x86emul: handle AVX512-FP16 insns encoded in 0f3a opcode map

In order to re-use (also in subsequent patches) existing code and tables
as much as possible, simply introduce a new boolean field in emulator
state indicating whether an insn is one with a half-precision source.
Everything else then follows "naturally".

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>

diff --git a/tools/tests/x86_emulator/evex-disp8.c b/tools/tests/x86_emulator/evex-disp8.c
index 78153847a009b087e759bca8cb8dbe25443937aa..63ed0fca401bce2b9fe31b2b680ca27621f324eb 100644 (file)
--- a/tools/tests/x86_emulator/evex-disp8.c
+++ b/tools/tests/x86_emulator/evex-disp8.c
@@ -76,6 +76,7 @@ enum esz {
     ESZ_b,
     ESZ_w,
     ESZ_bw,
+    ESZ_fp16,
 };
 
 #ifndef __i386__
@@ -601,6 +602,19 @@ static const struct test avx512_vpopcntdq_all[] = {
     INSN(popcnt, 66, 0f38, 55, vl, dq, vl)
 };
 
+static const struct test avx512_fp16_all[] = {
+    INSN(cmpph,           , 0f3a, c2,    vl, fp16, vl),
+    INSN(cmpsh,         f3, 0f3a, c2,    el, fp16, el),
+    INSN(fpclassph,       , 0f3a, 66,    vl, fp16, vl),
+    INSN(fpclasssh,       , 0f3a, 67,    el, fp16, el),
+    INSN(getmantph,       , 0f3a, 26,    vl, fp16, vl),
+    INSN(getmantsh,       , 0f3a, 27,    el, fp16, el),
+    INSN(reduceph,        , 0f3a, 56,    vl, fp16, vl),
+    INSN(reducesh,        , 0f3a, 57,    el, fp16, el),
+    INSN(rndscaleph,      , 0f3a, 08,    vl, fp16, vl),
+    INSN(rndscalesh,      , 0f3a, 0a,    el, fp16, el),
+};
+
 static const struct test gfni_all[] = {
     INSN(gf2p8affineinvqb, 66, 0f3a, cf, vl, q, vl),
     INSN(gf2p8affineqb,    66, 0f3a, ce, vl, q, vl),
@@ -728,8 +742,10 @@ static void test_one(const struct test *test, enum vl vl,
         break;
 
     case ESZ_w:
-        esz = 2;
         evex.w = 1;
+        /* fall through */
+    case ESZ_fp16:
+        esz = 2;
         break;
 
 #ifdef __i386__
@@ -845,7 +861,7 @@ static void test_one(const struct test *test, enum vl vl,
     case ESZ_b: case ESZ_w: case ESZ_bw:
         return;
 
-    case ESZ_d: case ESZ_q:
+    case ESZ_d: case ESZ_q: case ESZ_fp16:
         break;
 
     default:
@@ -1002,6 +1018,7 @@ void evex_disp8_test(void *instr, struct x86_emulate_ctxt *ctxt,
     RUN(avx512_vnni, all);
     RUN(avx512_vp2intersect, all);
     RUN(avx512_vpopcntdq, all);
+    RUN(avx512_fp16, all);
 
