[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v3 1/5] target-ppc: add vector insert instructio
From: |
David Gibson |
Subject: |
Re: [Qemu-devel] [PATCH v3 1/5] target-ppc: add vector insert instructions |
Date: |
Tue, 23 Aug 2016 11:08:15 -0400 |
User-agent: |
Mutt/1.6.2 (2016-07-01) |
On Fri, Aug 19, 2016 at 11:16:03AM +0530, Rajalakshmi Srinivasaraghavan wrote:
>
>
> On 08/16/2016 09:48 AM, David Gibson wrote:
> > On Thu, Aug 11, 2016 at 01:06:44PM +0530, Rajalakshmi Srinivasaraghavan
> > wrote:
> > > The following vector insert instructions are added from ISA 3.0.
> > >
> > > vinsertb - Vector Insert Byte
> > > vinserth - Vector Insert Halfword
> > > vinsertw - Vector Insert Word
> > > vinsertd - Vector Insert Doubleword
> > >
> > > Signed-off-by: Rajalakshmi Srinivasaraghavan <address@hidden>
> > > ---
> > > target-ppc/helper.h | 4 ++++
> > > target-ppc/int_helper.c | 27 +++++++++++++++++++++++++++
> > > target-ppc/translate.c | 2 ++
> > > target-ppc/translate/vmx-impl.c | 32 ++++++++++++++++++++++++++++++++
> > > target-ppc/translate/vmx-ops.c | 18 +++++++++++++-----
> > > 5 files changed, 78 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/target-ppc/helper.h b/target-ppc/helper.h
> > > index 93ac9e1..0923779 100644
> > > --- a/target-ppc/helper.h
> > > +++ b/target-ppc/helper.h
> > > @@ -250,6 +250,10 @@ DEF_HELPER_2(vspltisw, void, avr, i32)
> > > DEF_HELPER_3(vspltb, void, avr, avr, i32)
> > > DEF_HELPER_3(vsplth, void, avr, avr, i32)
> > > DEF_HELPER_3(vspltw, void, avr, avr, i32)
> > > +DEF_HELPER_3(vinsertb, void, avr, avr, i32)
> > > +DEF_HELPER_3(vinserth, void, avr, avr, i32)
> > > +DEF_HELPER_3(vinsertw, void, avr, avr, i32)
> > > +DEF_HELPER_3(vinsertd, void, avr, avr, i32)
> > > DEF_HELPER_2(vupkhpx, void, avr, avr)
> > > DEF_HELPER_2(vupklpx, void, avr, avr)
> > > DEF_HELPER_2(vupkhsb, void, avr, avr)
> > > diff --git a/target-ppc/int_helper.c b/target-ppc/int_helper.c
> > > index 552b2e0..3f8e439 100644
> > > --- a/target-ppc/int_helper.c
> > > +++ b/target-ppc/int_helper.c
> > > @@ -1792,6 +1792,33 @@ VSPLT(w, u32)
> > > #undef VSPLT
> > > #undef SPLAT_ELEMENT
> > > #undef _SPLAT_MASKED
> > > +#if defined(HOST_WORDS_BIGENDIAN)
> > > +#define VINSERT(suffix, element, index)
> > > \
> > > + void helper_vinsert##suffix(ppc_avr_t *r, ppc_avr_t *b, uint32_t
> > > splat) \
> > > + {
> > > \
> > > + ppc_avr_t result;
> > > \
> > > + result = *r;
> > > \
> > > + memcpy(&result.u8[splat], &b->element[index],
> > > \
> > > + sizeof(result.element[0]));
> > > \
> > > + *r = result;
> > > \
> > Using a temporary for the result means two extra full vector copies,
> > which seems unfortunate. Couldn't you just use memmove() instead of
> > memcpy() to handle the overlapping cases?
> If the registers r and b are same, using memove() can overwrite
> some of the values depending on index(third arg).
You're only transferring a single element, and memmove() can handle
overlapping source and destination, so I'm not seeing the problem even
if r == b.
> > > + }
> > > +#else
> > > +#define VINSERT(suffix, element, index)
> > > \
> > > + void helper_vinsert##suffix(ppc_avr_t *r, ppc_avr_t *b, uint32_t
> > > splat) \
> > > + {
> > > \
> > > + ppc_avr_t result;
> > > \
> > > + result = *r;
> > > \
> > > + uint32_t s = (ARRAY_SIZE(r->element) - index) - 1;
> > > \
> > The logic with index seems a bit convoluted. AFAICT the index is
> > always the least significant element of the most-significant half of
> > the vector b. So for BE &result.u8[8] should always be right and for
> > LE &result.u8[8 - sizeof(r->element)].
