Detect NaN in FADD and FSUB

[2011eric]

I used the pattern specified in the RISC-V ISA manual: 0x7fc00000.

[jserv]

List the difference of the results generated by make arch-test RISCV_DEVICE=F between the original implementation and the proposed one.

[2011eric]

Moved the definition to riscv_private.h

[jserv]

Moved the definition to riscv_private.h

Check CI regressions and attempt to fix.

[jserv]

Use "git rebase -i" to squash/rework the commits.

[jserv]

List the difference of the results generated by make arch-test RISCV_DEVICE=F between the original implementation and the proposed one.

@2011eric, you should mention the results within git commit messages.
See https://cbea.ms/git-commit/

[2011eric]

Passed 5 more tests in RV32F, listed in the commit message.

[jserv]

https://cbea.ms/git-commit/

Read How to Write a Git Commit Message carefully and enforce the rules:

• Do not end the subject line with a period

[2011eric]

I noticed that in the expected output_signature of test fadd_b1-01, one of the results is ff800000, which should be -INF.
However, the overflow flag (OF) and inexact flag (NX) are not set.

[jserv]

I noticed that in the expected output_signature of test fadd_b1-01, one of the results is ff800000, which should be -INF. However, the overflow flag (OF) and inexact flag (NX) are not set.

You can submit issue(s) about floating-point compliance tests if you can recognize.

[2011eric]

I noticed that in the expected output_signature of test fadd_b1-01, one of the results is ff800000, which should be -INF. However, the overflow flag (OF) and inexact flag (NX) are not set.

You can submit issue(s) about floating-point compliance tests if you can recognize.

I'm not sure if I can state the problem.
In some cases, the result of the operation differs in LSB. And I don't know if it is related to rounding.
In other cases, our simulator reports zero while the golden data is 'subnormal 0' .
And in some cases, the result of FADD is correct, but the flag is set incorectly.

[jserv]

Rewrite as following:

uint32_t out = *c;
if (csr == CSR_FFLAGS)
    out &= FFLAG_MASK;

[2011eric]

I rewrote the function and also add the same code in csr_csrrw and csr_csrrc.

Since FCSR is defined only if ENABLE_RV32F is defined:

static uint32_t *csr_get_ptr(struct riscv_t *rv, uint32_t csr)
{
    switch (csr) {
    case CSR_CYCLE:
        return (uint32_t *) (&rv->csr_cycle) + 0;
    case CSR_CYCLEH:
        return (uint32_t *) (&rv->csr_cycle) + 1;
    case CSR_MSTATUS:
        return (uint32_t *) (&rv->csr_mstatus);
    case CSR_MTVEC:
        return (uint32_t *) (&rv->csr_mtvec);
    case CSR_MISA:
        return (uint32_t *) (&rv->csr_misa);
    case CSR_MSCRATCH:
        return (uint32_t *) (&rv->csr_mscratch);
    case CSR_MEPC:
        return (uint32_t *) (&rv->csr_mepc);
    case CSR_MCAUSE:
        return (uint32_t *) (&rv->csr_mcause);
    case CSR_MTVAL:
        return (uint32_t *) (&rv->csr_mtval);
    case CSR_MIP:
        return (uint32_t *) (&rv->csr_mip);
#ifdef ENABLE_RV32F
    case CSR_FFLAGS:
        return (uint32_t *) (&rv->csr_fcsr);
    case CSR_FCSR:
        return (uint32_t *) (&rv->csr_fcsr);
#endif
    default:
        return NULL;
    }
}

Should I also use macro inside csr_csrrs ?

[jserv]

Should I also use macro inside csr_csrrs ?

You can do it in another pull request.

[jserv]

If ENABLE_RV32F is turned off, build system would complain:

emulate.c: In function ‘csr_csrrw’:
emulate.c:644:16: error: ‘FFLAG_MASK’ undeclared (first use in this function)
  644 |         out &= FFLAG_MASK;
      |                ^~~~~~~~~~
emulate.c:644:16: note: each undeclared identifier is reported only once for each function it appears in
emulate.c: In function ‘csr_csrrs’:
emulate.c:661:16: error: ‘FFLAG_MASK’ undeclared (first use in this function)
  661 |         out &= FFLAG_MASK;
      |                ^~~~~~~~~~
emulate.c: In function ‘csr_csrrc’:
emulate.c:678:16: error: ‘FFLAG_MASK’ undeclared (first use in this function)
  678 |         out &= FFLAG_MASK;
      |                ^~~~~~~~~~

Resolve the above compilation errors.

[jserv]

gcc-9 raises the following warnings:

emulate.c:918:15: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
  918 |             *((uint32_t *) &rv->F[rd]) = RV_NAN;
      |              ~^~~~~~~~~~~~~~~~~~~~~~~~
emulate.c:930:15: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
  930 |             *((uint32_t *) &rv->F[rd]) = RV_NAN;
      |              ~^~~~~~~~~~~~~~~~~~~~~~~~

[jserv]

Also, rebase the latest master branch.