On Thu, May 13, 2021 at 11:18 AM Hongtao Liu <crazy...@gmail.com> wrote:
>
> Hi:
> When __builtin_ia32_vzeroupper is called explicitly, the corresponding
> vzeroupper pattern does not carry any CLOBBERS or SETs before LRA,
> which leads to incorrect optimization in pass_reload.
> In order to solve this problem, this patch introduces a pre_reload
> splitter which adds CLOBBERS to vzeroupper's pattern, it can solve the
> problem in pr.
>
> At the same time, in order to optimize the low 128 bits in
> post_reload CSE, this patch also transforms those CLOBBERS to SETs in
> pass_vzeroupper.
>
> It works fine except for TARGET_64BIT_MS_ABI, under which xmm6-xmm15
> are callee-saved, so even if there're no other uses of xmm6-xmm15 in the
> function, because of vzeroupper's pattern, pro_epilog will save and
> restore those registers, which is obviously redundant. In order to
> eliminate this redundancy, a post_reload splitter is introduced, which
> drops those SETs, until epilogue_completed splitter adds those SETs
> back, it looks to be safe since there's no CSE between post_reload
> split2 and epilogue_completed split3??? Also frame info needs to be
> updated in pro_epilog, which saves and restores xmm6-xmm15 only if
> there's usage other than explicit vzeroupper pattern.
>
> Bootstrapped and regtested on X86_64-linux-gnu{-m32,}
> Ok for trunk?
Some time ago a support for CLOBBER_HIGH RTX was added (and later
removed for some reason). Perhaps we could resurrect the patch for the
purpose of ferrying 128bit modes via vzeroupper RTX?
+(define_split
+ [(match_parallel 0 "vzeroupper_pattern"
+ [(unspec_volatile [(const_int 0)] UNSPECV_VZEROUPPER)])]
+ "TARGET_AVX && ix86_pre_reload_split ()"
+ [(match_dup 0)]
+{
+ /* When vzeroupper is explictly used, for LRA purpose, make it clear
+ the instruction kills sse registers. */
+ gcc_assert (cfun->machine->has_explicit_vzeroupper);
+ unsigned int nregs = TARGET_64BIT ? 16 : 8;
+ rtvec vec = rtvec_alloc (nregs + 1);
+ RTVEC_ELT (vec, 0) = gen_rtx_UNSPEC_VOLATILE (VOIDmode,
+ gen_rtvec (1, const1_rtx),
+ UNSPECV_VZEROUPPER);
+ for (unsigned int i = 0; i < nregs; ++i)
+ {
+ unsigned int regno = GET_SSE_REGNO (i);
+ rtx reg = gen_rtx_REG (V2DImode, regno);
+ RTVEC_ELT (vec, i + 1) = gen_rtx_CLOBBER (VOIDmode, reg);
+ }
+ operands[0] = gen_rtx_PARALLEL (VOIDmode, vec);
+})
Wouldn't this also kill lower 128bit values that are not touched by
vzeroupper? A CLOBBER_HIGH would be more appropriate here.
Uros.
> gcc/ChangeLog:
>
> PR target/82735
> * config/i386/i386-expand.c (ix86_expand_builtin): Count
> number of __builtin_ia32_vzeroupper.
> * config/i386/i386-features.c (ix86_add_reg_usage_to_vzerouppers):
> Transform CLOBBERs to SETs for explicit vzeroupper pattern so
> that CSE can optimize lower 128 bits.
> * config/i386/i386.c (ix86_handle_explicit_vzeroupper_in_pro_epilog):
> New.
> (ix86_save_reg): If there's no use of xmm6~xmm15 other than
> explicit vzeroupper under TARGET_64BIT_MS_ABI, no need to save
> REGNO.
> (ix86_finalize_stack_frame_flags): Recompute frame layout if
> there's explicit vzeroupper under TARGET_64BIT_MS_ABI.
> * config/i386/i386.h (struct machine_function): Change type of
> has_explicit_vzeroupper from BOOL_BITFILED to unsigned int.
> * config/i386/sse.md (*avx_vzeroupper_2): New post-reload
> splitter which will drop all SETs for explicit vzeroupper
> patterns.
> (*avx_vzeroupper_1): Generate SET reg to reg instead of
> CLOBBER, and add pre-reload splitter after it.
>
> gcc/testsuite/ChangeLog:
>
> PR target/82735
> * gcc.target/i386/pr82735-1.c: New test.
> * gcc.target/i386/pr82735-2.c: New test.
> * gcc.target/i386/pr82735-3.c: New test.
> * gcc.target/i386/pr82735-4.c: New test.
> * gcc.target/i386/pr82735-5.c: New test.
>
>
> --
> BR,
> Hongtao
