[risc-v][vector] Restore vstart after vector configuration

9 files changed, 262 insertions, 8 deletions

diff --git a/bsp/qemu-virt64-riscv/.ci/attachconfig/ci.attachconfig.yml b/bsp/qemu-virt64-riscv/.ci/attachconfig/ci.attachconfig.yml
new file mode 100644
index 0000000000..9385b3498f
--- /dev/null
+++ b/bsp/qemu-virt64-riscv/.ci/attachconfig/ci.attachconfig.yml
@@ -0,0 +1,20 @@
+utest.rvv.vlen128:
+    scons_arg:
+      - '--strict'
+    kconfig:
+      - CONFIG_ENABLE_VECTOR=y
+      - CONFIG_RT_USING_HEAP=y
+      - CONFIG_RT_USING_UTEST=y
+      - CONFIG_ARCH_VECTOR_VLEN_128=y
+      - CONFIG_ARCH_VECTOR_VLEN_256=n
+      - CONFIG_RT_USING_SMP=n
+utest.rvv.vlen256:
+    scons_arg:
+      - '--strict'
+    kconfig:
+      - CONFIG_ENABLE_VECTOR=y
+      - CONFIG_RT_USING_HEAP=y
+      - CONFIG_RT_USING_UTEST=y
+      - CONFIG_ARCH_VECTOR_VLEN_128=n
+      - CONFIG_ARCH_VECTOR_VLEN_256=y
+      - CONFIG_RT_USING_SMP=n
diff --git a/bsp/qemu-virt64-riscv/SConstruct b/bsp/qemu-virt64-riscv/SConstruct
index ae0e3375e1..06d2a468df 100644
--- a/bsp/qemu-virt64-riscv/SConstruct
+++ b/bsp/qemu-virt64-riscv/SConstruct
@@ -27,6 +27,9 @@ rtconfig.ARCH='risc-v'
 # prepare building environment
 objs = PrepareBuilding(env, RTT_ROOT, has_libcpu = False)
 
+if GetDepend('ENABLE_VECTOR'):
+    env['ASFLAGS'] = str(env['ASFLAGS']).replace('-march=rv64imafdc ', '-march=rv64imafdcv ')
+
 stack_size = 4096
 
 if GetDepend('RT_USING_SMART'):
diff --git a/bsp/qemu-virt64-riscv/applications/test/test_vector/SConscript b/bsp/qemu-virt64-riscv/applications/test/test_vector/SConscript
index 0827b048b2..0a287adadb 100644
--- a/bsp/qemu-virt64-riscv/applications/test/test_vector/SConscript
+++ b/bsp/qemu-virt64-riscv/applications/test/test_vector/SConscript
@@ -3,7 +3,12 @@ from building import *
 cwd     = GetCurrentDir()
 src = Glob('*.c') + Glob('*.cpp')
 CPPPATH = [cwd]
+LOCAL_CCFLAGS = ''
 
-group = DefineGroup('Vector', src, depend = [''], CPPPATH = CPPPATH)
+if GetDepend(['RT_USING_UTEST', 'ENABLE_VECTOR']):
+    src += ['vector_context_gcc.S']
+    LOCAL_CCFLAGS = ' -march=rv64gcv'
+
+group = DefineGroup('Vector', src, depend = [''], CPPPATH = CPPPATH, LOCAL_CCFLAGS = LOCAL_CCFLAGS)
 
 Return('group')
diff --git a/bsp/qemu-virt64-riscv/applications/test/test_vector/test_vector.c b/bsp/qemu-virt64-riscv/applications/test/test_vector/test_vector.c
index 50aae80248..b84e1c42b3 100644
--- a/bsp/qemu-virt64-riscv/applications/test/test_vector/test_vector.c
+++ b/bsp/qemu-virt64-riscv/applications/test/test_vector/test_vector.c
@@ -55,8 +55,8 @@ static void test_feature(void)
  * not modified by each other
  * ==============================================================
  */
-#define TEST_THREAD 2
-#define VECTOR_CTX_BYTES (CTX_VECTOR_REG_NR * REGBYTES)
+#define TEST_THREAD      2
+#define VECTOR_CTX_BYTES (CTX_VECTOR_REG_NR * sizeof(rt_ubase_t))
 void *ctx_vector[TEST_THREAD * 2];
 
