target/ppc: Update vector insns to use 128 bit
v2 cover target/ppc: Update vector insns to use 128 bit
Updating a bunch of VMX and VSX storage access instructions to use tcg_gen_qemu_ld/st_i128 instead of using tcg_gen_qemu_ld/st_i64 in succession; as suggested by Richard, in my decodetree patches. Plus some minor clean-ups to facilitate the above in case of VMX insns.
Change log:
v2 : Applied IFALIGN_PAIR memop changes in patches 2/3 and 3/3, based on review comments by Richard in v1.
v1 : https://lore.kernel.org/qemu-devel/[email protected]/
Chinmay Rath (3):
target/ppc: Move get/set_avr64 functions to vmx-impl.c.inc. target/ppc: Update VMX storage access insns to use
tcg_gen_qemu_ld/st_i128.
target/ppc : Update VSX storage access insns to use tcg_gen_qemu
_ld/st_i128.target/ppc/translate.c | 10 ----- target/ppc/translate/vmx-impl.c.inc | 52 ++++++++++++--------- target/ppc/translate/vsx-impl.c.inc | 70 +++++++++++++---------------- 3 files changed, 61 insertions(+), 71 deletions(-)
Signature · 1 line
2.39.3
PATCH v2 2/3 target/ppc: Update VMX storage access insns to use tcg_gen_qemu_ld/st_i128.
Updated instructions {l, st}vx to use tcg_gen_qemu_ld/st_i128, instead of using 64 bits loads/stores in succession. Introduced functions {get, set}_avr_full in vmx-impl.c.inc to facilitate the above, and potential future usage.
--- target/ppc/translate/vmx-impl.c.inc | 42 ++++++++++++++--------------- 1 file changed, 20 insertions(+), 22 deletions(-)
target/ppc/translate/vmx-impl.c.inc +20 -22
@@ -24,25 +24,29 @@ static inline void set_avr64(int regno, TCGv_i64 src, bool high) tcg_gen_st_i64(src, tcg_env, avr64_offset(regno, high)); } +static inline void get_avr_full(TCGv_i128 dst, int regno) +{ + tcg_gen_ld_i128(dst, tcg_env, avr_full_offset(regno)); +} + +static inline void set_avr_full(int regno, TCGv_i128 src) +{ + tcg_gen_st_i128(src, tcg_env, avr_full_offset(regno)); +} + static bool trans_LVX(DisasContext *ctx, arg_X *a) { TCGv EA; - TCGv_i64 avr; + TCGv_i128 avr; REQUIRE_INSNS_FLAGS(ctx, ALTIVEC); REQUIRE_VECTOR(ctx); gen_set_access_type(ctx, ACCESS_INT); - avr = tcg_temp_new_i64(); + avr = tcg_temp_new_i128(); EA = do_ea_calc(ctx, a->ra, cpu_gpr[a->rb]); tcg_gen_andi_tl(EA, EA, ~0xf); - /* - * We only need to swap high and low halves. gen_qemu_ld64_i64 - * does necessary 64-bit byteswap already. - */ - gen_qemu_ld64_i64(ctx, avr, EA); - set_avr64(a->rt, avr, !ctx->le_mode); - tcg_gen_addi_tl(EA, EA, 8); - gen_qemu_ld64_i64(ctx, avr, EA); - set_avr64(a->rt, avr, ctx->le_mode); + tcg_gen_qemu_ld_i128(avr, EA, ctx->mem_idx, + DEF_MEMOP(MO_128 | MO_ATOM_IFALIGN_PAIR)); + set_avr_full(a->rt, avr); return true; }
@@ -56,22 +60,16 @@ static bool trans_LVXL(DisasContext *ctx, arg_LVXL *a) static bool trans_STVX(DisasContext *ctx, arg_STVX *a) { TCGv EA; - TCGv_i64 avr; + TCGv_i128 avr; REQUIRE_INSNS_FLAGS(ctx, ALTIVEC); REQUIRE_VECTOR(ctx); gen_set_access_type(ctx, ACCESS_INT); - avr = tcg_temp_new_i64(); + avr = tcg_temp_new_i128(); EA = do_ea_calc(ctx, a->ra, cpu_gpr[a->rb]); tcg_gen_andi_tl(EA, EA, ~0xf); - /* - * We only need to swap high and low halves. gen_qemu_st64_i64 - * does necessary 64-bit byteswap already. - */ - get_avr64(avr, a->rt, !ctx->le_mode); - gen_qemu_st64_i64(ctx, avr, EA); - tcg_gen_addi_tl(EA, EA, 8); - get_avr64(avr, a->rt, ctx->le_mode); - gen_qemu_st64_i64(ctx, avr, EA); + get_avr_full(avr, a->rt); + tcg_gen_qemu_st_i128(avr, EA, ctx->mem_idx, + DEF_MEMOP(MO_128 | MO_ATOM_IFALIGN_PAIR)); return true; }
Signature · 1 line
2.39.3
PATCH v2 3/3 target/ppc : Update VSX storage access insns to use tcg_gen_qemu _ld/st_i128.
