Generating tests for VLEN=64 fails

[EAlexJ]

When building executables with VLEN equal to 64, faulty patches are generated for at least one instruction. For XLEN=32:

build/merger -stage1output out/v64x32machine/tests/stage1/ -stage2output out/v64x32machine/tests/stage2/ -stage2patch out/v64x32machine/patches/stage2/
fatal: wrong patch generated for file vloxei16_v-0.S, expected 190, got 156

and for XLEN=64:

build/merger -stage1output out/v64x64machine/tests/stage1/ -stage2output out/v64x64machine/tests/stage2/ -stage2patch out/v64x64machine/patches/stage2/
fatal: wrong patch generated for file vloxei16_v-0.S, expected 191, got 181

[ksco]

Thanks for the report, I'll take a look next week.

[zhongchengyong]

Similar issue：fatal: wrong patch generated for file vsse16_v-0.S, expected 625, got 334

[nadime15]

I’m not entirely sure, but this could be Spike related. In some cases, even when I set VLEN=64, Spike seemed to internally use VLEN=128.

I discovered this while running Spike in debug mode and printing VLEN and ELEN.

[nadime15]

I can confirm that the issue is setting illegal combinations of V, Zvl and Zve as described in #78.

I got VLEN=64, XLEN=64 working for vloxei16_v

When compiled with

isa=rv64gc_zvbb_zvbc_zfh_zvfh_zvkg_zvkned_zvknha_zvksed_zvksh_zvl64b_zve64d_zfh_zfhmin_zvfh

(I removed V, Zvfbfmin and Zvfbfwma)

[ksco]

So a change to the isa string in Makefile would resolve this issue?

[nadime15]

For VLEN=64 and VLEN=32 (I think we can start supporting VLEN=32), yes. I will open a PR later.

[nadime15]

We can close this one.