https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90235
--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to H.J. Lu from comment #0)
> From PR 90202:
>
> [hjl@gnu-cfl-1 pr90202]$ cat x.ii
> struct v {
> int val[16];
> };
>
> struct v test(struct v a, struct v b) {
> struct v res;
>
> for (int i = 0; i < 16; i++)
> res.val[i] = a.val[i] + b.val[i];
>
> return res;
> }
> [hjl@gnu-cfl-1 pr90202]$ make CC=gcc
> gcc -O3 -march=skylake -S x.ii
> [hjl@gnu-cfl-1 pr90202]$ cat x.s
> .file "x.ii"
> .text
> .p2align 4,,15
> .globl _Z4test1vS_
> .type _Z4test1vS_, @function
> _Z4test1vS_:
> .LFB0:
> .cfi_startproc
> pushq %rbp
> .cfi_def_cfa_offset 16
> .cfi_offset 6, -16
> movq %rdi, %rax
> movq %rsp, %rbp
> .cfi_def_cfa_register 6
> vmovdqu 16(%rbp), %ymm1
> vmovdqu 48(%rbp), %ymm2
> vpaddd 80(%rbp), %ymm1, %ymm0
> vmovdqu %ymm0, (%rdi)
> vpaddd 112(%rbp), %ymm2, %ymm0
> vmovdqu %ymm0, 32(%rdi)
> vzeroupper
> popq %rbp
> .cfi_def_cfa 7, 8
> ret
> .cfi_endproc
>
> Since there is
>
> rtx
> gen_reg_rtx (machine_mode mode)
> {
> rtx val;
> unsigned int align = GET_MODE_ALIGNMENT (mode);
>
> gcc_assert (can_create_pseudo_p ());
>
> /* If a virtual register with bigger mode alignment is generated,
> increase stack alignment estimation because it might be spilled
> to stack later. */
> if (SUPPORTS_STACK_ALIGNMENT
> && crtl->stack_alignment_estimated < align
> && !crtl->stack_realign_processed)
> {
> unsigned int min_align = MINIMUM_ALIGNMENT (NULL, mode, align);
> if (crtl->stack_alignment_estimated < min_align)
> crtl->stack_alignment_estimated = min_align;
> }
>
> and IRA has
>
> frame_pointer_needed
> = (! flag_omit_frame_pointer
> || (cfun->calls_alloca && EXIT_IGNORE_STACK)
> /* We need the frame pointer to catch stack overflow exceptions if
> the stack pointer is moving (as for the alloca case just above).
> */
> || (STACK_CHECK_MOVING_SP
> && flag_stack_check
> && flag_exceptions
> && cfun->can_throw_non_call_exceptions)
> || crtl->accesses_prior_frames
> || (SUPPORTS_STACK_ALIGNMENT && crtl->stack_realign_needed)
> || targetm.frame_pointer_required ());
>
> generate AVX/AVX512 pseudo registers via gen_reg_rtx will mark frame
> pointer as needed. Stack realignment is needed to
>
> 1. Align the outgoing stack.
> 2. Support aligned spill of AVX/AVX512 registers.
>
> But we won't know if spill is needed before RA. As the result, we
> save and restore frame pointer even if not needed. Since
>
> (define_insn "mov<mode>_internal"
> [(set (match_operand:VMOVE 0 "nonimmediate_operand"
> "=v,v ,v ,m")
> (match_operand:VMOVE 1 "nonimmediate_or_sse_const_operand"
> " C,BC,vm,v"))]
> "TARGET_SSE
> && (register_operand (operands[0], <MODE>mode)
> || register_operand (operands[1], <MODE>mode))"
>
> now supports both aligned and unaligned load/store of AVX/AVX512
> registers, we can change gen_reg_rtx to
>
> /* If a virtual register with bigger mode alignment is generated,
> increase stack alignment estimation because it might be spilled
> to stack later. */
> if (SUPPORTS_STACK_ALIGNMENT
> && !SUPPORTS_MISALIGNED_SPILL
> && crtl->stack_alignment_estimated < align
> && !crtl->stack_realign_processed)
> {
> unsigned int min_align = MINIMUM_ALIGNMENT (NULL, mode, align);
> if (crtl->stack_alignment_estimated < min_align)
> crtl->stack_alignment_estimated = min_align;
> }
Would this generate more unaligned_loads/stores, and does harm to performance?
