AArch64 RTX cost for vector SET is causing PR65375. Lower subreg is using this rtx_cost to compute the cost of moves, and splitting anything larger than word size, 64-bits in this case. The aarch64 rtx_costs is returning 2 * COST_N_INSNS(1) for vector moves, so they get split. Attach patch fixes this.
With the patch the testcase in the PR: #include <arm_neon.h> void hello_vst2(float* fout, float *fin) { float32x4x2_t a; a = vld2q_f32 (fin); vst2q_f32 (fout, a); } Changes to: hello_vst2: - ld2 {v0.4s - v1.4s}, [x1] - sub sp, sp, #32 - umov x1, v0.d[0] - umov x2, v0.d[1] - str q1, [sp, 16] - mov x5, x1 - stp x5, x2, [sp] - ld1 {v0.16b - v1.16b}, [sp] + ld2 {v2.4s - v3.4s}, [x1] + orr v0.16b, v2.16b, v2.16b + orr v1.16b, v3.16b, v3.16b st2 {v0.4s - v1.4s}, [x0] - add sp, sp, 32 ret lower-subreg.c:compute_costs() only cares about the cost of a (set (reg) (const_int )) move but I think the intention, at least for now, is to return extra_cost->vect.alu for all the vector operations. Regression tested on aarch64-linux-gnu with no new regression. Is this OK for trunk? Thanks, Kugan gcc/ChangeLog: 2015-03-16 Kugan Vivekanandarajah <kug...@linaro.org> Jim Wilson <jim.wil...@linaro.org> PR target/65375 * config/aarch64/aarch64.c (aarch64_rtx_costs): Return extra_cost->vect.alu for SET.
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index cba3c1a..db69979 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -5517,6 +5517,13 @@ aarch64_rtx_costs (rtx x, int code, int outer ATTRIBUTE_UNUSED, op0 = SET_DEST (x); op1 = SET_SRC (x); + /* Sets don't have a mode, so we must recompute this here. */ + if (VECTOR_MODE_P (GET_MODE (op0))) + { + *cost += extra_cost->vect.alu; + return true; + } + switch (GET_CODE (op0)) { case MEM: