1.1 --- a/src/cpu/x86/vm/assembler_x86.cpp Mon Sep 17 17:02:10 2012 -0700
1.2 +++ b/src/cpu/x86/vm/assembler_x86.cpp Mon Sep 17 19:39:07 2012 -0700
1.3 @@ -3496,6 +3496,33 @@
1.4 emit_byte(0x01);
1.5 }
1.6
1.7 +void Assembler::vinsertf128h(XMMRegister dst, Address src) {
1.8 + assert(VM_Version::supports_avx(), "");
1.9 + InstructionMark im(this);
1.10 + bool vector256 = true;
1.11 + assert(dst != xnoreg, "sanity");
1.12 + int dst_enc = dst->encoding();
1.13 + // swap src<->dst for encoding
1.14 + vex_prefix(src, dst_enc, dst_enc, VEX_SIMD_66, VEX_OPCODE_0F_3A, false, vector256);
1.15 + emit_byte(0x18);
1.16 + emit_operand(dst, src);
1.17 + // 0x01 - insert into upper 128 bits
1.18 + emit_byte(0x01);
1.19 +}
1.20 +
1.21 +void Assembler::vextractf128h(Address dst, XMMRegister src) {
1.22 + assert(VM_Version::supports_avx(), "");
1.23 + InstructionMark im(this);
1.24 + bool vector256 = true;
1.25 + assert(src != xnoreg, "sanity");
1.26 + int src_enc = src->encoding();
1.27 + vex_prefix(dst, 0, src_enc, VEX_SIMD_66, VEX_OPCODE_0F_3A, false, vector256);
1.28 + emit_byte(0x19);
1.29 + emit_operand(src, dst);
1.30 + // 0x01 - extract from upper 128 bits
1.31 + emit_byte(0x01);
1.32 +}
1.33 +
1.34 void Assembler::vinserti128h(XMMRegister dst, XMMRegister nds, XMMRegister src) {
1.35 assert(VM_Version::supports_avx2(), "");
1.36 bool vector256 = true;
1.37 @@ -3507,6 +3534,33 @@
1.38 emit_byte(0x01);
1.39 }
1.40
1.41 +void Assembler::vinserti128h(XMMRegister dst, Address src) {
1.42 + assert(VM_Version::supports_avx2(), "");
1.43 + InstructionMark im(this);
1.44 + bool vector256 = true;
1.45 + assert(dst != xnoreg, "sanity");
1.46 + int dst_enc = dst->encoding();
1.47 + // swap src<->dst for encoding
1.48 + vex_prefix(src, dst_enc, dst_enc, VEX_SIMD_66, VEX_OPCODE_0F_3A, false, vector256);
1.49 + emit_byte(0x38);
1.50 + emit_operand(dst, src);
1.51 + // 0x01 - insert into upper 128 bits
1.52 + emit_byte(0x01);
1.53 +}
1.54 +
1.55 +void Assembler::vextracti128h(Address dst, XMMRegister src) {
1.56 + assert(VM_Version::supports_avx2(), "");
1.57 + InstructionMark im(this);
1.58 + bool vector256 = true;
1.59 + assert(src != xnoreg, "sanity");
1.60 + int src_enc = src->encoding();
1.61 + vex_prefix(dst, 0, src_enc, VEX_SIMD_66, VEX_OPCODE_0F_3A, false, vector256);
1.62 + emit_byte(0x39);
1.63 + emit_operand(src, dst);
1.64 + // 0x01 - extract from upper 128 bits
1.65 + emit_byte(0x01);
1.66 +}
1.67 +
1.68 void Assembler::vzeroupper() {
1.69 assert(VM_Version::supports_avx(), "");
1.70 (void)vex_prefix_and_encode(xmm0, xmm0, xmm0, VEX_SIMD_NONE);
1.71 @@ -8907,11 +8961,9 @@
1.72 pusha();
1.73
1.74 // if we are coming from c1, xmm registers may be live
1.75 - if (UseSSE >= 1) {
1.76 - subptr(rsp, sizeof(jdouble)* LP64_ONLY(16) NOT_LP64(8));
1.77 - }
1.78 int off = 0;
1.79 if (UseSSE == 1) {
1.80 + subptr(rsp, sizeof(jdouble)*8);
1.81 movflt(Address(rsp,off++*sizeof(jdouble)),xmm0);
1.82 movflt(Address(rsp,off++*sizeof(jdouble)),xmm1);
1.83 movflt(Address(rsp,off++*sizeof(jdouble)),xmm2);
1.84 @@ -8921,23 +8973,50 @@
1.85 movflt(Address(rsp,off++*sizeof(jdouble)),xmm6);