     if ( cpu_has_avx512f )
     {

diff --git a/tools/tests/x86_emulator/predicates.c b/tools/tests/x86_emulator/predicates.c
index 80db0e4cf3bdc6f6173d0b1deb33bcd03ae30a03..50b5c75cbf7d7c87d31977c7538535b5867f72c8 100644 (file)
--- a/tools/tests/x86_emulator/predicates.c
+++ b/tools/tests/x86_emulator/predicates.c
@@ -1988,8 +1988,10 @@ static const struct evex {
     { { 0x03 }, 3, T, R, pfx_66, Wn, Ln }, /* valign{d,q} */
     { { 0x04 }, 3, T, R, pfx_66, W0, Ln }, /* vpermilps */
     { { 0x05 }, 3, T, R, pfx_66, W1, Ln }, /* vpermilpd */
+    { { 0x08 }, 3, T, R, pfx_no, W0, Ln }, /* vrndscaleph */
     { { 0x08 }, 3, T, R, pfx_66, W0, Ln }, /* vrndscaleps */
     { { 0x09 }, 3, T, R, pfx_66, W1, Ln }, /* vrndscalepd */
+    { { 0x0a }, 3, T, R, pfx_no, W0, LIG }, /* vrndscalesh */
     { { 0x0a }, 3, T, R, pfx_66, W0, LIG }, /* vrndscaless */
     { { 0x0b }, 3, T, R, pfx_66, W1, LIG }, /* vrndscalesd */
     { { 0x0f }, 3, T, R, pfx_66, WIG, Ln }, /* vpalignr */
@@ -2009,7 +2011,9 @@ static const struct evex {
     { { 0x22 }, 3, T, R, pfx_66, Wn, L0 }, /* vpinsr{d,q} */
     { { 0x23 }, 3, T, R, pfx_66, Wn, L1|L2 }, /* vshuff{32x4,64x2} */
     { { 0x25 }, 3, T, R, pfx_66, Wn, Ln }, /* vpternlog{d,q} */
+    { { 0x26 }, 3, T, R, pfx_no, W0, Ln }, /* vgetmantph */
     { { 0x26 }, 3, T, R, pfx_66, Wn, Ln }, /* vgetmantp{s,d} */
+    { { 0x27 }, 3, T, R, pfx_no, W0, LIG }, /* vgetmantsh */
     { { 0x27 }, 3, T, R, pfx_66, Wn, LIG }, /* vgetmants{s,d} */
     { { 0x38 }, 3, T, R, pfx_66, Wn, L1|L2 }, /* vinserti{32x4,64x2} */
     { { 0x39 }, 3, T, W, pfx_66, Wn, L1|L2 }, /* vextracti{32x4,64x2} */
@@ -2024,14 +2028,20 @@ static const struct evex {
     { { 0x51 }, 3, T, R, pfx_66, Wn, LIG }, /* vranges{s,d} */
     { { 0x54 }, 3, T, R, pfx_66, Wn, Ln }, /* vfixupimmp{s,d} */
     { { 0x55 }, 3, T, R, pfx_66, Wn, LIG }, /* vfixumpimms{s,d} */
+    { { 0x56 }, 3, T, R, pfx_no, W0, Ln }, /* vreduceph */
     { { 0x56 }, 3, T, R, pfx_66, Wn, Ln }, /* vreducep{s,d} */
+    { { 0x57 }, 3, T, R, pfx_no, W0, LIG }, /* vreducesh */
     { { 0x57 }, 3, T, R, pfx_66, Wn, LIG }, /* vreduces{s,d} */
+    { { 0x66 }, 3, T, R, pfx_no, W0, Ln }, /* vfpclassph */
     { { 0x66 }, 3, T, R, pfx_66, Wn, Ln }, /* vfpclassp{s,d} */
+    { { 0x67 }, 3, T, R, pfx_no, W0, LIG }, /* vfpclasssh */
     { { 0x67 }, 3, T, R, pfx_66, Wn, LIG }, /* vfpclasss{s,d} */
     { { 0x70 }, 3, T, R, pfx_66, W1, Ln }, /* vshldw */
     { { 0x71 }, 3, T, R, pfx_66, Wn, Ln }, /* vshld{d,q} */
     { { 0x72 }, 3, T, R, pfx_66, W1, Ln }, /* vshrdw */
     { { 0x73 }, 3, T, R, pfx_66, Wn, Ln }, /* vshrd{d,q} */
+    { { 0xc2 }, 3, T, R, pfx_no, W0, Ln }, /* vcmpph */
+    { { 0xc2 }, 3, T, R, pfx_f3, W0, LIG }, /* vcmpsh */
     { { 0xce }, 3, T, R, pfx_66, W1, Ln }, /* vgf2p8affineqb */
     { { 0xcf }, 3, T, R, pfx_66, W1, Ln }, /* vgf2p8affineinvqb */
 };

