@@ -836,6 +836,13 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \
{ \
intptr_t i = 0, opr_sz = simd_oprsz(desc); \
intptr_t opr_sz_n = opr_sz / sizeof(TYPED); \
+ /* \
+ * Special case: opr_sz == 8 from AA64/AA32 advsimd means the \
+ * first iteration might not be a full 16 byte segment. But \
+ * for vector lengths beyond that this must be SVE and we know \
+ * opr_sz is a multiple of 16, so we need not clamp segend \
+ * to opr_sz_n when we advance it at the end of the loop. \
+ */ \
intptr_t segend = MIN(16 / sizeof(TYPED), opr_sz_n); \
intptr_t index = simd_data(desc); \
TYPED *d = vd, *a = va; \
@@ -853,7 +860,7 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \
n[i * 4 + 2] * m2 + \
n[i * 4 + 3] * m3); \
} while (++i < segend); \
- segend = i + 4; \
+ segend = i + (16 / sizeof(TYPED)); \
} while (i < opr_sz_n); \
clear_tail(d, opr_sz, simd_maxsz(desc)); \
}
Our implementation of the indexed version of SVE SDOT/UDOT/USDOT got the calculation of the inner loop terminator wrong. Although we correctly account for the element size when we calculate the terminator for the first iteration: intptr_t segend = MIN(16 / sizeof(TYPED), opr_sz_n); we don't do that when we move it forward after the first inner loop completes. The intention is that we process the vector in 128-bit segments, which for a 64-bit element size should mean (1, 2), (3, 4), (5, 6), etc. This bug meant that we would iterate (1, 2), (3, 4, 5, 6), (7, 8, 9, 10) etc and apply the wrong indexed element to some of the operations, and also index off the end of the vector. You don't see this bug if the vector length is small enough that we don't need to iterate the outer loop, i.e. if it is only 128 bits, or if it is the 64-bit special case from AA32/AA64 AdvSIMD. If the vector length is 256 bits then we calculate the right results for the elements in the vector but do index off the end of the vector. Vector lengths greater than 256 bits see wrong answers. The instructions that produce 32-bit results behave correctly. Fix the recalculation of 'segend' for subsequent iterations, and restore a version of the comment that was lost in the refactor of commit 7020ffd656a5 that explains why we only need to clamp segend to opr_sz_n for the first iteration, not the later ones. Cc: qemu-stable@nongnu.org Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2595 Fixes: 7020ffd656a5 ("target/arm: Macroize helper_gvec_{s,u}dot_idx_{b,h}") Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20241101185544.2130972-1-peter.maydell@linaro.org --- target/arm/tcg/vec_helper.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)