target/i386: implement F16C instructions

F16C only consists of two instructions, which are a bit peculiar
nevertheless.

First, they access only the low half of an YMM or XMM register for the
packed-half operand; the exact size still depends on the VEX.L flag.
This is similar to the existing avx_movx flag, but not exactly because
avx_movx is hardcoded to affect operand 2.  To this end I added a "ph"
format name; it's possible to reuse this approach for the VPMOVSX and
VPMOVZX instructions, though that would also require adding two more
formats for the low-quarter and low-eighth of an operand.

Second, VCVTPS2PH is somewhat weird because it *stores* the result of
the instruction into memory rather than loading it.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

7 files changed, 66 insertions, 4 deletions

diff --git a/target/i386/cpu.c b/target/i386/cpu.c
index 0ebd610faa..6292b7e12f 100644
--- a/target/i386/cpu.c
+++ b/target/i386/cpu.c
@@ -625,13 +625,12 @@ void x86_cpu_vendor_words2str(char *dst, uint32_t vendor1,
           CPUID_EXT_SSE41 | CPUID_EXT_SSE42 | CPUID_EXT_POPCNT | \
           CPUID_EXT_XSAVE | /* CPUID_EXT_OSXSAVE is dynamic */   \
           CPUID_EXT_MOVBE | CPUID_EXT_AES | CPUID_EXT_HYPERVISOR | \
-          CPUID_EXT_RDRAND | CPUID_EXT_AVX)
+          CPUID_EXT_RDRAND | CPUID_EXT_AVX | CPUID_EXT_F16C)
           /* missing:
           CPUID_EXT_DTES64, CPUID_EXT_DSCPL, CPUID_EXT_VMX, CPUID_EXT_SMX,
           CPUID_EXT_EST, CPUID_EXT_TM2, CPUID_EXT_CID, CPUID_EXT_FMA,
           CPUID_EXT_XTPR, CPUID_EXT_PDCM, CPUID_EXT_PCID, CPUID_EXT_DCA,
-          CPUID_EXT_X2APIC, CPUID_EXT_TSC_DEADLINE_TIMER,
-          CPUID_EXT_F16C */
+          CPUID_EXT_X2APIC, CPUID_EXT_TSC_DEADLINE_TIMER */
 
 #ifdef TARGET_X86_64
 #define TCG_EXT2_X86_64_FEATURES (CPUID_EXT2_SYSCALL | CPUID_EXT2_LM)
diff --git a/target/i386/cpu.h b/target/i386/cpu.h
index dad2b2db8d..d4bc19577a 100644
--- a/target/i386/cpu.h
+++ b/target/i386/cpu.h
@@ -1258,6 +1258,7 @@ typedef union ZMMReg {
     uint16_t _w_ZMMReg[512 / 16];
     uint32_t _l_ZMMReg[512 / 32];
     uint64_t _q_ZMMReg[512 / 64];
+    float16  _h_ZMMReg[512 / 16];
     float32  _s_ZMMReg[512 / 32];
     float64  _d_ZMMReg[512 / 64];
     XMMReg   _x_ZMMReg[512 / 128];
@@ -1282,6 +1283,7 @@ typedef struct BNDCSReg {
 #define ZMM_B(n) _b_ZMMReg[63 - (n)]
 #define ZMM_W(n) _w_ZMMReg[31 - (n)]
 #define ZMM_L(n) _l_ZMMReg[15 - (n)]
+#define ZMM_H(n) _h_ZMMReg[31 - (n)]
 #define ZMM_S(n) _s_ZMMReg[15 - (n)]
 #define ZMM_Q(n) _q_ZMMReg[7 - (n)]
 #define ZMM_D(n) _d_ZMMReg[7 - (n)]
@@ -1301,6 +1303,7 @@ typedef struct BNDCSReg {
 #define ZMM_B(n) _b_ZMMReg[n]
 #define ZMM_W(n) _w_ZMMReg[n]
 #define ZMM_L(n) _l_ZMMReg[n]
+#define ZMM_H(n) _h_ZMMReg[n]
 #define ZMM_S(n) _s_ZMMReg[n]
 #define ZMM_Q(n) _q_ZMMReg[n]
 #define ZMM_D(n) _d_ZMMReg[n]
diff --git a/target/i386/ops_sse.h b/target/i386/ops_sse.h
index 0799712f6e..33c61896ee 100644
--- a/target/i386/ops_sse.h
+++ b/target/i386/ops_sse.h
@@ -586,6 +586,35 @@ void glue(helper_cvtpd2ps, SUFFIX)(CPUX86State *env, Reg *d, Reg *s)
     }
 }
 
