inline asm issue

[salcideon]

using the toolchain:

static inline float32x4_t
test(float32x4_t _acc_, float32x4_t _op1_, float32x4_t _op2_, int _lane_)
{
    __asm__ __volatile__("fmla %0.4s, %1.4s, %2.s[%3]" : "+w"(_acc_) : "w"(_op1_), "w"(_op2_), "I"(_lane_));
    return _acc_;
}

generates fmla q21.4s, q19.4s, q27.s[#1] instead of fmla q21.4s, q19.4s, q27.s[1] as it should.

Then compilation fails with error: unexpected token in argument list

[pirama-arumuga-nainar]

Below is a simpler repro. The original fails because lane is a variable instead of a constant (presumably it was a constant once inlined).

void test() {
  __asm__ ("fmla v2.4s, v0.4s, v1.s[%0]" :: "I"(0x1));
}

Fix under review: https://reviews.llvm.org/D65550

[pirama-arumuga-nainar]

From the LLVM review linked above, another option is to use the 'c' modifier for the lane specifier, which is supported by gcc as well. So if we change the above repro to

void test() {
  __asm__ ("fmla v2.4s, v0.4s, v1.s[%c0]" :: "I"(0x1));
}

it works.

[pirama-arumuga-nainar]

The upstream change was merged and will be part of r21.