Fix ARM32 JIT errors with native int-div

[clamp03]

Fix errors when USE_HELPERS_FOR_INT_DIV is disabled to use native ARM division instructions, and add the missing GT_UDIV case.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[JulieLeeMSFT]

@dhartglassMSFT, PTAL.

[am11]

when USE_HELPERS_FOR_INT_DIV is disabled

It's enabled with hardcoded value:

runtime/src/coreclr/jit/targetarm.h

Line 14 in 503cdd1

#define USE_HELPERS_FOR_INT_DIV 1 // BeagleBoard (ARMv7A) doesn't support SDIV/UDIV

Do we need to add a runtime detection?

// in src/native/minipal/cpufeatures.h

#if defined(HOST_ARM)
#define ArmIntrinsicConstants_IDIV (1 << 0)
#endif

// in src/native/minipal/cpufeatures.c

#elif defined(HOST_ARM)

#if defined(HOST_UNIX)

    unsigned long hwcaps = getauxval(AT_HWCAP);

    if (hwcaps & HWCAP_IDIV)
    {
        // Our baseline support is ARMv7 with SDIV/UDIV
        result |= ArmIntrinsicConstants_IDIV;
    }

and wire it like Zbs was done in 65e9497.

[clamp03]

@am11 Thanks for the suggestion! Regarding TODO-ARM-CQ: Check for sdiv/udiv at runtime and generate it if available, I agree that implementing a runtime checker would be better. @jkotas @jakobbotsch What do you think?

[jkotas]

Our bar for linux arm32 is to keep it functional. We are not spending any cycles to make it better. I do not think we are interested in investing into new arm32-specific optimizations.

[clamp03]

@jkotas Thank you for your comment. I think PR like error fix can still be reviewed and merged if there is no problem. Could you review this PR?

[jkotas]

@dhartglassMSFT is going to review this PR

[clamp03]

@dhartglassMSFT Could you please review this PR when you get a chance?

[dhartglassMSFT]

Should the if be checking the inverse? This seems like a typo and we want if(varTypeIsIntegral... without the !

or get rid of the if, and add the assert message to the TypeIsFloating assert, seems like we don't need both of these

[dhartglassMSFT]

I think the following code in gentree.cpp must also be updated - as-is something strange could happen if we try to generate a MOD/UMOD helper call when "USE_HELPERS_FOR_INT_DIV == 0"

in void GenTree::InitNodeSize()

#if USE_HELPERS_FOR_INT_DIV
    GenTree::s_gtNodeSizes[GT_DIV]           = TREE_NODE_SZ_LARGE;
    GenTree::s_gtNodeSizes[GT_UDIV]          = TREE_NODE_SZ_LARGE;
    GenTree::s_gtNodeSizes[GT_MOD]           = TREE_NODE_SZ_LARGE;
    GenTree::s_gtNodeSizes[GT_UMOD]          = TREE_NODE_SZ_LARGE;
#endif

the GT_MOD/GT_UMOD cases should be under the ARM32 target ifdef

[am11]

Our bar for linux arm32 is to keep it functional. We are not spending any cycles to make it better. I do not think we are interested in investing into new arm32-specific optimizations.

BTW, that hwcap suggestion is about changing buildtime check to runtime check with a single if HWCAP_IDIV, rest of the code is already in JIT (under #if !USE_HELPERS_FOR_INT_DIV).