diff --git a/tools/tests/x86_emulator/test_x86_emulator.c b/tools/tests/x86_emulator/test_x86_emulator.c
index 5592b91914705e677ded3130817773771b0ce3df..c0d908cc4fb305243655764c03dc2e88ca0b8bc2 100644 (file)
--- a/tools/tests/x86_emulator/test_x86_emulator.c
+++ b/tools/tests/x86_emulator/test_x86_emulator.c
@@ -4710,6 +4710,44 @@ int main(int argc, char **argv)
     else
         printf("skipped\n");
 
+    printf("%-40s", "Testing vfpclassphz $0x46,128(%ecx),%k3...");
+    if ( stack_exec && cpu_has_avx512_fp16 )
+    {
+        decl_insn(vfpclassph);
+
+        asm volatile ( put_insn(vfpclassph,
+                                /* 0x46: check for +/- 0 and neg. */
+                                /* vfpclassphz $0x46, 128(%0), %%k3 */
+                                ".byte 0x62, 0xf3, 0x7c, 0x48\n\t"
+                                ".byte 0x66, 0x59, 0x02, 0x46")
+                       :: "c" (NULL) );
+
+        set_insn(vfpclassph);
+        for ( i = 0; i < 3; ++i )
+        {
+            res[16 + i * 5 + 0] = 0x7fff0000; /* +0 / +NaN */
+            res[16 + i * 5 + 1] = 0xffff8000; /* -0 / -NaN */
+            res[16 + i * 5 + 2] = 0x80010001; /* +DEN / -DEN */
+            res[16 + i * 5 + 3] = 0xfc00f800; /* -FIN / -INF */
+            res[16 + i * 5 + 4] = 0x7c007800; /* +FIN / +INF */
+        }
+        res[31] = 0;
+        regs.ecx = (unsigned long)res - 64;
+        rc = x86_emulate(&ctxt, &emulops);
+        if ( rc != X86EMUL_OKAY || !check_eip(vfpclassph) )
+            goto fail;
+        asm volatile ( "kmovd %%k3, %0" : "=g" (rc) );
+        /*
+         * 0b11(0001100101)*3
+         * 0b1100_0110_0101_0001_1001_0100_0110_0101
+         */
+        if ( rc != 0xc6519465 )
+            goto fail;
+        printf("okay\n");
+    }
+    else
+        printf("skipped\n");
+
     /*
      * The following compress/expand tests are not only making sure the
      * accessed data is correct, but they also verify (by placing operands

diff --git a/tools/tests/x86_emulator/x86-emulate.h b/tools/tests/x86_emulator/x86-emulate.h
index 256ccb9fefc18f752457c8ce1cf82b07e11884b9..de1f826680106d769d5f3212898261a3dc92ff07 100644 (file)
--- a/tools/tests/x86_emulator/x86-emulate.h
+++ b/tools/tests/x86_emulator/x86-emulate.h
@@ -183,6 +183,7 @@ void wrpkru(unsigned int val);
 #define cpu_has_avx512_4fmaps (cp.feat.avx512_4fmaps && xcr0_mask(0xe6))
 #define cpu_has_avx512_vp2intersect (cp.feat.avx512_vp2intersect && xcr0_mask(0xe6))
 #define cpu_has_serialize  cp.feat.serialize
+#define cpu_has_avx512_fp16 (cp.feat.avx512_fp16 && xcr0_mask(0xe6))
 #define cpu_has_avx_vnni   (cp.feat.avx_vnni && xcr0_mask(6))
 #define cpu_has_avx512_bf16 (cp.feat.avx512_bf16 && xcr0_mask(0xe6))
 #define cpu_has_avx_ifma   (cp.feat.avx_ifma && xcr0_mask(6))

