[aarch64] Prefer use callee save register to avoid repeat load when cross call

[vfdff]

• test: https://gcc.godbolt.org/z/hdPf8veEc

float foo (float num[], float r2inv, int n) {
   float sum = 0.0;
   for (int i=0; i < 1000; i++) {
       float a = num[i];
       // const float r = a / std::sqrt (r2inv);
       const float expm2 = std::exp (a);

       float tmp = expm2 * num[i];
       sum += tmp * tmp;
   }
    
    return sum;
}

• llvm: When built with O3, use 2 load to get num[i] before and after the call expf

.LBB0_1:                                // =>This Inner Loop Header: Depth=1
        ldr     s0, [x19, x20]  // load num[i]
        bl      expf
        ldr     s1, [x19, x20]   // repeat load the num[i]
        add     x20, x20, #4
        cmp     x20, #4000
        fmul    s0, s0, s1
        fmadd   s8, s0, s0, s8
        b.ne    .LBB0_1

• gcc: Only one load for num[i] because it use the callee save register

.L2:
        ldr     s15, [x19], 4
        fmov    s0, s15
        bl      expf
        fmul    s15, s15, s0
        fmadd   s14, s15, s15, s14
        cmp     x19, x20
        bne     .L2

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Allen (vfdff)

* test: https://gcc.godbolt.org/z/hdPf8veEc ``` float foo (float num[], float r2inv, int n) { float sum = 0.0; for (int i=0; i < 1000; i++) { float a = num[i]; // const float r = a / std::sqrt (r2inv); const float expm2 = std::exp (a);

   float tmp = expm2 * num[i];
   sum += tmp * tmp;

}

return sum;

}

* llvm:  When built with O3, use 2 load to get num[i] **before** and **after** the call expf

.LBB0_1: // =>This Inner Loop Header: Depth=1
ldr s0, [x19, x20] // load num[i]
bl expf
ldr s1, [x19, x20] // repeat load the num[i]
add x20, x20, #4
cmp x20, #4000
fmul s0, s0, s1
fmadd s8, s0, s0, s8
b.ne .LBB0_1

* gcc: Only one load for num[i] because it use the callee save register

.L2:
ldr s15, [x19], 4
fmov s0, s15
bl expf
fmul s15, s15, s0
fmadd s14, s15, s15, s14
cmp x19, x20
bne .L2

</details>

[jroelofs]

I bet this is alias-analysis related, since adding __restrict fixes it: https://gcc.godbolt.org/z/s9x7hcvrM

Marking the called function pure also fixes it: https://gcc.godbolt.org/z/vedjadMh3

[vfdff]

Thanks @jroelofs for your input
so it more likely gcc will add pure attribute for buildin math function? gcc also use 2 load if we delete the pure attribute for fn,
https://gcc.godbolt.org/z/f1W98eehq

[vfdff]

• It seems math libraries will access (read & write) global variable math-errno if without -ffast-math, so they are not pure in that case, and it is more likely a alias-analysis related.

• I find function Sema::AddKnownFunctionAttributes will add const attribute for implicitly-declared builtins, but we still don't get some difference between with/without the -ffast-math
  

  llvm-project/clang/lib/Sema/SemaDecl.cpp

  Line 16741 in 78ee473

  FD->addAttr(ConstAttr::CreateImplicit(Context, FD->getLocation()));

[vfdff]

Add option -fno-math-errno can avoid the issue for clang, https://gcc.godbolt.org/z/hcqrWhhMb