+#if SHIFT >= 1
+void glue(helper_cvtph2ps, SUFFIX)(CPUX86State *env, Reg *d, Reg *s)
+{
+    int i;
+
+    for (i = 2 << SHIFT; --i >= 0; ) {
+         d->ZMM_S(i) = float16_to_float32(s->ZMM_H(i), true, &env->sse_status);
+    }
+}
+
+void glue(helper_cvtps2ph, SUFFIX)(CPUX86State *env, Reg *d, Reg *s, int mode)
+{
+    int i;
+    FloatRoundMode prev_rounding_mode = env->sse_status.float_rounding_mode;
+    if (!(mode & (1 << 2))) {
+        set_x86_rounding_mode(mode & 3, &env->sse_status);
+    }
+
+    for (i = 0; i < 2 << SHIFT; i++) {
+        d->ZMM_H(i) = float32_to_float16(s->ZMM_S(i), true, &env->sse_status);
+    }
+    for (i >>= 2; i < 1 << SHIFT; i++) {
+        d->Q(i) = 0;
+    }
+
+    env->sse_status.float_rounding_mode = prev_rounding_mode;
+}
+#endif
+
 #if SHIFT == 1
 void helper_cvtss2sd(CPUX86State *env, Reg *d, Reg *v, Reg *s)
 {
diff --git a/target/i386/ops_sse_header.h b/target/i386/ops_sse_header.h
index 2f1f811f9f..c4c41976c0 100644
--- a/target/i386/ops_sse_header.h
+++ b/target/i386/ops_sse_header.h
@@ -353,6 +353,12 @@ DEF_HELPER_4(glue(aeskeygenassist, SUFFIX), void, env, Reg, Reg, i32)
 DEF_HELPER_5(glue(pclmulqdq, SUFFIX), void, env, Reg, Reg, Reg, i32)
 #endif
 
+/* F16C helpers */
+#if SHIFT >= 1
+DEF_HELPER_3(glue(cvtph2ps, SUFFIX), void, env, Reg, Reg)
+DEF_HELPER_4(glue(cvtps2ph, SUFFIX), void, env, Reg, Reg, int)
+#endif
+
 /* AVX helpers */
 #if SHIFT >= 1
 DEF_HELPER_4(glue(vpermilpd, SUFFIX), void, env, Reg, Reg, Reg)
diff --git a/target/i386/tcg/decode-new.c.inc b/target/i386/tcg/decode-new.c.inc
index 8e1eb9db42..8baee9018a 100644
--- a/target/i386/tcg/decode-new.c.inc
+++ b/target/i386/tcg/decode-new.c.inc
@@ -336,6 +336,7 @@ static const X86OpEntry opcodes_0F38_00toEF[240] = {
     [0x07] = X86_OP_ENTRY3(PHSUBSW,   V,x,  H,x,   W,x,  vex4 cpuid(SSSE3) mmx avx2_256 p_00_66),
 
     [0x10] = X86_OP_ENTRY2(PBLENDVB,  V,x,         W,x,  vex4 cpuid(SSE41) avx2_256 p_66),
+    [0x13] = X86_OP_ENTRY2(VCVTPH2PS, V,x,         W,ph, vex11 cpuid(F16C) p_66),
     [0x14] = X86_OP_ENTRY2(BLENDVPS,  V,x,         W,x,  vex4 cpuid(SSE41) p_66),
     [0x15] = X86_OP_ENTRY2(BLENDVPD,  V,x,         W,x,  vex4 cpuid(SSE41) p_66),
     /* Listed incorrectly as type 4 */
@@ -525,6 +526,7 @@ static const X86OpEntry opcodes_0F3A[256] = {
     [0x15] = X86_OP_ENTRY3(PEXTRW,     E,w,  V,dq, I,b,  vex5 cpuid(SSE41) zext0 p_66),
     [0x16] = X86_OP_ENTRY3(PEXTR,      E,y,  V,dq, I,b,  vex5 cpuid(SSE41) p_66),
     [0x17] = X86_OP_ENTRY3(VEXTRACTPS, E,d,  V,dq, I,b,  vex5 cpuid(SSE41) p_66),
+    [0x1d] = X86_OP_ENTRY3(VCVTPS2PH,  W,ph, V,x,  I,b,  vex11 cpuid(F16C) p_66),
 
