
+    TCGv i = tcg_temp_new();
+    TCGv i_esz = tcg_temp_new();
+    TCGv k = tcg_temp_new();
+    TCGv k_esz = tcg_temp_new();
+    TCGv k_max = tcg_temp_new();
+    TCGv mask = tcg_temp_new();
+    TCGv mask_offs = tcg_temp_new();
+    TCGv mask_offs_64 = tcg_temp_new();
+    TCGv mask_elem = tcg_temp_new();
+    TCGv mask_offs_rem = tcg_temp_new();
+    TCGv vreg = tcg_temp_new();
+    TCGv dest_offs = tcg_temp_new();
+    TCGv stride_offs = tcg_temp_new();
+
+    uint32_t max_elems = MAXSZ(s) >> s->sew;
+
+    TCGLabel *start = gen_new_label();
+    TCGLabel *end = gen_new_label();
+    TCGLabel *start_k = gen_new_label();
+    TCGLabel *inc_k = gen_new_label();
+    TCGLabel *end_k = gen_new_label();
+
+    MemOp atomicity = MO_ATOM_NONE;
+    if (s->sew == 0) {
+        atomicity = MO_ATOM_NONE;
+    } else {
+        atomicity = MO_ATOM_IFALIGN_PAIR;
+    }
+
+    mark_vs_dirty(s);
+
+    tcg_gen_addi_tl(mask, (TCGv)tcg_env, vreg_ofs(s, 0));
+
+    /* Start of outer loop. */
+    tcg_gen_mov_tl(i, cpu_vstart);
+    gen_set_label(start);
+    tcg_gen_brcond_tl(TCG_COND_GE, i, cpu_vl, end);
+    tcg_gen_shli_tl(i_esz, i, s->sew);
+    /* Start of inner loop. */
+    tcg_gen_movi_tl(k, 0);
+    gen_set_label(start_k);
+    tcg_gen_brcond_tl(TCG_COND_GE, k, tcg_constant_tl(nf), end_k);
+    /*
+     * If we are in mask agnostic regime and the operation is not unmasked we
+     * set the inactive elements to 1.
+     */
+    if (!vm && s->vma) {
+        TCGLabel *active_element = gen_new_label();
+        /* (i + k * max_elems) * esz */
+        tcg_gen_shli_tl(mask_offs, k, get_log2(max_elems << s->sew));
+        tcg_gen_add_tl(mask_offs, mask_offs, i_esz);
+
+        /*
+         * Check whether the i bit of the mask is 0 or 1.
+         *
+         * static inline int vext_elem_mask(void *v0, int index)
+         * {
+         *     int idx = index / 64;
+         *     int pos = index  % 64;
+         *     return (((uint64_t *)v0)[idx] >> pos) & 1;
+         * }
+         */
+        tcg_gen_shri_tl(mask_offs_64, mask_offs, 3);
+        tcg_gen_add_tl(mask_offs_64, mask_offs_64, mask);
+        tcg_gen_ld_i64((TCGv_i64)mask_elem, (TCGv_ptr)mask_offs_64, 0);
+        tcg_gen_rem_tl(mask_offs_rem, mask_offs, tcg_constant_tl(8));
+        tcg_gen_shr_tl(mask_elem, mask_elem, mask_offs_rem);
+        tcg_gen_andi_tl(mask_elem, mask_elem, 1);
+        tcg_gen_brcond_tl(TCG_COND_NE, mask_elem, tcg_constant_tl(0),
+                          active_element);
+        /*
+         * Set masked-off elements in the destination vector register to 1s.
+         * Store instructions simply skip this bit as memory ops access memory
+         * only for active elements.
+         */
+        if (is_load) {
+            tcg_gen_shli_tl(mask_offs, mask_offs, s->sew);
+            tcg_gen_add_tl(mask_offs, mask_offs, dest);
+            st_fn(tcg_constant_tl(-1), (TCGv_ptr)mask_offs, 0);
+        }
+        tcg_gen_br(inc_k);
+        gen_set_label(active_element);
+    }
+    /*
+     * The element is active, calculate the address with stride:
+     * target_ulong addr = base + stride * i + (k << log2_esz);
