target/riscv: fix vector register address calculation in strided LD/ST
PATCH 2/2 tests/tcg/riscv64: Add test for vlsseg8e32 instruction
In reply to Chao LiuThis case, it copied 64 bytes from a0 to a1 with vlsseg8e32.
--- tests/tcg/riscv64/Makefile.softmmu-target | 8 +- tests/tcg/riscv64/test-vlsseg8e32.S | 108 ++++++++++++++++++++++ 2 files changed, 114 insertions(+), 2 deletions(-) create mode 100644 tests/tcg/riscv64/test-vlsseg8e32.S
tests/tcg/riscv64/Makefile.softmmu-target +6 -2
@@ -7,18 +7,22 @@ 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 $(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
tests/tcg/riscv64/test-vlsseg8e32.S new file +108 -0
108 lines not on this page · raw message
Signature · 1 line
2.50.1
PATCH 1/2 target/riscv: fix vector register address calculation in strided LD/ST
In reply to Chao Liuhis patch fixes a critical bug in the RISC-V vector instruction translation that caused incorrect data handling in strided load operations (e.g., vlsseg8e32).
Problem Description:
The get_log2 function in trans_rvv.c.inc returned a value 1 higher than the actual log2 value. For example, get_log2(4) incorrectly returned 3 instead of 2.
This led to erroneous vector register offset calculations, resulting in data overlap where bytes 32-47 were incorrectly copied to positions 16-31 in ChaCha20 encryption code.
rvv_test_func:
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
retAnalysis:
The original implementation counted the number of right shifts until zero, including the final shift that reduced the value to zero:
static inline uint32_t get_log2(uint32_t a)
{
uint32_t i = 0;
for (; a > 0;) {
a >>= 1;
i++;
}
return i; // Returns 3 for a=4 (0b100 → 0b10 → 0b1 → 0b0)
}Fix:
The corrected function stops shifting when only the highest bit remains and handles the special case of a=0:
static inline uint32_t get_log2(uint32_t a)
{
uint32_t i = 0;
if (a == 0) {
return i; // Handle edge case
}
for (; a > 1; a >>= 1) {
i++;
}
return i; // Now returns 2 for a=4
}--- target/riscv/insn_trans/trans_rvv.c.inc | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
target/riscv/insn_trans/trans_rvv.c.inc +4 -1
@@ -877,7 +877,10 @@ static inline uint32_t MAXSZ(DisasContext *s) static inline uint32_t get_log2(uint32_t a) { uint32_t i = 0; - for (; a > 0;) { + if (a == 0) { + return i; + } + for (; a > 1;) { a >>= 1; i++; }
Signature · 1 line
2.50.1
cover target/riscv: fix vector register address calculation in strided LD/ST
Hi Paolo, Eric, Daniel,
I have attempted to fix this issue. Thanks to Eric for providing the test case.
This patch fixes a critical bug in the RISC-V vector instruction translation that caused incorrect data handling in strided load operations (e.g.,vlsseg8e32).
#### Problem Description
The get_log2 function in trans_rvv.c.inc returned a value 1 higher than the actual log2 value. For example, get_log2(4) incorrectly returned 3 instead of 2.
This led to erroneous vector register offset calculations, resulting in data overlap where bytes 32-47 were incorrectly copied to positions 16-31 in ChaCha20 encryption code.
rvv_test_func:
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
ret#### Root Cause Analysis
The original implementation counted the number of right shifts until zero, including the final shift that reduced the value to zero:
static inline uint32_t get_log2(uint32_t a)
{
uint32_t i = 0;
for (; a > 0;) {
a >>= 1;
i++;
}
return i; // Returns 3 for a=4 (0b100 → 0b10 → 0b1 → 0b0)
}#### Fix Implementation
The corrected function stops shifting when only the highest bit remains and handles the special case of a=0:
static inline uint32_t get_log2(uint32_t a)
{
uint32_t i = 0;
if (a == 0) {
return i; // Handle edge case
}
for (; a > 1; a >>= 1) {
i++;
}
return i; // Now returns 2 for a=4
}#### Testing
This fix has been verified with:
1. The provided ChaCha20 vector optimization test case
2. RVV strided load instruction tests in test-vlsseg8e32.S
All tests now pass with correct data handling and no memory overlap.
Test using the following command:
./configure --target-list=riscv64-softmmu \
--cross-prefix-riscv64=riscv64-unknown-elf-
ninja -j$(nproc) -C build && make check-tcgExpected result:
BUILD riscv64-softmmu guest-tests
RUN riscv64-softmmu guest-tests
TEST issue1060 on riscv64
TEST test-vlsseg8e32 on riscv64Best regards,
Chao
Chao Liu (2):
target/riscv: fix vector register address calculation in strided LD/ST tests/tcg/riscv64: Add test for vlsseg8e32 instruction
target/riscv/insn_trans/trans_rvv.c.inc | 5 +- tests/tcg/riscv64/Makefile.softmmu-target | 8 +- tests/tcg/riscv64/test-vlsseg8e32.S | 108 ++++++++++++++++++++++ 3 files changed, 118 insertions(+), 3 deletions(-) create mode 100644 tests/tcg/riscv64/test-vlsseg8e32.S
Signature · 1 line
2.50.1