diff --git a/xen/arch/x86/x86_emulate/decode.c b/xen/arch/x86/x86_emulate/decode.c
index 362de670bcf673ecae8bc588d14c48e138a2473f..13d71bd7e179b700129240c44aff466630becbbf 100644 (file)
--- a/xen/arch/x86/x86_emulate/decode.c
+++ b/xen/arch/x86/x86_emulate/decode.c
@@ -508,6 +508,7 @@ static const struct ext0f3a_table {
     [0x7a ... 0x7b] = { .simd_size = simd_scalar_opc, .four_op = 1 },
     [0x7c ... 0x7d] = { .simd_size = simd_packed_fp, .four_op = 1 },
     [0x7e ... 0x7f] = { .simd_size = simd_scalar_opc, .four_op = 1 },
+    [0xc2] = { .simd_size = simd_any_fp, .d8s = d8s_vl },
     [0xcc] = { .simd_size = simd_other },
     [0xce ... 0xcf] = { .simd_size = simd_packed_int, .d8s = d8s_vl },
     [0xdf] = { .simd_size = simd_packed_int, .two_op = 1 },
@@ -569,7 +570,7 @@ static unsigned int decode_disp8scale(enum disp8scale scale,
         if ( s->evex.brs )
         {
     case d8s_dq:
-            return 2 + s->evex.w;
+            return 1 + !s->fp16 + s->evex.w;
         }
         break;
 
@@ -586,7 +587,7 @@ static unsigned int decode_disp8scale(enum disp8scale scale,
         /* fall through */
     case simd_scalar_opc:
     case simd_scalar_vexw:
-        return 2 + s->evex.w;
+        return 1 + !s->fp16 + s->evex.w;
 
     case simd_128:
         /* These should have an explicit size specified. */
@@ -1406,7 +1407,29 @@ int x86emul_decode(struct x86_emulate_state *s,
              */
             s->simd_size = ext0f3a_table[b].simd_size;
             if ( evex_encoded() )
+            {
+                switch ( b )
+                {
+                case 0x08: /* vrndscaleph */
+                case 0x0a: /* vrndscalesh */
+                case 0x26: /* vfpclassph */
+                case 0x27: /* vfpclasssh */
+                case 0x56: /* vgetmantph */
+                case 0x57: /* vgetmantsh */
+                case 0x66: /* vreduceph */
+                case 0x67: /* vreducesh */
+                    if ( !s->evex.pfx )
+                        s->fp16 = true;
+                    break;
+
+                case 0xc2: /* vpcmp{p,s}h */
+                    if ( !(s->evex.pfx & VEX_PREFIX_DOUBLE_MASK) )
+                        s->fp16 = true;
+                    break;
+                }
+
                 disp8scale = decode_disp8scale(ext0f3a_table[b].d8s, s);
+            }
             break;
 
         case ext_8f09:
@@ -1701,7 +1724,7 @@ int x86emul_decode(struct x86_emulate_state *s,
             break;
         case vex_f3:
             generate_exception_if(evex_encoded() && s->evex.w, X86_EXC_UD);
-            s->op_bytes = 4;
+            s->op_bytes = 4 >> s->fp16;
             break;
         case vex_f2:
             generate_exception_if(evex_encoded() && !s->evex.w, X86_EXC_UD);
@@ -1711,11 +1734,11 @@ int x86emul_decode(struct x86_emulate_state *s,
         break;
 
     case simd_scalar_opc:
-        s->op_bytes = 4 << (ctxt->opcode & 1);
+        s->op_bytes = 2 << (!s->fp16 + (ctxt->opcode & 1));
         break;
 
     case simd_scalar_vexw:
-        s->op_bytes = 4 << s->vex.w;
+        s->op_bytes = 2 << (!s->fp16 + s->vex.w);
         break;
 