+     */
+    tcg_gen_mul_tl(stride_offs, stride, i);
+    tcg_gen_shli_tl(k_esz, k, s->sew);
+    tcg_gen_add_tl(stride_offs, stride_offs, k_esz);
+    tcg_gen_add_tl(addr, base, stride_offs);
+    /* Calculate the offset in the dst/src vector register. */
+    tcg_gen_shli_tl(k_max, k, get_log2(max_elems));
+    tcg_gen_add_tl(dest_offs, i, k_max);
+    tcg_gen_shli_tl(dest_offs, dest_offs, s->sew);
+    tcg_gen_add_tl(dest_offs, dest_offs, dest);
+    if (is_load) {
+        tcg_gen_qemu_ld_tl(vreg, addr, s->mem_idx, MO_LE | s->sew | atomicity);
+        st_fn((TCGv)vreg, (TCGv_ptr)dest_offs, 0);
+    } else {
+        ld_fn((TCGv)vreg, (TCGv_ptr)dest_offs, 0);
+        tcg_gen_qemu_st_tl(vreg, addr, s->mem_idx, MO_LE | s->sew | atomicity);
+    }
+    /*
+     * We don't execute the load/store above if the element was inactive.
+     * We jump instead directly to incrementing k and continuing the loop.
+     */
+    if (!vm && s->vma) {
+        gen_set_label(inc_k);
+    }
+    tcg_gen_addi_tl(k, k, 1);
+    tcg_gen_br(start_k);
+    /* End of the inner loop. */
+    gen_set_label(end_k);
+
+    tcg_gen_addi_tl(i, i, 1);
+    tcg_gen_mov_tl(cpu_vstart, i);
+    tcg_gen_br(start);
+
+    /* End of the outer loop. */
+    gen_set_label(end);
+
+    return;
+}
+
+
+/*
+ * Set the tail bytes of the strided loads/stores to 1:
+ *
+ * for (k = 0; k < nf; ++k) {
+ *     cnt = (k * max_elems + vl) * esz;
+ *     tot = (k * max_elems + max_elems) * esz;
+ *     for (i = cnt; i < tot; i += esz) {
+ *         store_1s(-1, vd[vl+i]);
+ *     }
+ * }
+ */
+static void gen_ldst_stride_tail_loop(DisasContext *s, TCGv dest, uint32_t nf,
+                                      gen_tl_ldst *st_fn)
+{
+    TCGv i = tcg_temp_new();
+    TCGv k = tcg_temp_new();
+    TCGv tail_cnt = tcg_temp_new();
+    TCGv tail_tot = tcg_temp_new();
+    TCGv tail_addr = tcg_temp_new();
+
+    TCGLabel *start = gen_new_label();
+    TCGLabel *end = gen_new_label();
+    TCGLabel *start_i = gen_new_label();
+    TCGLabel *end_i = gen_new_label();
+
+    uint32_t max_elems_b = MAXSZ(s);
+    uint32_t esz = 1 << s->sew;
+
+    /* Start of the outer loop. */
+    tcg_gen_movi_tl(k, 0);
+    tcg_gen_shli_tl(tail_cnt, cpu_vl, s->sew);
+    tcg_gen_movi_tl(tail_tot, max_elems_b);
+    tcg_gen_add_tl(tail_addr, dest, tail_cnt);
+    gen_set_label(start);
+    tcg_gen_brcond_tl(TCG_COND_GE, k, tcg_constant_tl(nf), end);
+    /* Start of the inner loop. */
+    tcg_gen_mov_tl(i, tail_cnt);
+    gen_set_label(start_i);
+    tcg_gen_brcond_tl(TCG_COND_GE, i, tail_tot, end_i);
+    /* store_1s(-1, vd[vl+i]); */
+    st_fn(tcg_constant_tl(-1), (TCGv_ptr)tail_addr, 0);
+    tcg_gen_addi_tl(tail_addr, tail_addr, esz);
+    tcg_gen_addi_tl(i, i, esz);
+    tcg_gen_br(start_i);
+    /* End of the inner loop. */
+    gen_set_label(end_i);
+    /* Update the counts */
+    tcg_gen_addi_tl(tail_cnt, tail_cnt, max_elems_b);
+    tcg_gen_addi_tl(tail_tot, tail_cnt, max_elems_b);
+    tcg_gen_addi_tl(k, k, 1);
+    tcg_gen_br(start);
+    /* End of the outer loop. */
+    gen_set_label(end);
+
+    return;
+}
 
