On Thu, Jul 20, 2023 at 9:44 AM Roger Sayle <ro...@nextmovesoftware.com> wrote:
>
>
> Hi Uros,
>
> > From: Uros Bizjak <ubiz...@gmail.com>
> > Sent: 20 July 2023 07:50
> >
> > On Wed, Jul 19, 2023 at 10:07 PM Roger Sayle <ro...@nextmovesoftware.com>
> > wrote:
> > >
> > > This patch is the next piece of a solution to the x86_64 ABI issues in
> > > PR 88873. This splits the *concat<mode><dwi>3_3 define_insn_and_split
> > > into two patterns, a TARGET_64BIT *concatditi3_3 and a !TARGET_64BIT
> > > *concatsidi3_3. This allows us to add an additional alternative to
> > > the the 64-bit version, enabling the register allocator to perform
> > > this operation using SSE registers, which is implemented/split after
> > > reload using vec_concatv2di.
> > >
> > > To demonstrate the improvement, the test case from PR88873:
> > >
> > > typedef struct { double x, y; } s_t;
> > >
> > > s_t foo (s_t a, s_t b, s_t c)
> > > {
> > > return (s_t){ __builtin_fma(a.x, b.x, c.x), __builtin_fma (a.y, b.y,
> > > c.y) }; }
> > >
> > > when compiled with -O2 -march=cascadelake, currently generates:
> > >
> > > foo: vmovq %xmm2, -56(%rsp)
> > > movq -56(%rsp), %rax
> > > vmovq %xmm3, -48(%rsp)
> > > vmovq %xmm4, -40(%rsp)
> > > movq -48(%rsp), %rcx
> > > vmovq %xmm5, -32(%rsp)
> > > vmovq %rax, %xmm6
> > > movq -40(%rsp), %rax
> > > movq -32(%rsp), %rsi
> > > vpinsrq $1, %rcx, %xmm6, %xmm6
> > > vmovq %xmm0, -24(%rsp)
> > > vmovq %rax, %xmm7
> > > vmovq %xmm1, -16(%rsp)
> > > vmovapd %xmm6, %xmm2
> > > vpinsrq $1, %rsi, %xmm7, %xmm7
> > > vfmadd132pd -24(%rsp), %xmm7, %xmm2
> > > vmovapd %xmm2, -56(%rsp)
> > > vmovsd -48(%rsp), %xmm1
> > > vmovsd -56(%rsp), %xmm0
> > > ret
> > >
> > > with this change, we avoid many of the reloads via memory,
> > >
> > > foo: vpunpcklqdq %xmm3, %xmm2, %xmm7
> > > vpunpcklqdq %xmm1, %xmm0, %xmm6
> > > vpunpcklqdq %xmm5, %xmm4, %xmm2
> > > vmovdqa %xmm7, -24(%rsp)
> > > vmovdqa %xmm6, %xmm1
> > > movq -16(%rsp), %rax
> > > vpinsrq $1, %rax, %xmm7, %xmm4
> > > vmovapd %xmm4, %xmm6
> > > vfmadd132pd %xmm1, %xmm2, %xmm6
> > > vmovapd %xmm6, -24(%rsp)
> > > vmovsd -16(%rsp), %xmm1
> > > vmovsd -24(%rsp), %xmm0
> > > ret
> > >
> > >
> > > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> > > and make -k check, both with and without --target_board=unix{-m32}
> > > with no new failures. Ok for mainline?
> > >
> > >
> > > 2023-07-19 Roger Sayle <ro...@nextmovesoftware.com>
> > >
> > > gcc/ChangeLog
> > > * config/i386/i386-expand.cc (ix86_expand_move): Don't call
> > > force_reg, to use SUBREG rather than create a new pseudo when
> > > inserting DFmode fields into TImode with insvti_{high,low}part.
> > > (*concat<mode><dwi>3_3): Split into two define_insn_and_split...
> > > (*concatditi3_3): 64-bit implementation. Provide alternative
> > > that allows register allocation to use SSE registers that is
> > > split into vec_concatv2di after reload.
> > > (*concatsidi3_3): 32-bit implementation.
> > >
> > > gcc/testsuite/ChangeLog
> > > * gcc.target/i386/pr88873.c: New test case.
> >
> > diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> > index f9b0dc6..9c3febe 100644
> > --- a/gcc/config/i386/i386-expand.cc
> > +++ b/gcc/config/i386/i386-expand.cc
> > @@ -558,7 +558,7 @@ ix86_expand_move (machine_mode mode, rtx
> > operands[])
> > op0 = SUBREG_REG (op0);
> > tmp = gen_rtx_AND (TImode, copy_rtx (op0), tmp);
> > if (mode == DFmode)
> > - op1 = force_reg (DImode, gen_lowpart (DImode, op1));
> > + op1 = gen_lowpart (DImode, op1);
> >
> > Please note that gen_lowpart will ICE when op1 is a SUBREG. This is the
> > reason
> > that we need to first force a SUBREG to a register and then perform
> > gen_lowpart,
> > and it is necessary to avoid ICE.
>
> The good news is that we know op1 is a register, as this is tested by
> "&& REG_P (op1)" on line 551. You'll also notice that I'm not removing
> the force_reg from before the call to gen_lowpart, but removing the call
> to force_reg after the call to gen_lowpart. When I originally wrote this,
> the hope was that placing this SUBREG in its own pseudo would help
> with register allocation/CSE. Unfortunately, increasing the number of
> pseudos (in this case) increases compile-time (due to quadratic behaviour
> in LRA), as shown by PR rtl-optimization/110587, and keeping the DF->DI
> conversion in a SUBREG inside the insvti_{high,low}part allows the
> register allocator to see the DF->DI->TI sequence in a single pattern,
> and hence choose to keep the TI mode in SSE registers, rather than use
> a pair of reloads, to write the DF value to memory, then read it back as
> a scalar in DImode, and perhaps the same again to go the other way.
This was my only concern with the patch, with that cleared, the patch is OK.
Thanks,
Uros.