     case simd_128:

diff --git a/xen/arch/x86/x86_emulate/private.h b/xen/arch/x86/x86_emulate/private.h
index bd05a37183d89e155b195fab7bea2e9006ab8c24..3644b5fbee2d7b506a8ba9cfa2a8ff0ea82fb849 100644 (file)
--- a/xen/arch/x86/x86_emulate/private.h
+++ b/xen/arch/x86/x86_emulate/private.h
@@ -293,6 +293,7 @@ struct x86_emulate_state {
     bool lock_prefix;
     bool not_64bit; /* Instruction not available in 64bit. */
     bool fpu_ctrl;  /* Instruction is an FPU control one. */
+    bool fp16;      /* Instruction has half-precision FP source operand. */
     opcode_desc_t desc;
     union vex vex;
     union evex evex;
@@ -581,6 +582,7 @@ amd_like(const struct x86_emulate_ctxt *ctxt)
 #define vcpu_has_avx512_vp2intersect() (ctxt->cpuid->feat.avx512_vp2intersect)
 #define vcpu_has_serialize()   (ctxt->cpuid->feat.serialize)
 #define vcpu_has_tsxldtrk()    (ctxt->cpuid->feat.tsxldtrk)
+#define vcpu_has_avx512_fp16() (ctxt->cpuid->feat.avx512_fp16)
 #define vcpu_has_avx_vnni()    (ctxt->cpuid->feat.avx_vnni)
 #define vcpu_has_avx512_bf16() (ctxt->cpuid->feat.avx512_bf16)
 #define vcpu_has_wrmsrns()     (ctxt->cpuid->feat.wrmsrns)

diff --git a/xen/arch/x86/x86_emulate/x86_emulate.c b/xen/arch/x86/x86_emulate/x86_emulate.c
index 6ff12384ec03319cd10aa319d7e94edb7e5e3114..5f1c43934680bee4b8d8c0ef4ecfe3b50d53a8f5 100644 (file)
--- a/xen/arch/x86/x86_emulate/x86_emulate.c
+++ b/xen/arch/x86/x86_emulate/x86_emulate.c
@@ -1266,7 +1266,7 @@ x86_emulate(
     b = ctxt->opcode;
     d = state.desc;
 #define state (&state)
-    elem_bytes = 4 << evex.w;
+    elem_bytes = 2 << (!state->fp16 + evex.w);
 
     generate_exception_if(state->not_64bit && mode_64bit(), X86_EXC_UD);
 
@@ -7148,6 +7148,15 @@ x86_emulate(
         avx512_vlen_check(b & 2);
         goto simd_imm8_zmm;
 
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x0a): /* vrndscalesh $imm8,xmm/mem,xmm,xmm{k} */
+        generate_exception_if(ea.type != OP_REG && evex.brs, X86_EXC_UD);
+        /* fall through */
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x08): /* vrndscaleph $imm8,[xyz]mm/mem,[xyz]mm{k} */
+        host_and_vcpu_must_have(avx512_fp16);
+        generate_exception_if(evex.w, X86_EXC_UD);
+        avx512_vlen_check(b & 2);
+        goto simd_imm8_zmm;
+
 #endif /* X86EMUL_NO_SIMD */
 
     CASE_SIMD_PACKED_INT(0x0f3a, 0x0f): /* palignr $imm8,{,x}mm/mem,{,x}mm */
@@ -7458,6 +7467,14 @@ x86_emulate(
             avx512_vlen_check(false);
         goto simd_imm8_zmm;
 