1.86 movflt(Address(rsp,off++*sizeof(jdouble)),xmm7);
1.87 } else if (UseSSE >= 2) {
1.88 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm0);
1.89 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm1);
1.90 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm2);
1.91 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm3);
1.92 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm4);
1.93 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm5);
1.94 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm6);
1.95 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm7);
1.96 +#ifdef COMPILER2
1.97 + if (MaxVectorSize > 16) {
1.98 + assert(UseAVX > 0, "256bit vectors are supported only with AVX");
1.99 + // Save upper half of YMM registes
1.100 + subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
1.101 + vextractf128h(Address(rsp, 0),xmm0);
1.102 + vextractf128h(Address(rsp, 16),xmm1);
1.103 + vextractf128h(Address(rsp, 32),xmm2);
1.104 + vextractf128h(Address(rsp, 48),xmm3);
1.105 + vextractf128h(Address(rsp, 64),xmm4);
1.106 + vextractf128h(Address(rsp, 80),xmm5);
1.107 + vextractf128h(Address(rsp, 96),xmm6);
1.108 + vextractf128h(Address(rsp,112),xmm7);
1.109 #ifdef _LP64
1.110 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm8);
1.111 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm9);
1.112 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm10);
1.113 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm11);
1.114 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm12);
1.115 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm13);
1.116 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm14);
1.117 - movdbl(Address(rsp,off++*sizeof(jdouble)),xmm15);
1.118 + vextractf128h(Address(rsp,128),xmm8);
1.119 + vextractf128h(Address(rsp,144),xmm9);
1.120 + vextractf128h(Address(rsp,160),xmm10);
1.121 + vextractf128h(Address(rsp,176),xmm11);
1.122 + vextractf128h(Address(rsp,192),xmm12);
1.123 + vextractf128h(Address(rsp,208),xmm13);
1.124 + vextractf128h(Address(rsp,224),xmm14);
1.125 + vextractf128h(Address(rsp,240),xmm15);
1.126 +#endif
1.127 + }
1.128 +#endif
1.129 + // Save whole 128bit (16 bytes) XMM regiters
1.130 + subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
1.131 + movdqu(Address(rsp,off++*16),xmm0);
1.132 + movdqu(Address(rsp,off++*16),xmm1);
1.133 + movdqu(Address(rsp,off++*16),xmm2);
1.134 + movdqu(Address(rsp,off++*16),xmm3);
1.135 + movdqu(Address(rsp,off++*16),xmm4);
1.136 + movdqu(Address(rsp,off++*16),xmm5);
1.137 + movdqu(Address(rsp,off++*16),xmm6);
1.138 + movdqu(Address(rsp,off++*16),xmm7);
1.139 +#ifdef _LP64
1.140 + movdqu(Address(rsp,off++*16),xmm8);
1.141 + movdqu(Address(rsp,off++*16),xmm9);
1.142 + movdqu(Address(rsp,off++*16),xmm10);
1.143 + movdqu(Address(rsp,off++*16),xmm11);
1.144 + movdqu(Address(rsp,off++*16),xmm12);
1.145 + movdqu(Address(rsp,off++*16),xmm13);
1.146 + movdqu(Address(rsp,off++*16),xmm14);
1.147 + movdqu(Address(rsp,off++*16),xmm15);
1.148 #endif
1.149 }
1.150
1.151 @@ -9015,28 +9094,52 @@
1.152 movflt(xmm5, Address(rsp,off++*sizeof(jdouble)));
