diff options
Diffstat (limited to 'target/arm/translate-neon.c.inc')
| -rw-r--r-- | target/arm/translate-neon.c.inc | 117 |
1 files changed, 94 insertions, 23 deletions
diff --git a/target/arm/translate-neon.c.inc b/target/arm/translate-neon.c.inc index f6c68e30ab..a02b8369a1 100644 --- a/target/arm/translate-neon.c.inc +++ b/target/arm/translate-neon.c.inc @@ -429,7 +429,7 @@ static bool trans_VLDST_multiple(DisasContext *s, arg_VLDST_multiple *a) { /* Neon load/store multiple structures */ int nregs, interleave, spacing, reg, n; - MemOp endian = s->be_data; + MemOp mop, align, endian; int mmu_idx = get_mem_index(s); int size = a->size; TCGv_i64 tmp64; @@ -473,20 +473,36 @@ static bool trans_VLDST_multiple(DisasContext *s, arg_VLDST_multiple *a) } /* For our purposes, bytes are always little-endian. */ + endian = s->be_data; if (size == 0) { endian = MO_LE; } + + /* Enforce alignment requested by the instruction */ + if (a->align) { + align = pow2_align(a->align + 2); /* 4 ** a->align */ + } else { + align = s->align_mem ? MO_ALIGN : 0; + } + /* * Consecutive little-endian elements from a single register * can be promoted to a larger little-endian operation. */ if (interleave == 1 && endian == MO_LE) { + /* Retain any natural alignment. */ + if (align == MO_ALIGN) { + align = pow2_align(size); + } size = 3; } + tmp64 = tcg_temp_new_i64(); addr = tcg_temp_new_i32(); tmp = tcg_const_i32(1 << size); load_reg_var(s, addr, a->rn); + + mop = endian | size | align; for (reg = 0; reg < nregs; reg++) { for (n = 0; n < 8 >> size; n++) { int xs; @@ -494,13 +510,16 @@ static bool trans_VLDST_multiple(DisasContext *s, arg_VLDST_multiple *a) int tt = a->vd + reg + spacing * xs; if (a->l) { - gen_aa32_ld_i64(s, tmp64, addr, mmu_idx, endian | size); + gen_aa32_ld_internal_i64(s, tmp64, addr, mmu_idx, mop); neon_store_element64(tt, n, size, tmp64); } else { neon_load_element64(tmp64, tt, n, size); - gen_aa32_st_i64(s, tmp64, addr, mmu_idx, endian | size); + gen_aa32_st_internal_i64(s, tmp64, addr, mmu_idx, mop); } tcg_gen_add_i32(addr, addr, tmp); + + /* Subsequent memory operations inherit alignment */ + mop &= ~MO_AMASK; } } } @@ -520,6 +539,7 @@ static bool trans_VLD_all_lanes(DisasContext *s, arg_VLD_all_lanes *a) int size = a->size; int nregs = a->n + 1; TCGv_i32 addr, tmp; + MemOp mop, align; if (!arm_dc_feature(s, ARM_FEATURE_NEON)) { return false; @@ -530,18 +550,33 @@ static bool trans_VLD_all_lanes(DisasContext *s, arg_VLD_all_lanes *a) return false; } + align = 0; if (size == 3) { if (nregs != 4 || a->a == 0) { return false; } /* For VLD4 size == 3 a == 1 means 32 bits at 16 byte alignment */ - size = 2; - } - if (nregs == 1 && a->a == 1 && size == 0) { - return false; - } - if (nregs == 3 && a->a == 1) { - return false; + size = MO_32; + align = MO_ALIGN_16; + } else if (a->a) { + switch (nregs) { + case 1: + if (size == 0) { + return false; + } + align = MO_ALIGN; + break; + case 2: + align = pow2_align(size + 1); + break; + case 3: + return false; + case 4: + align = pow2_align(size + 2); + break; + default: + g_assert_not_reached(); + } } if (!vfp_access_check(s)) { @@ -554,13 +589,12 @@ static bool trans_VLD_all_lanes(DisasContext *s, arg_VLD_all_lanes *a) */ stride = a->t ? 