 static rt_sem_t sem;
diff --git a/bsp/qemu-virt64-riscv/applications/test/test_vector/vector_context_gcc.S b/bsp/qemu-virt64-riscv/applications/test/test_vector/vector_context_gcc.S
new file mode 100644
index 0000000000..42e60f323d
--- /dev/null
+++ b/bsp/qemu-virt64-riscv/applications/test/test_vector/vector_context_gcc.S
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2006-2026, RT-Thread Development Team
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifndef __ASSEMBLY__
+#define __ASSEMBLY__
+#endif
+#define REGBYTES 8
+#define STORE sd
+#define LOAD ld
+
+#include <rtconfig.h>
+#include "vector_encoding.h"
+#include "rvv_context.h"
+
+    .text
+    .align 2
+
+    /* Test-only helpers; SAVE_VECTOR must not change the caller's controls. */
+    .globl rt_hw_vector_ctx_save
+    .type rt_hw_vector_ctx_save, @function
+rt_hw_vector_ctx_save:
+    csrrci  a3, sstatus, 2
+    li      t0, SSTATUS_VS
+    csrs    sstatus, t0
+    SAVE_VECTOR a0
+    RESTORE_VECTOR a0
+    csrw    sstatus, a3
+    ret
+    .size rt_hw_vector_ctx_save, . - rt_hw_vector_ctx_save
+
+    .globl rt_hw_vector_ctx_restore
+    .type rt_hw_vector_ctx_restore, @function
+rt_hw_vector_ctx_restore:
+    csrrci  a3, sstatus, 2
+    li      t0, SSTATUS_VS
+    csrs    sstatus, t0
+    RESTORE_VECTOR a0
+    csrw    sstatus, a3
+    ret
+    .size rt_hw_vector_ctx_restore, . - rt_hw_vector_ctx_restore
diff --git a/bsp/qemu-virt64-riscv/applications/utest/SConscript b/bsp/qemu-virt64-riscv/applications/utest/SConscript
new file mode 100644
index 0000000000..8e56d71ed7
--- /dev/null
+++ b/bsp/qemu-virt64-riscv/applications/utest/SConscript
@@ -0,0 +1,10 @@
+from building import *
+
+cwd = GetCurrentDir()
+src = ['rvv_vstart_tc.c', 'rvv_vstart_probe_gcc.S']
+
+group = DefineGroup('RVV utestcases', src,
+                    depend = ['RT_USING_UTEST', 'ENABLE_VECTOR', 'RT_USING_HEAP'],
+                    CPPPATH = [cwd])
+
+Return('group')
diff --git a/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_probe_gcc.S b/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_probe_gcc.S
new file mode 100644
index 0000000000..e02cbaf393
--- /dev/null
+++ b/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_probe_gcc.S
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2006-2026, RT-Thread Development Team
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Change Logs:
+ * Date           Author       Notes
+ * 2026-10-10     Piyush Patle  Add RVV restore regression probe
+ */
+
+#ifndef __ASSEMBLY__
+#define __ASSEMBLY__
+#endif
+#define REGBYTES 8
+#define STORE sd
+#define LOAD ld
+
+#include <rtconfig.h>
+#include "vector_encoding.h"
+#include "rvv_context.h"
+
+    .text
+    .align 2
+    .option push
+    .option arch, +v
+
+    /* a0: expected frame, a1: observed frame, a2: caller's vector frame. */
+    .globl rvv_vstart_probe
+    .type rvv_vstart_probe, @function
+rvv_vstart_probe:
+    /* Keep interrupts masked until the caller's complete state is restored. */
+    csrrci  a3, sstatus, 2
+    li      t0, SSTATUS_VS
+    csrs    sstatus, t0
+    SAVE_VECTOR a2
+
+    RESTORE_VECTOR a0
+    /* SAVE_VECTOR records vstart before issuing any vector instruction. */
+    SAVE_VECTOR a1
+
+    RESTORE_VECTOR a2
+    csrw    sstatus, a3
+    ret
+    .size rvv_vstart_probe, . - rvv_vstart_probe
+
+    .option pop
diff --git a/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_tc.c b/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_tc.c
new file mode 100644
index 0000000000..1377644718
--- /dev/null
+++ b/bsp/qemu-virt64-riscv/applications/utest/rvv_vstart_tc.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright (c) 2006-2026, RT-Thread Development Team
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Change Logs:
+ * Date           Author       Notes
+ * 2026-10-10     Piyush Patle  Test zero and nonzero RVV restart indices
+ */
+
+#include <rtthread.h>
+
+/* Run with: utest_run testcases.libcpu.rvv_vstart */
+#if defined(ENABLE_VECTOR) && defined(RT_USING_HEAP) && defined(RT_USING_UTEST)
+#include "rvv_context.h"
+
+/* RV64 vector frames contain four control words followed by v0 through v31. */
+#define VECTOR_CONTROL_WORDS 4
+#define VECTOR_GUARD         0x5a5a5a5aUL
+
+struct vector_test_frame
+{
+    rt_ubase_t before[2];
+    rt_ubase_t image[CTX_VECTOR_REG_NR];
+    rt_ubase_t after[2];
+};
+
+void rvv_vstart_probe(const rt_ubase_t *input, rt_ubase_t *output, rt_ubase_t *backup);
+
+static void frame_init(struct vector_test_frame *frame)
+{
+    rt_memset(frame, 0, sizeof(*frame));
+    frame->before[0] = frame->before[1] = VECTOR_GUARD;
+    frame->after[0] = frame->after[1] = VECTOR_GUARD;
+}
+
+static rt_bool_t frame_guards_valid(const struct vector_test_frame *frame)
+{
+    return frame->before[0] == VECTOR_GUARD && frame->before[1] == VECTOR_GUARD &&
+           frame->after[0] == VECTOR_GUARD && frame->after[1] == VECTOR_GUARD;
+}
+
+static int rvv_vstart_test(void)
+{
+    rt_ubase_t vlenb;
+    rt_ubase_t starts[3];
+    struct vector_test_frame *frames;
+    struct vector_test_frame *expected;
+    struct vector_test_frame *observed;
+    struct vector_test_frame *saved;
+    rt_uint8_t *payload;
+    rt_size_t payload_size = sizeof(expected->image) - VECTOR_CONTROL_WORDS * sizeof(rt_ubase_t);
+    rt_size_t i;
+    rt_size_t j;
+
+    __asm__ volatile("csrr %0, vlenb" : "=r"(vlenb));
+    if (payload_size != 32 * vlenb)
+    {
+        rt_kprintf("RVV test: configured frame does not match vlenb=%lu\n", vlenb);
+        return -RT_ERROR;
+    }
+
+    /* Each invocation owns its frames, including the caller-state backup. */
+    frames = rt_malloc(3 * sizeof(*frames));
+    if (frames == RT_NULL)
+    {
+        return -RT_ENOMEM;
+    }
+    expected = &frames[0];
+    observed = &frames[1];
+    saved = &frames[2];
+    payload = (rt_uint8_t *)&expected->image[VECTOR_CONTROL_WORDS];
+    frame_init(expected);
+    for (i = 0; i < payload_size; i++)
+    {
+        payload[i] = (rt_uint8_t)(i * 37 + i / vlenb + 0x5a);
+    }
+
+    starts[0] = 0;
+    starts[1] = 1;
+    starts[2] = vlenb - 1;
+
+    for (i = 0; i < sizeof(starts) / sizeof(starts[0]); i++)
+    {
+        for (j = 0; j < 3; j++)
+        {
+            expected->image[0] = starts[i];
+            expected->image[1] = 0;       /* e8, m1, tail/mask undisturbed */
+            expected->image[2] = vlenb;
+            expected->image[3] = j * 3;  /* Exercise vxrm and vxsat through vcsr. */
+            frame_init(observed);
+            frame_init(saved);
+
+            rvv_vstart_probe(expected->image, observed->image, saved->image);
+
+            if (!frame_guards_valid(expected) || !frame_guards_valid(observed) ||
+                !frame_guards_valid(saved) ||
+                rt_memcmp(expected->image, observed->image, sizeof(expected->image)) != 0)
+            {
+                rt_kprintf("RVV restore failed: vstart expected=%lu observed=%lu\n",
+                           starts[i], observed->image[0]);
+                rt_free(frames);
+                return -RT_ERROR;
+            }
+        }
+    }
+
+    rt_free(frames);
+    rt_kprintf("RVV restore passed: vlenb=%lu, 9 control/register checks\n", vlenb);
+    return RT_EOK;
+}
+
+#include <utest.h>
+
+static void testcase(void)
+{
+    uassert_int_equal(rvv_vstart_test(), RT_EOK);
+}
+
+UTEST_TC_EXPORT(testcase, "testcases.libcpu.rvv_vstart", RT_NULL, RT_NULL, 10);
+#endif
diff --git a/libcpu/risc-v/vector/rvv-1.0/rvv_context.h b/libcpu/risc-v/vector/rvv-1.0/rvv_context.h
index 4466b3aeff..f5ec4d4ae8 100644
--- a/libcpu/risc-v/vector/rvv-1.0/rvv_context.h
+++ b/libcpu/risc-v/vector/rvv-1.0/rvv_context.h
@@ -12,17 +12,20 @@
 #define __RVV_CONTEXT_1_0_H__
 
