I noticed that the fmul8x16 instruction did not appear to be emulated correctly[1]. It would seem that emulation was not using a single- precision input register like the real hardware does, but rather a double-precision register, causing it to operate on the wrong data.
Every other VIS instruction which contains one or more single-precision inputs and a double-precision output has the exact same problem. A few computational problems are found and fixed by this series too. All patches can be applied independently, except patch 2 adds some helpers which are subsequently needed by patches 3, 4 and 5. Emulation results are tested by manually comparing the output of a small Linux test program on an UltraSparc II against the output of running the same binary under qemu-sparc32plus on a ppc64le host system. [1] https://gitlab.com/qemu-project/qemu/-/issues/1901 Nick Bowler (8): target/sparc: Fix VIS fmul8x16 input register. target/sparc: Fix VIS fmul8x16au instruction. target/sparc: Fix VIS fmul8x16al instruction. target/sparc: Fix VIS fmuld8sux16 instruction. target/sparc: Fix VIS fmuld8ulx16 instruction. target/sparc: Fix VIS fpmerge input registers. target/sparc: Fix VIS fexpand input register. target/sparc: Fix VIS subtraction instructions. target/sparc/helper.h | 14 ++--- target/sparc/translate.c | 42 +++++++++++--- target/sparc/vis_helper.c | 119 +++++++++++++++++++------------------- 3 files changed, 101 insertions(+), 74 deletions(-) -- 2.41.0