> Ack. One change in your comment for BE its u8[8 - sizeof(r->element)]
> and LE its u8[8].
Ah, yes, sorry.
>
> >
> > > + uint32_t d = (16 - splat) - sizeof(r->element[0]);
> > > \
> > > + memcpy(&result.u8[d], &b->element[s],
> > > sizeof(result.element[0])); \
> > > + *r = result;
> > > \
> > > + }
> > > +#endif
> > > +VINSERT(b, u8, 7)
> > > +VINSERT(h, u16, 3)
> > > +VINSERT(w, u32, 1)
> > > +VINSERT(d, u64, 0)
> > > +#undef VINSERT
> > > #define VSPLTI(suffix, element, splat_type) \
> > > void helper_vspltis##suffix(ppc_avr_t *r, uint32_t splat) \
> > > diff --git a/target-ppc/translate.c b/target-ppc/translate.c
> > > index fc3d371..dbe952e 100644
> > > --- a/target-ppc/translate.c
> > > +++ b/target-ppc/translate.c
> > > @@ -498,6 +498,8 @@ EXTRACT_HELPER(UIMM, 0, 16);
> > > EXTRACT_HELPER(SIMM5, 16, 5);
> > > /* 5 bits signed immediate value */
> > > EXTRACT_HELPER(UIMM5, 16, 5);
> > > +/* 4 bits unsigned immediate value */
> > > +EXTRACT_HELPER(UIMM4, 16, 4);
> > > /* Bit count */
> > > EXTRACT_HELPER(NB, 11, 5);
> > > /* Shift count */
> > > diff --git a/target-ppc/translate/vmx-impl.c
> > > b/target-ppc/translate/vmx-impl.c
> > > index ac78caf..f6a97ac 100644
> > > --- a/target-ppc/translate/vmx-impl.c
> > > +++ b/target-ppc/translate/vmx-impl.c
> > > @@ -623,13 +623,45 @@ static void glue(gen_, name)(DisasContext *ctx)
> > > \
> > > tcg_temp_free_ptr(rd);
> > > \
> > > }
> > > +#define GEN_VXFORM_UIMM_SPLAT(name, opc2, opc3, splat_max) \
> > > +static void glue(gen_, name)(DisasContext *ctx) \
> > > + { \
> > > + TCGv_ptr rb, rd; \
> > > + uint8_t uimm = UIMM4(ctx->opcode); \
> > > + TCGv_i32 t0 = tcg_temp_new_i32(); \
> > > + if (unlikely(!ctx->altivec_enabled)) { \
> > > + gen_exception(ctx, POWERPC_EXCP_VPU); \
> > > + return; \
> > > + } \
> > > + if (uimm > splat_max) { \
> > > + uimm = 0; \
> > > + } \
> > > + tcg_gen_movi_i32(t0, uimm); \
> > > + rb = gen_avr_ptr(rB(ctx->opcode)); \
> > > + rd = gen_avr_ptr(rD(ctx->opcode)); \
> > > + gen_helper_##name(rd, rb, t0); \
> > > + tcg_temp_free_i32(t0); \
> > > + tcg_temp_free_ptr(rb); \
> > > + tcg_temp_free_ptr(rd); \
> > > + }
> > > +
> > > GEN_VXFORM_UIMM(vspltb, 6, 8);
> > > GEN_VXFORM_UIMM(vsplth, 6, 9);
> > > GEN_VXFORM_UIMM(vspltw, 6, 10);
> > > +GEN_VXFORM_UIMM_SPLAT(vinsertb, 6, 12, 15);
> > > +GEN_VXFORM_UIMM_SPLAT(vinserth, 6, 13, 14);
> > > +GEN_VXFORM_UIMM_SPLAT(vinsertw, 6, 14, 12);
> > > +GEN_VXFORM_UIMM_SPLAT(vinsertd, 6, 15, 8);
> > > GEN_VXFORM_UIMM_ENV(vcfux, 5, 12);
> > > GEN_VXFORM_UIMM_ENV(vcfsx, 5, 13);
> > > GEN_VXFORM_UIMM_ENV(vctuxs, 5, 14);
> > > GEN_VXFORM_UIMM_ENV(vctsxs, 5, 15);
> > > +GEN_VXFORM_DUAL(vspltisb, PPC_NONE, PPC2_ALTIVEC_207,
> > > + vinsertb, PPC_NONE, PPC2_ISA300);
> > > +GEN_VXFORM_DUAL(vspltish, PPC_NONE, PPC2_ALTIVEC_207,
