[AArch64] llvm has more add intructions to update the array index

[vfdff]

• https://gcc.godbolt.org/z/8rcKrrr33

  
  #include <complex.h>
  #include <math.h>
  using namespace std;

void cmul2 (complex __restrict a, complex restrict b, complex * restrict c) { for (int i=0; i<1000; i++) { c[i] = a[i] * b[i]; } }

* gcc:

cmul2(std::complex, std::complex, std::complex*): mov x3, 16000 add x3, x2, x3 .L2: ld2 {v30.2d - v31.2d}, [x0], 32 ld2 {v28.2d - v29.2d}, [x1], 32 fmul v27.2d, v31.2d, v28.2d fmul v26.2d, v30.2d, v28.2d fmla v27.2d, v30.2d, v29.2d fmls v26.2d, v29.2d, v31.2d st2 {v26.2d - v27.2d}, [x2], 32 cmp x3, x2 bne .L2 ret

*llvm:

cmul2(std::complex, std::complex, std::complex): // @cmul2(std::complex, std::complex, std::complex) mov x8, #-16000 // =0xffffffffffffc180 mov w9, #16000 // =0x3e80 .LBB0_1: // =>This Inner Loop Header: Depth=1 add x10, x0, x8 add x11, x1, x8 add x10, x10, x9 add x11, x11, x9 ld2 { v0.2d, v1.2d }, [x10] ld2 { v2.2d, v3.2d }, [x11] add x10, x2, x8 add x10, x10, x9 adds x8, x8, #32 fmul v4.2d, v2.2d, v0.2d fmul v5.2d, v2.2d, v1.2d fmls v4.2d, v3.2d, v1.2d fmla v5.2d, v3.2d, v0.2d st2 { v4.2d, v5.2d }, [x10] b.ne .LBB0_1 ret

* it seems the issue related to type **std::complex<double>**, when we change it to **int**, the issue doesn't apprear
,https://gcc.godbolt.org/z/zzxeKPsvo

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Allen (vfdff)

* https://gcc.godbolt.org/z/8rcKrrr33 ``` #include <complex.h> #include <math.h> using namespace std; void cmul2 (complex<double> * __restrict a, complex<double> * __restrict b, complex<double> * __restrict c) { for (int i=0; i<1000; i++) { c[i] = a[i] * b[i]; } } ``` * gcc: ``` cmul2(std::complex<double>*, std::complex<double>*, std::complex<double>*): mov x3, 16000 add x3, x2, x3 .L2: ld2 {v30.2d - v31.2d}, [x0], 32 ld2 {v28.2d - v29.2d}, [x1], 32 fmul v27.2d, v31.2d, v28.2d fmul v26.2d, v30.2d, v28.2d fmla v27.2d, v30.2d, v29.2d fmls v26.2d, v29.2d, v31.2d st2 {v26.2d - v27.2d}, [x2], 32 cmp x3, x2 bne .L2 ret ``` *llvm: ``` cmul2(std::complex<double>*, std::complex<double>*, std::complex<double>*): // @cmul2(std::complex<double>*, std::complex<double>*, std::complex<double>*) mov x8, #-16000 // =0xffffffffffffc180 mov w9, #16000 // =0x3e80 .LBB0_1: // =>This Inner Loop Header: Depth=1 add x10, x0, x8 add x11, x1, x8 add x10, x10, x9 add x11, x11, x9 ld2 { v0.2d, v1.2d }, [x10] ld2 { v2.2d, v3.2d }, [x11] add x10, x2, x8 add x10, x10, x9 adds x8, x8, #32 fmul v4.2d, v2.2d, v0.2d fmul v5.2d, v2.2d, v1.2d fmls v4.2d, v3.2d, v1.2d fmla v5.2d, v3.2d, v0.2d st2 { v4.2d, v5.2d }, [x10] b.ne .LBB0_1 ret ``` * it seems the issue related to type **std::complex<double>**, when we change it to **int**, the issue doesn't apprear ,https://gcc.godbolt.org/z/zzxeKPsvo

[vfdff]

• a similar but simple case, https://gcc.godbolt.org/z/eoqdhEaxK

[madhur13490]

Have you measured the perf difference?