On 5/20/19 6:00 AM, Christophe Lyon wrote:
> Hi,
>
> As discussed in the PR, the new cdce3.c test fails on
> arm-linux-gnueabi, and passes on arm-linux-gnueabihf.
>
> On arm-linux-gnueabi, the generated code is:
> .arm
> .fpu softvfp
> .type foo, %function
> foo:
> b sqrtf
>
> on arm-liunx-gnueabihf, it is:
> .arm
> .fpu neon
> .type foo, %function
> foo:
> vcmp.f32 s0, #0
> vmrs APSR_nzcv, FPSCR
> vmov.f32 s15, s0
> bmi .L10
> vsqrt.f32 s0, s15
> bx lr
> .L10:
> b sqrtf
>
> Actually both generate a tail-call, but only the latter says "function
> call is shrink-wrapped into error conditions." in the dumps.
>
> According to what JunMa says in the PR, it seems we want to restrict
> this testcase to hard-float cases, so that's what this small patch
> does.
>
> OK?
OK
jeff