1.153 movflt(xmm6, Address(rsp,off++*sizeof(jdouble)));
1.154 movflt(xmm7, Address(rsp,off++*sizeof(jdouble)));
1.155 + addptr(rsp, sizeof(jdouble)*8);
1.156 } else if (UseSSE >= 2) {
1.157 - movdbl(xmm0, Address(rsp,off++*sizeof(jdouble)));
1.158 - movdbl(xmm1, Address(rsp,off++*sizeof(jdouble)));
1.159 - movdbl(xmm2, Address(rsp,off++*sizeof(jdouble)));
1.160 - movdbl(xmm3, Address(rsp,off++*sizeof(jdouble)));
1.161 - movdbl(xmm4, Address(rsp,off++*sizeof(jdouble)));
1.162 - movdbl(xmm5, Address(rsp,off++*sizeof(jdouble)));
1.163 - movdbl(xmm6, Address(rsp,off++*sizeof(jdouble)));
1.164 - movdbl(xmm7, Address(rsp,off++*sizeof(jdouble)));
1.165 + // Restore whole 128bit (16 bytes) XMM regiters
1.166 + movdqu(xmm0, Address(rsp,off++*16));
1.167 + movdqu(xmm1, Address(rsp,off++*16));
1.168 + movdqu(xmm2, Address(rsp,off++*16));
1.169 + movdqu(xmm3, Address(rsp,off++*16));
1.170 + movdqu(xmm4, Address(rsp,off++*16));
1.171 + movdqu(xmm5, Address(rsp,off++*16));
1.172 + movdqu(xmm6, Address(rsp,off++*16));
1.173 + movdqu(xmm7, Address(rsp,off++*16));
1.174 #ifdef _LP64
1.175 - movdbl(xmm8, Address(rsp,off++*sizeof(jdouble)));
1.176 - movdbl(xmm9, Address(rsp,off++*sizeof(jdouble)));
1.177 - movdbl(xmm10, Address(rsp,off++*sizeof(jdouble)));
1.178 - movdbl(xmm11, Address(rsp,off++*sizeof(jdouble)));
1.179 - movdbl(xmm12, Address(rsp,off++*sizeof(jdouble)));
1.180 - movdbl(xmm13, Address(rsp,off++*sizeof(jdouble)));
1.181 - movdbl(xmm14, Address(rsp,off++*sizeof(jdouble)));
1.182 - movdbl(xmm15, Address(rsp,off++*sizeof(jdouble)));
1.183 + movdqu(xmm8, Address(rsp,off++*16));
1.184 + movdqu(xmm9, Address(rsp,off++*16));
1.185 + movdqu(xmm10, Address(rsp,off++*16));
1.186 + movdqu(xmm11, Address(rsp,off++*16));
1.187 + movdqu(xmm12, Address(rsp,off++*16));
1.188 + movdqu(xmm13, Address(rsp,off++*16));
1.189 + movdqu(xmm14, Address(rsp,off++*16));
1.190 + movdqu(xmm15, Address(rsp,off++*16));
1.191 #endif
1.192 - }
1.193 - if (UseSSE >= 1) {
1.194 - addptr(rsp, sizeof(jdouble)* LP64_ONLY(16) NOT_LP64(8));
1.195 + addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
1.196 +#ifdef COMPILER2
1.197 + if (MaxVectorSize > 16) {
1.198 + // Restore upper half of YMM registes.
1.199 + vinsertf128h(xmm0, Address(rsp, 0));
1.200 + vinsertf128h(xmm1, Address(rsp, 16));
1.201 + vinsertf128h(xmm2, Address(rsp, 32));
1.202 + vinsertf128h(xmm3, Address(rsp, 48));
1.203 + vinsertf128h(xmm4, Address(rsp, 64));
1.204 + vinsertf128h(xmm5, Address(rsp, 80));
1.205 + vinsertf128h(xmm6, Address(rsp, 96));
1.206 + vinsertf128h(xmm7, Address(rsp,112));
1.207 +#ifdef _LP64
1.208 + vinsertf128h(xmm8, Address(rsp,128));
1.209 + vinsertf128h(xmm9, Address(rsp,144));
1.210 + vinsertf128h(xmm10, Address(rsp,160));
1.211 + vinsertf128h(xmm11, Address(rsp,176));
1.212 + vinsertf128h(xmm12, Address(rsp,192));
1.213 + vinsertf128h(xmm13, Address(rsp,208));
1.214 + vinsertf128h(xmm14, Address(rsp,224));
1.215 + vinsertf128h(xmm15, Address(rsp,240));
1.216 +#endif
1.217 + addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
1.218 + }
1.219 +#endif
1.220 }
1.221 popa();
1.222 }