 static bool ldst_stride_trans(uint32_t vd, uint32_t rs1, uint32_t rs2,
-                              uint32_t data, gen_helper_ldst_stride *fn,
-                              DisasContext *s)
+                              uint32_t data, DisasContext *s, bool is_load)
 {
-    TCGv_ptr dest, mask;
-    TCGv base, stride;
-    TCGv_i32 desc;
+    if (!s->vstart_eq_zero) {
+        return false;
+    }
 
-    dest = tcg_temp_new_ptr();
-    mask = tcg_temp_new_ptr();
-    base = get_gpr(s, rs1, EXT_NONE);
-    stride = get_gpr(s, rs2, EXT_NONE);
-    desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb,
-                                      s->cfg_ptr->vlenb, data));
+    TCGv dest = tcg_temp_new();
 
-    tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, vd));
-    tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0));
+    uint32_t nf = FIELD_EX32(data, VDATA, NF);
+    uint32_t vm = FIELD_EX32(data, VDATA, VM);
+
+    /* Destination register and mask register */
+    tcg_gen_addi_tl(dest, (TCGv)tcg_env, vreg_ofs(s, vd));
+
+    /*
+     * Select the appropriate load/tore to retrieve data from the vector
+     * register given a specific sew.
+     */
+    static gen_tl_ldst * const ld_fns[4] = {
+        tcg_gen_ld8u_tl, tcg_gen_ld16u_tl,
+        tcg_gen_ld32u_tl, tcg_gen_ld_tl
+    };
+
+    static gen_tl_ldst * const st_fns[4] = {
+        tcg_gen_st8_tl, tcg_gen_st16_tl,
+        tcg_gen_st32_tl, tcg_gen_st_tl
+    };
+
+    gen_tl_ldst *ld_fn = ld_fns[s->sew];
+    gen_tl_ldst *st_fn = st_fns[s->sew];
+
+    if (ld_fn == NULL || st_fn == NULL) {
+        return false;
+    }
 
     mark_vs_dirty(s);
 
-    fn(dest, mask, base, stride, tcg_env, desc);
+    gen_ldst_stride_main_loop(s, dest, rs1, rs2, vm, nf, ld_fn, st_fn, is_load);
+
+    tcg_gen_movi_tl(cpu_vstart, 0);
+
+    /*
+     * Set the tail bytes to 1 if tail agnostic:
+     */
+    if (s->vta != 0 && is_load) {
+        gen_ldst_stride_tail_loop(s, dest, nf, st_fn);
+    }
 
     finalize_rvv_inst(s);
     return true;
@@ -898,16 +1148,6 @@ static bool ldst_stride_trans(uint32_t vd, uint32_t rs1, uint32_t rs2,
 static bool ld_stride_op(DisasContext *s, arg_rnfvm *a, uint8_t eew)
 {
     uint32_t data = 0;
-    gen_helper_ldst_stride *fn;
-    static gen_helper_ldst_stride * const fns[4] = {
-        gen_helper_vlse8_v, gen_helper_vlse16_v,
-        gen_helper_vlse32_v, gen_helper_vlse64_v
-    };
-
-    fn = fns[eew];
-    if (fn == NULL) {
-        return false;
-    }
 