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x26): /* vgetmantph $imm8,[xyz]mm/mem,[xyz]mm{k} */
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x56): /* vreduceph $imm8,[xyz]mm/mem,[xyz]mm{k} */
+        host_and_vcpu_must_have(avx512_fp16);
+        generate_exception_if(evex.w, X86_EXC_UD);
+        if ( ea.type != OP_REG || !evex.brs )
+            avx512_vlen_check(false);
+        goto simd_imm8_zmm;
+
     case X86EMUL_OPC_EVEX_66(0x0f3a, 0x51): /* vranges{s,d} $imm8,xmm/mem,xmm,xmm{k} */
     case X86EMUL_OPC_EVEX_66(0x0f3a, 0x57): /* vreduces{s,d} $imm8,xmm/mem,xmm,xmm{k} */
         host_and_vcpu_must_have(avx512dq);
@@ -7470,6 +7487,16 @@ x86_emulate(
             avx512_vlen_check(true);
         goto simd_imm8_zmm;
 
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x27): /* vgetmantsh $imm8,xmm/mem,xmm,xmm{k} */
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x57): /* vreducesh $imm8,xmm/mem,xmm,xmm{k} */
+        host_and_vcpu_must_have(avx512_fp16);
+        generate_exception_if(evex.w, X86_EXC_UD);
+        if ( !evex.brs )
+            avx512_vlen_check(true);
+        else
+            generate_exception_if(ea.type != OP_REG, X86_EXC_UD);
+        goto simd_imm8_zmm;
+
     case X86EMUL_OPC_VEX_66(0x0f3a, 0x30): /* kshiftr{b,w} $imm8,k,k */
     case X86EMUL_OPC_VEX_66(0x0f3a, 0x32): /* kshiftl{b,w} $imm8,k,k */
         if ( !vex.w )
@@ -7633,6 +7660,16 @@ x86_emulate(
         avx512_vlen_check(true);
         goto simd_imm8_zmm;
 
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x66): /* vfpclassph $imm8,[xyz]mm/mem,k{k} */
+    case X86EMUL_OPC_EVEX(0x0f3a, 0x67): /* vfpclasssh $imm8,xmm/mem,k{k} */
+        host_and_vcpu_must_have(avx512_fp16);
+        generate_exception_if(evex.w || !evex.r || !evex.R || evex.z, X86_EXC_UD);
+        if ( !(b & 1) )
+            goto avx512f_imm8_no_sae;
+        generate_exception_if(evex.brs, X86_EXC_UD);
+        avx512_vlen_check(true);
+        goto simd_imm8_zmm;
+
     case X86EMUL_OPC_EVEX_66(0x0f3a, 0x70): /* vpshldw $imm8,[xyz]mm/mem,[xyz]mm,[xyz]mm{k} */
     case X86EMUL_OPC_EVEX_66(0x0f3a, 0x72): /* vpshrdw $imm8,[xyz]mm/mem,[xyz]mm,[xyz]mm{k} */
         generate_exception_if(!evex.w, X86_EXC_UD);
@@ -7643,6 +7680,16 @@ x86_emulate(
         host_and_vcpu_must_have(avx512_vbmi2);
         goto avx512f_imm8_no_sae;
 
+    case X86EMUL_OPC_EVEX_F3(0x0f3a, 0xc2): /* vcmpsh $imm8,xmm/mem,xmm,k{k} */
+        generate_exception_if(ea.type != OP_REG && evex.brs, X86_EXC_UD);
+        /* fall through */
+    case X86EMUL_OPC_EVEX(0x0f3a, 0xc2): /* vcmpph $imm8,[xyz]mm/mem,[xyz]mm,k{k} */
+        host_and_vcpu_must_have(avx512_fp16);
+        generate_exception_if(evex.w || !evex.r || !evex.R || evex.z, X86_EXC_UD);
+        if ( ea.type != OP_REG || !evex.brs )
+            avx512_vlen_check(evex.pfx & VEX_PREFIX_SCALAR_MASK);
+        goto simd_imm8_zmm;
+
     case X86EMUL_OPC(0x0f3a, 0xcc):     /* sha1rnds4 $imm8,xmm/m128,xmm */
         host_and_vcpu_must_have(sha);
         op_bytes = 16;