In reply to Chinmay RathUpdated many VSX instructions to use tcg_gen_qemu_ld/st_i128, instead of using tcg_gen_qemu_ld/st_i64 consecutively.
Introduced functions {get,set}_vsr_full to facilitate the above & for future use.
--- target/ppc/translate/vsx-impl.c.inc | 70 +++++++++++++---------------- 1 file changed, 31 insertions(+), 39 deletions(-)
target/ppc/translate/vsx-impl.c.inc +31 -39
@@ -10,6 +10,16 @@ static inline void set_cpu_vsr(int n, TCGv_i64 src, bool high) tcg_gen_st_i64(src, tcg_env, vsr64_offset(n, high)); } +static inline void get_vsr_full(TCGv_i128 dst, int reg) +{ + tcg_gen_ld_i128(dst, tcg_env, vsr_full_offset(reg)); +} + +static inline void set_vsr_full(int reg, TCGv_i128 src) +{ + tcg_gen_st_i128(src, tcg_env, vsr_full_offset(reg)); +} + static inline TCGv_ptr gen_vsr_ptr(int reg) { TCGv_ptr r = tcg_temp_new_ptr();
@@ -196,20 +206,17 @@ static bool trans_LXVH8X(DisasContext *ctx, arg_LXVH8X *a) static bool trans_LXVB16X(DisasContext *ctx, arg_LXVB16X *a) { TCGv EA; - TCGv_i64 xth, xtl; + TCGv_i128 data; REQUIRE_VSX(ctx); REQUIRE_INSNS_FLAGS2(ctx, ISA300); - xth = tcg_temp_new_i64(); - xtl = tcg_temp_new_i64(); + data = tcg_temp_new_i128(); gen_set_access_type(ctx, ACCESS_INT); EA = do_ea_calc(ctx, a->ra, cpu_gpr[a->rb]); - tcg_gen_qemu_ld_i64(xth, EA, ctx->mem_idx, MO_BEUQ); - tcg_gen_addi_tl(EA, EA, 8); - tcg_gen_qemu_ld_i64(xtl, EA, ctx->mem_idx, MO_BEUQ); - set_cpu_vsr(a->rt, xth, true); - set_cpu_vsr(a->rt, xtl, false); + tcg_gen_qemu_ld_i128(data, EA, ctx->mem_idx, + MO_BE | MO_128 | MO_ATOM_IFALIGN_PAIR); + set_vsr_full(a->rt, data); return true; }
@@ -385,20 +392,17 @@ static bool trans_STXVH8X(DisasContext *ctx, arg_STXVH8X *a) static bool trans_STXVB16X(DisasContext *ctx, arg_STXVB16X *a) { TCGv EA; - TCGv_i64 xsh, xsl; + TCGv_i128 data; REQUIRE_VSX(ctx); REQUIRE_INSNS_FLAGS2(ctx, ISA300); - xsh = tcg_temp_new_i64(); - xsl = tcg_temp_new_i64(); - get_cpu_vsr(xsh, a->rt, true); - get_cpu_vsr(xsl, a->rt, false); + data = tcg_temp_new_i128(); gen_set_access_type(ctx, ACCESS_INT); EA = do_ea_calc(ctx, a->ra, cpu_gpr[a->rb]); - tcg_gen_qemu_st_i64(xsh, EA, ctx->mem_idx, MO_BEUQ); - tcg_gen_addi_tl(EA, EA, 8); - tcg_gen_qemu_st_i64(xsl, EA, ctx->mem_idx, MO_BEUQ); + get_vsr_full(data, a->rt); + tcg_gen_qemu_st_i128(data, EA, ctx->mem_idx, + MO_BE | MO_128 | MO_ATOM_IFALIGN_PAIR); return true; }
@@ -2175,13 +2179,13 @@ static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, int rt, bool store, bool paired) { TCGv ea; - TCGv_i64 xt; + TCGv_i128 data; MemOp mop; int rt1, rt2; - xt = tcg_temp_new_i64(); + data = tcg_temp_new_i128(); - mop = DEF_MEMOP(MO_UQ); + mop = DEF_MEMOP(MO_128 | MO_ATOM_IFALIGN_PAIR); gen_set_access_type(ctx, ACCESS_INT); ea = do_ea_calc(ctx, ra, displ);