> > > + vinserth, PPC_NONE, PPC2_ISA300);
> > > +GEN_VXFORM_DUAL(vspltisw, PPC_NONE, PPC2_ALTIVEC_207,
> > > + vinsertw, PPC_NONE, PPC2_ISA300);
> > > static void gen_vsldoi(DisasContext *ctx)
> > > {
> > > diff --git a/target-ppc/translate/vmx-ops.c
> > > b/target-ppc/translate/vmx-ops.c
> > > index 7449396..ca69e56 100644
> > > --- a/target-ppc/translate/vmx-ops.c
> > > +++ b/target-ppc/translate/vmx-ops.c
> > > @@ -41,6 +41,9 @@ GEN_HANDLER_E(name, 0x04, opc2, opc3, 0x00000000,
> > > PPC_NONE, PPC2_ALTIVEC_207)
> > > #define GEN_VXFORM_300(name, opc2, opc3)
> > > \
> > > GEN_HANDLER_E(name, 0x04, opc2, opc3, 0x00000000, PPC_NONE, PPC2_ISA300)
> > > +#define GEN_VXFORM_300_EXT(name, opc2, opc3, inval) \
> > > +GEN_HANDLER_E(name, 0x04, opc2, opc3, inval, PPC_NONE, PPC2_ISA300)
> > > +
> > > #define GEN_VXFORM_DUAL(name0, name1, opc2, opc3, type0, type1) \
> > > GEN_HANDLER_E(name0##_##name1, 0x4, opc2, opc3, 0x00000000, type0,
> > > type1)
> > > @@ -191,11 +194,16 @@ GEN_VXRFORM(vcmpgefp, 3, 7)
> > > GEN_VXRFORM_DUAL(vcmpgtfp, vcmpgtud, 3, 11, PPC_ALTIVEC, PPC_NONE)
> > > GEN_VXRFORM_DUAL(vcmpbfp, vcmpgtsd, 3, 15, PPC_ALTIVEC, PPC_NONE)
> > > -#define GEN_VXFORM_SIMM(name, opc2, opc3) \
> > > - GEN_HANDLER(name, 0x04, opc2, opc3, 0x00000000, PPC_ALTIVEC)
> > > -GEN_VXFORM_SIMM(vspltisb, 6, 12),
> > > -GEN_VXFORM_SIMM(vspltish, 6, 13),
> > > -GEN_VXFORM_SIMM(vspltisw, 6, 14),
> > > +#define GEN_VXFORM_DUAL_INV(name0, name1, opc2, opc3, inval0, inval1,
> > > type) \
> > > +GEN_OPCODE_DUAL(name0##_##name1, 0x04, opc2, opc3, inval0, inval1, type,
> > > \
> > > + PPC_NONE)
> > > +GEN_VXFORM_DUAL_INV(vspltisb, vinsertb, 6, 12, 0x00000000, 0x100000,
> > > + PPC2_ALTIVEC_207),
> > > +GEN_VXFORM_DUAL_INV(vspltish, vinserth, 6, 13, 0x00000000, 0x100000,
> > > + PPC2_ALTIVEC_207),
> > > +GEN_VXFORM_DUAL_INV(vspltisw, vinsertw, 6, 14, 0x00000000, 0x100000,
> > > + PPC2_ALTIVEC_207),
> > > +GEN_VXFORM_300_EXT(vinsertd, 6, 15, 0x100000),
> > > #define GEN_VXFORM_NOA(name, opc2, opc3)
> > > \
> > > GEN_HANDLER(name, 0x04, opc2, opc3, 0x001f0000, PPC_ALTIVEC)
>
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
- [Qemu-devel] [PATCH v3 0/5] POWER9 TCG enablement - part3, Rajalakshmi Srinivasaraghavan, 2016/08/11
- [Qemu-devel] [PATCH v3 2/5] target-ppc: add vector extract instructions, Rajalakshmi Srinivasaraghavan, 2016/08/11
- [Qemu-devel] [PATCH v3 3/5] target-ppc: add vector count trailing zeros instructions, Rajalakshmi Srinivasaraghavan, 2016/08/11
- [Qemu-devel] [PATCH v3 4/5] target-ppc: add vector bit permute doubleword instruction, Rajalakshmi Srinivasaraghavan, 2016/08/11
- [Qemu-devel] [PATCH v3 5/5] target-ppc: add vector permute right indexed instruction, Rajalakshmi Srinivasaraghavan, 2016/08/11