     [0x20] = X86_OP_ENTRY4(PINSRB,     V,dq, H,dq, E,b,  vex5 cpuid(SSE41) zext2 p_66),
     [0x21] = X86_OP_GROUP0(VINSERTPS),
@@ -1051,6 +1053,10 @@ static bool decode_op_size(DisasContext *s, X86OpEntry *e, X86OpSize size, MemOp
         *ot = s->vex_l ? MO_256 : MO_128;
         return true;
 
+    case X86_SIZE_ph: /* SSE/AVX packed half precision */
+        *ot = s->vex_l ? MO_128 : MO_64;
+        return true;
+
     case X86_SIZE_d64:  /* Default to 64-bit in 64-bit mode */
         *ot = CODE64(s) && s->dflag == MO_32 ? MO_64 : s->dflag;
         return true;
@@ -1342,6 +1348,8 @@ static bool has_cpuid_feature(DisasContext *s, X86CPUIDFeature cpuid)
     switch (cpuid) {
     case X86_FEAT_None:
         return true;
+    case X86_FEAT_F16C:
+        return (s->cpuid_ext_features & CPUID_EXT_F16C);
     case X86_FEAT_MOVBE:
         return (s->cpuid_ext_features & CPUID_EXT_MOVBE);
     case X86_FEAT_PCLMULQDQ:
diff --git a/target/i386/tcg/decode-new.h b/target/i386/tcg/decode-new.h
index f159c26850..0ef54628ee 100644
--- a/target/i386/tcg/decode-new.h
+++ b/target/i386/tcg/decode-new.h
@@ -92,6 +92,7 @@ typedef enum X86OpSize {
     /* Custom */
     X86_SIZE_d64,
     X86_SIZE_f64,
+    X86_SIZE_ph, /* SSE/AVX packed half precision */
 } X86OpSize;
 
 typedef enum X86CPUIDFeature {
@@ -103,6 +104,7 @@ typedef enum X86CPUIDFeature {
     X86_FEAT_AVX2,
     X86_FEAT_BMI1,
     X86_FEAT_BMI2,
+    X86_FEAT_F16C,
     X86_FEAT_MOVBE,
     X86_FEAT_PCLMULQDQ,
     X86_FEAT_SSE,
diff --git a/target/i386/tcg/emit.c.inc b/target/i386/tcg/emit.c.inc
index ebf299451d..9334f0939d 100644
--- a/target/i386/tcg/emit.c.inc
+++ b/target/i386/tcg/emit.c.inc
@@ -296,7 +296,7 @@ static void gen_writeback(DisasContext *s, X86DecodedInsn *decode, int opn, TCGv
     case X86_OP_MMX:
         break;
     case X86_OP_SSE:
-        if (!op->has_ea && (s->prefix & PREFIX_VEX) && op->ot == MO_128) {
+        if (!op->has_ea && (s->prefix & PREFIX_VEX) && op->ot <= MO_128) {
             tcg_gen_gvec_dup_imm(MO_64,
                                  offsetof(CPUX86State, xmm_regs[op->n].ZMM_X(1)),
                                  16, 16, 0);
@@ -852,6 +852,7 @@ UNARY_INT_SSE(VCVTTPD2DQ, cvttpd2dq)
 UNARY_INT_SSE(VCVTDQ2PS, cvtdq2ps)
 UNARY_INT_SSE(VCVTPS2DQ, cvtps2dq)
 UNARY_INT_SSE(VCVTTPS2DQ, cvttps2dq)
+UNARY_INT_SSE(VCVTPH2PS, cvtph2ps)
 
 
 static inline void gen_unary_imm_sse(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode,
@@ -1868,6 +1869,20 @@ static void gen_VCVTfp2fp(DisasContext *s, CPUX86State *env, X86DecodedInsn *dec
                      gen_helper_cvtsd2ss, gen_helper_cvtss2sd);
 }
 
+static void gen_VCVTPS2PH(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode)
+{
+    gen_unary_imm_fp_sse(s, env, decode,
+                      gen_helper_cvtps2ph_xmm,
+                      gen_helper_cvtps2ph_ymm);
+    /*
+     * VCVTPS2PH is the only instruction that performs an operation on a
+     * register source and then *stores* into memory.
+     */
+    if (decode->op[0].has_ea) {
+        gen_store_sse(s, decode, decode->op[0].offset);
+    }
+}
+
 static void gen_VCVTSI2Sx(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode)
 {
     int vec_len = vector_len(s, decode);