@@ -2195,32 +2199,20 @@ static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, } if (store) { - get_cpu_vsr(xt, rt1, !ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); - gen_addr_add(ctx, ea, ea, 8); - get_cpu_vsr(xt, rt1, ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); + get_vsr_full(data, rt1); + tcg_gen_qemu_st_i128(data, ea, ctx->mem_idx, mop); if (paired) { gen_addr_add(ctx, ea, ea, 8); - get_cpu_vsr(xt, rt2, !ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); - gen_addr_add(ctx, ea, ea, 8); - get_cpu_vsr(xt, rt2, ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); + get_vsr_full(data, rt2); + tcg_gen_qemu_st_i128(data, ea, ctx->mem_idx, mop); } } else { - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt1, xt, !ctx->le_mode); - gen_addr_add(ctx, ea, ea, 8); - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt1, xt, ctx->le_mode); + tcg_gen_qemu_ld_i128(data, ea, ctx->mem_idx, mop); + set_vsr_full(rt1, data); if (paired) { gen_addr_add(ctx, ea, ea, 8); - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt2, xt, !ctx->le_mode); - gen_addr_add(ctx, ea, ea, 8); - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt2, xt, ctx->le_mode); + tcg_gen_qemu_ld_i128(data, ea, ctx->mem_idx, mop); + set_vsr_full(rt2, data); } } return true;
Signature · 1 line
2.39.3
PATCH v2 1/3 target/ppc: Move get/set_avr64 functions to vmx-impl.c.inc.
In reply to Chinmay RathThose functions are used to ld/st data to and from Altivec registers, in 64 bits chunks, and are only used in vmx-impl.c.inc file, hence the clean-up movement.
--- target/ppc/translate.c | 10 ---------- target/ppc/translate/vmx-impl.c.inc | 10 ++++++++++ 2 files changed, 10 insertions(+), 10 deletions(-)
target/ppc/translate.c +0 -10
@@ -6200,16 +6200,6 @@ static inline void set_fpr(int regno, TCGv_i64 src)tcg_gen_st_i64 (tcg_constant_i64 (0 ), tcg_env,vsr64_offset (regno, false)); } -static inline void get_avr64 (TCGv_i64 dst,int regno,bool high) -{ -tcg_gen_ld_i64 (dst, tcg_env,avr64_offset (regno, high)); -} - -static inline void set_avr64 (int regno,TCGv_i64 src,bool high) -{ -tcg_gen_st_i64 (src, tcg_env,avr64_offset (regno, high)); -} -/* * Helpers for decodetree used by !function for decoding arguments. */
target/ppc/translate/vmx-impl.c.inc +10 -0
@@ -14,6 +14,16 @@ static inline TCGv_ptr gen_avr_ptr(int reg) return r; } +static inline void get_avr64(TCGv_i64 dst, int regno, bool high) +{ + tcg_gen_ld_i64(dst, tcg_env, avr64_offset(regno, high)); +} + +static inline void set_avr64(int regno, TCGv_i64 src, bool high) +{ + tcg_gen_st_i64(src, tcg_env, avr64_offset(regno, high)); +} + static bool trans_LVX(DisasContext *ctx, arg_X *a) { TCGv EA;
Signature · 1 line
2.39.3
Re: [PATCH v2 1/3] target/ppc: Move get/set_avr64 functions to vmx-impl.c.inc.
