Hi All, The cost model is currently treating multiplication by element as being more expensive than 3 same multiplication. This means that if the value is on the SIMD side we add an unneeded DUP. If the value is on the genreg side we use the more expensive DUP instead of fmov.
This patch corrects the costs such that the two multiplies are costed the same which allows us to generate fmul v3.4s, v3.4s, v0.s[0] instead of dup v0.4s, v0.s[0] fmul v3.4s, v3.4s, v0.4s Bootstrapped Regtested on aarch64-none-linux-gnu and no issues. Ok for master? Thanks, Tamar gcc/ChangeLog: * config/aarch64/aarch64.c (aarch64_rtx_mult_cost): Adjust costs for mul. gcc/testsuite/ChangeLog: * gcc.target/aarch64/asimd-mull-elem.c: New test. --
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index 97da60762390db81df9cffaf316b909cd1609130..9cc8da338125afa01bc9fb645f4112d2d7ef548c 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -11279,6 +11279,14 @@ aarch64_rtx_mult_cost (rtx x, enum rtx_code code, int outer, bool speed) if (VECTOR_MODE_P (mode)) mode = GET_MODE_INNER (mode); + /* The by element versions of the instruction has the same costs as the + normal 3 vector version. So don't add the costs of the duplicate into + the costs of the multiply. */ + if (GET_CODE (op0) == VEC_DUPLICATE) + op0 = XEXP (op0, 0); + else if (GET_CODE (op1) == VEC_DUPLICATE) + op1 = XEXP (op1, 0); + /* Integer multiply/fma. */ if (GET_MODE_CLASS (mode) == MODE_INT) { diff --git a/gcc/testsuite/gcc.target/aarch64/asimd-mull-elem.c b/gcc/testsuite/gcc.target/aarch64/asimd-mull-elem.c new file mode 100644 index 0000000000000000000000000000000000000000..513721cee0c8372781e6daf33bc06e256cab8cb8 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/asimd-mull-elem.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-require-effective-target vect_int } */ +/* { dg-require-effective-target vect_float } */ +/* { dg-options "-Ofast" } */ + +#include <arm_neon.h> + +void s_mult_i (int32_t* restrict res, int32_t* restrict a, int32_t b) +{ + for (int x = 0; x < 16; x++) + res[x] = a[x] * b; +} + +void s_mult_f (float32_t* restrict res, float32_t* restrict a, float32_t b) +{ + for (int x = 0; x < 16; x++) + res[x] = a[x] * b; +} + +/* { dg-final { scan-assembler-times {\s+mul\tv[0-9]+\.4s, v[0-9]+\.4s, v[0-9]+\.s\[0\]} 4 } } */ +/* { dg-final { scan-assembler-times {\s+fmul\tv[0-9]+\.4s, v[0-9]+\.4s, v[0-9]+\.s\[0\]} 4 } } */