 #if defined(ARCH_VECTOR_VLEN_128)
-    #define CTX_VECTOR_REGS 64
+#define CTX_VECTOR_REGS 64
 #elif defined(ARCH_VECTOR_VLEN_256)
-    #define CTX_VECTOR_REGS 128
+#define CTX_VECTOR_REGS 128
 #else
 #error "No supported VLEN"
 #endif /* VLEN */
 
-#define CTX_VECTOR_REG_NR  (CTX_VECTOR_REGS + 4)
+#define CTX_VECTOR_REG_NR (CTX_VECTOR_REGS + 4)
 
 #ifdef __ASSEMBLY__
 
+/* The following macros contain assembler, not C. */
+/* clang-format off */
+
 /**
  * ==================================
  * VECTOR EXTENSION
@@ -97,16 +100,19 @@
 
     mv      t1, t0
 
-    LOAD    t0, VEC_FRAME_VSTART(t1)
-    csrw    vstart, t0
     LOAD    t0, VEC_FRAME_VCSR(t1)
     csrw    vcsr, t0
 
     LOAD    t0, VEC_FRAME_VTYPE(t1)
     LOAD    t3, VEC_FRAME_VL(t1)
     VEC_CONFIG_SET_VL_VTYPE(t3, t0)
+
+    /* vsetvl resets vstart, so restore the restart index last. */
+    LOAD    t0, VEC_FRAME_VSTART(t1)
+    csrw    vstart, t0
 .endm
 
+/* clang-format on */
 #endif
 
 #endif /* __RVV_CONTEXT_H__ */