2 : 1; vec_size = nregs == 1 ? stride * 8 : 8; - + mop = size | align; tmp = tcg_temp_new_i32(); addr = tcg_temp_new_i32(); load_reg_var(s, addr, a->rn); for (reg = 0; reg < nregs; reg++) { - gen_aa32_ld_i32(s, tmp, addr, get_mem_index(s), - s->be_data | size); + gen_aa32_ld_i32(s, tmp, addr, get_mem_index(s), mop); if ((vd & 1) && vec_size == 16) { /* * We cannot write 16 bytes at once because the @@ -576,6 +610,9 @@ static bool trans_VLD_all_lanes(DisasContext *s, arg_VLD_all_lanes *a) } tcg_gen_addi_i32(addr, addr, 1 << size); vd += stride; + + /* Subsequent memory operations inherit alignment */ + mop &= ~MO_AMASK; } tcg_temp_free_i32(tmp); tcg_temp_free_i32(addr); @@ -592,6 +629,7 @@ static bool trans_VLDST_single(DisasContext *s, arg_VLDST_single *a) int nregs = a->n + 1; int vd = a->vd; TCGv_i32 addr, tmp; + MemOp mop; if (!arm_dc_feature(s, ARM_FEATURE_NEON)) { return false; @@ -606,7 +644,7 @@ static bool trans_VLDST_single(DisasContext *s, arg_VLDST_single *a) switch (nregs) { case 1: if (((a->align & (1 << a->size)) != 0) || - (a->size == 2 && ((a->align & 3) == 1 || (a->align & 3) == 2))) { + (a->size == 2 && (a->align == 1 || a->align == 2))) { return false; } break; @@ -621,7 +659,7 @@ static bool trans_VLDST_single(DisasContext *s, arg_VLDST_single *a) } break; case 4: - if ((a->size == 2) && ((a->align & 3) == 3)) { + if (a->size == 2 && a->align == 3) { return false; } break; @@ -641,25 +679,58 @@ static bool trans_VLDST_single(DisasContext *s, arg_VLDST_single *a) return true; } + /* Pick up SCTLR settings */ + mop = finalize_memop(s, a->size); + + if (a->align) { + MemOp align_op; + + switch (nregs) { + case 1: + /* For VLD1, use natural alignment. */ + align_op = MO_ALIGN; + break; + case 2: + /* For VLD2, use double alignment. */ + align_op = pow2_align(a->size + 1); + break; + case 4: + if (a->size == MO_32) { + /* + * For VLD4.32, align = 1 is double alignment, align = 2 is + * quad alignment; align = 3 is rejected above. + */ + align_op = pow2_align(a->size + a->align); + } else { + /* For VLD4.8 and VLD.16, we want quad alignment. */ + align_op = pow2_align(a->size + 2); + } + break; + default: + /* For VLD3, the alignment field is zero and rejected above. */ + g_assert_not_reached(); + } + + mop = (mop & ~MO_AMASK) | align_op; + } + tmp = tcg_temp_new_i32(); addr = tcg_temp_new_i32(); load_reg_var(s, addr, a->rn); - /* - * TODO: if we implemented alignment exceptions, we should check - * addr against the alignment encoded in a->align here. - */ + for (reg = 0; reg < nregs; reg++) { if (a->l) { - gen_aa32_ld_i32(s, tmp, addr, get_mem_index(s), - s->be_data | a->size); + gen_aa32_ld_internal_i32(s, tmp, addr, get_mem_index(s), mop); neon_store_element(vd, a->reg_idx, a->size, tmp); } else { /* Store */ neon_load_element(tmp, vd, a->reg_idx, a->size); - gen_aa32_st_i32(s, tmp, addr, get_mem_index(s), - s->be_data | a->size); + gen_aa32_st_internal_i32(s, tmp, addr, get_mem_index(s), mop); } vd += a->stride; tcg_gen_addi_i32(addr, addr, 1 << a->size); + + /* Subsequent memory operations inherit alignment */ + mop &= ~MO_AMASK; } tcg_temp_free_i32(addr); tcg_temp_free_i32(tmp); |