     uint8_t emul = vext_get_emul(s, eew);
     data = FIELD_DP32(data, VDATA, VM, a->vm);
@@ -915,7 +1155,7 @@ static bool ld_stride_op(DisasContext *s, arg_rnfvm *a, uint8_t eew)
     data = FIELD_DP32(data, VDATA, NF, a->nf);
     data = FIELD_DP32(data, VDATA, VTA, s->vta);
     data = FIELD_DP32(data, VDATA, VMA, s->vma);
-    return ldst_stride_trans(a->rd, a->rs1, a->rs2, data, fn, s);
+    return ldst_stride_trans(a->rd, a->rs1, a->rs2, data, s, true);
 }
 
 static bool ld_stride_check(DisasContext *s, arg_rnfvm* a, uint8_t eew)
@@ -933,23 +1173,13 @@ GEN_VEXT_TRANS(vlse64_v, MO_64, rnfvm, ld_stride_op, ld_stride_check)
 static bool st_stride_op(DisasContext *s, arg_rnfvm *a, uint8_t eew)
 {
     uint32_t data = 0;
-    gen_helper_ldst_stride *fn;
-    static gen_helper_ldst_stride * const fns[4] = {
-        /* masked stride store */
-        gen_helper_vsse8_v,  gen_helper_vsse16_v,
-        gen_helper_vsse32_v,  gen_helper_vsse64_v
-    };
 
     uint8_t emul = vext_get_emul(s, eew);
     data = FIELD_DP32(data, VDATA, VM, a->vm);
     data = FIELD_DP32(data, VDATA, LMUL, emul);
     data = FIELD_DP32(data, VDATA, NF, a->nf);
-    fn = fns[eew];
-    if (fn == NULL) {
-        return false;
-    }
 
-    return ldst_stride_trans(a->rd, a->rs1, a->rs2, data, fn, s);
+    return ldst_stride_trans(a->rd, a->rs1, a->rs2, data, s, false);
 }
 
 static bool st_stride_check(DisasContext *s, arg_rnfvm* a, uint8_t eew)
@@ -1300,17 +1530,6 @@ GEN_LDST_WHOLE_TRANS(vs8r_v, int8_t, 8, false)
  *** Vector Integer Arithmetic Instructions
  */
 
-/*
- * MAXSZ returns the maximum vector size can be operated in bytes,
- * which is used in GVEC IR when vl_eq_vlmax flag is set to true
- * to accelerate vector operation.
- */
-static inline uint32_t MAXSZ(DisasContext *s)
-{
-    int max_sz = s->cfg_ptr->vlenb * 8;
-    return max_sz >> (3 - s->lmul);
-}
-
 static bool opivv_check(DisasContext *s, arg_rmrr *a)
 {
     return require_rvv(s) &&
-- 
2.50.1

[PATCH v4 2/2] tests/tcg/riscv64: Add test for vlsseg8e32 instruction

[Chao Liu]

This case, it copied 64 bytes from a0 to a1 with vlsseg8e32.
---
 tests/tcg/riscv64/Makefile.softmmu-target |   8 +-
 tests/tcg/riscv64/test-vlsseg8e32.S       | 107 ++++++++++++++++++++++
 2 files changed, 113 insertions(+), 2 deletions(-)
 create mode 100644 tests/tcg/riscv64/test-vlsseg8e32.S

diff --git a/tests/tcg/riscv64/Makefile.softmmu-target b/tests/tcg/riscv64/Makefile.softmmu-target
index 3ca595335d..384c291554 100644
--- a/tests/tcg/riscv64/Makefile.softmmu-target
+++ b/tests/tcg/riscv64/Makefile.softmmu-target
@@ -7,14 +7,14 @@ VPATH += $(TEST_SRC)
 
 LINK_SCRIPT = $(TEST_SRC)/semihost.ld
 LDFLAGS = -T $(LINK_SCRIPT)
-CFLAGS += -g -Og
+CFLAGS += -march=rv64gcv -mabi=lp64d -g -Og
 
 %.o: %.S
 	$(CC) $(CFLAGS) $< -Wa,--noexecstack -c -o $@
 %: %.o $(LINK_SCRIPT)
 	$(LD) $(LDFLAGS) $< -o $@
 