Quoted 9 lines · Chinmay Rath
Those functions are used to ld/st data to and from Altivec registers, in 64 bits chunks, and are only used in vmx-impl.c.inc file, hence the clean-up movement.
Signed-off-by Chinmay Rath<[email protected]>--- target/ppc/translate.c | 10 ---------- target/ppc/translate/vmx-impl.c.inc | 10 ++++++++++ 2 files changed, 10 insertions(+), 10 deletions(-)
r~
Re: [PATCH v2 2/3] target/ppc: Update VMX storage access insns to use tcg_gen_qemu_ld/st_i128.
In reply to Chinmay RathQuoted 10 lines · Chinmay Rath
Updated instructions {l, st}vx to use tcg_gen_qemu_ld/st_i128, instead of using 64 bits loads/stores in succession. Introduced functions {get, set}_avr_full in vmx-impl.c.inc to facilitate the above, and potential future usage.
Suggested-by Richard Henderson<[email protected]>Signed-off-by Chinmay Rath<[email protected]>--- target/ppc/translate/vmx-impl.c.inc | 42 ++++++++++++++--------------- 1 file changed, 20 insertions(+), 22 deletions(-)
r~
Re: [PATCH v2 3/3] target/ppc : Update VSX storage access insns to use tcg_gen_qemu _ld/st_i128.
In reply to Chinmay RathQuoted 20 lines · Chinmay Rath
target/ppc/translate/vsx-impl.c.inc +3
-3@@ -2175,13 +2179,13 @@ static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, int rt, bool store, bool paired) { TCGv ea; - TCGv_i64 xt; + TCGv_i128 data; MemOp mop; int rt1, rt2; - xt = tcg_temp_new_i64(); + data = tcg_temp_new_i128(); - mop = DEF_MEMOP(MO_UQ); + mop = DEF_MEMOP(MO_128 | MO_ATOM_IFALIGN_PAIR); gen_set_access_type(ctx, ACCESS_INT); ea = do_ea_calc(ctx, ra, displ);@@ -2195,32 +2199,20 @@ static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, }
target/ppc/translate/vsx-impl.c.inc static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, · line 2205
Quoted 10 lines
target/ppc/translate/vsx-impl.c.inc +2
-5if (store) { - get_cpu_vsr(xt, rt1, !ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); - gen_addr_add(ctx, ea, ea, 8); - get_cpu_vsr(xt, rt1, ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); + get_vsr_full(data, rt1); + tcg_gen_qemu_st_i128(data, ea, ctx->mem_idx, mop); if (paired) { gen_addr_add(ctx, ea, ea, 8);
The increment needs updating to 16.
Quoted 8 lines
target/ppc/translate/vsx-impl.c.inc +2
-5- get_cpu_vsr(xt, rt2, !ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); - gen_addr_add(ctx, ea, ea, 8); - get_cpu_vsr(xt, rt2, ctx->le_mode); - tcg_gen_qemu_st_i64(xt, ea, ctx->mem_idx, mop); + get_vsr_full(data, rt2); + tcg_gen_qemu_st_i128(data, ea, ctx->mem_idx, mop); }
target/ppc/translate/vsx-impl.c.inc static bool do_lstxv(DisasContext *ctx, int ra, TCGv displ, · line 2213
Quoted 10 lines
target/ppc/translate/vsx-impl.c.inc +2
-5} else { - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt1, xt, !ctx->le_mode); - gen_addr_add(ctx, ea, ea, 8); - tcg_gen_qemu_ld_i64(xt, ea, ctx->mem_idx, mop); - set_cpu_vsr(rt1, xt, ctx->le_mode); + tcg_gen_qemu_ld_i128(data, ea, ctx->mem_idx, mop); + set_vsr_full(rt1, data); if (paired) { gen_addr_add(ctx, ea, ea, 8);
Likewise.
With those fixed,
r~