-QEMU_OPTS += -M virt -display none -semihosting -device loader,file=
+QEMU_OPTS += -M virt -cpu rv64,v=true -display none -semihosting -device loader,file=
 
 EXTRA_RUNS += run-issue1060
 run-issue1060: issue1060
@@ -24,5 +24,9 @@ EXTRA_RUNS += run-test-mepc-masking
 run-test-mepc-masking: test-mepc-masking
 	$(call run-test, $<, $(QEMU) $(QEMU_OPTS)$<)
 
+EXTRA_RUNS += run-vlsseg8e32
+run-vlsseg8e32: test-vlsseg8e32
+	$(call run-test, $<, $(QEMU) $(QEMU_OPTS)$<)
+
 # We don't currently support the multiarch system tests
 undefine MULTIARCH_TESTS
diff --git a/tests/tcg/riscv64/test-vlsseg8e32.S b/tests/tcg/riscv64/test-vlsseg8e32.S
new file mode 100644
index 0000000000..bbc79d5e8d
--- /dev/null
+++ b/tests/tcg/riscv64/test-vlsseg8e32.S
@@ -0,0 +1,107 @@
+#
+# QEMU RISC-V Vector Strided Load Instruction testcase
+#
+# Copyright (c) 2025 Chao Liu chao.liu@yeah.net
+#
+# SPDX-License-Identifier: GPL-2.0-or-later
+#
+	.option	norvc
+
+	.section .data
+	.align 4
+source_data:
+	.asciz "Test the vssseg8e32 insn by copy 64b and verifying correctness."
+	.equ source_len, 64
+
+	.text
+	.global _start
+_start:
+	lla	t0, trap
+	csrw	mtvec, t0
+
+enable_rvv:
+
+	li	x15, 0x800000000024112d
+	csrw	0x301, x15
+	li	x1, 0x2200
+	csrr	x2, mstatus
+	or	x2, x2, x1
+	csrw	mstatus, x2
+
+rvv_test_func:
+	la	a0, source_data
+	li	a1, 0x80020000
+	vsetivli	zero, 1, e32, m1, ta, ma
+	li	t0, 64
+
+	vlsseg8e32.v	v0, (a0), t0
+	addi	a0, a0, 32
+	vlsseg8e32.v	v8, (a0), t0
+
+	vssseg8e32.v	v0, (a1), t0
+	addi	a1, a1, 32
+	vssseg8e32.v	v8, (a1), t0
+
+compare_start:
+	la	a0, source_data
+	li	a1, 0x80020000
+	li	t0, 0
+	li	t1, source_len
+
+compare_loop:
+	# when t0 >= len, compare end
+	bge	 t0, t1, compare_done
+
+	lb	t2, 0(a0)
+	lb	t3, 0(a1)
+	bne	t2, t3, compare_fail
+
+	addi	a0, a0, 1
+	addi	a1, a1, 1
+	addi	t0, t0, 1
+	j	compare_loop
+
+compare_done:
+	# compare ok, return 0
+	li	a0, 0
+	j	_exit
+
+compare_fail:
+	# compare failed, return 2
+	li	a0, 2
+	j	_exit
+
+trap:
+	# When an instruction traps, compare it to the insn in memory.
+	csrr	t0, mepc
+	csrr	t1, mtval
+	lwu	t2, 0(t0)
+	bne	t1, t2, fail
+
+	# Skip the insn and continue.
+	addi	t0, t0, 4
+	csrw	mepc, t0
+	mret
+
+fail:
+	li	a0, 1
+
+# Exit code in a0
+_exit:
+	lla	a1, semiargs
+	li	t0, 0x20026	# ADP_Stopped_ApplicationExit
+	sd	t0, 0(a1)
+	sd	a0, 8(a1)
+	li	a0, 0x20	# TARGET_SYS_EXIT_EXTENDED
+
+	# Semihosting call sequence
+	.balign	16
+	slli	zero, zero, 0x1f
+	ebreak
+	srai	zero, zero, 0x7
+	j	.
+
+	.data
+	.balign	16
+semiargs:
+	.space	16
-- 
2.50.1