2008-10-13 Rodrigo Kumpera <rkumpera@novell.com>

	* mini-ops.h: Add remaining sse1 fp ops.

	* cpu-x86.md: Add remaining sse1 fp ops.

	* mini-x86.c (mono_arch_output_basic_block): Same.

	* mini.h: Add enum for simd FP compare conditions.

	* simd-intrinsics.c (vector4f_intrinsics): Add all new ops.

	* simd-intrinsics.c (simd_intrinsic_emit_binary): Set inst_c0 to flags
	so the backed can generate the appropriate op.

svn path=/trunk/mono/; revision=115701

diff --git a/mono/mini/ChangeLog b/mono/mini/ChangeLog
index 789ea43f17db66330a8fd28046fb69be0801e26c..09bfe3a6f924b2333d9be9761ce96b0ee03303b3 100644 (file)
--- a/mono/mini/ChangeLog
+++ b/mono/mini/ChangeLog
@@ -1,3 +1,18 @@
+2008-10-13  Rodrigo Kumpera  <rkumpera@novell.com>
+
+       * mini-ops.h: Add remaining sse1 fp ops.
+       
+       * cpu-x86.md: Add remaining sse1 fp ops.
+
+       * mini-x86.c (mono_arch_output_basic_block): Same.
+
+       * mini.h: Add enum for simd FP compare conditions.
+
+       * simd-intrinsics.c (vector4f_intrinsics): Add all new ops.
+
+       * simd-intrinsics.c (simd_intrinsic_emit_binary): Set inst_c0 to flags
+       so the backed can generate the appropriate op.
+
 2008-10-13  Rodrigo Kumpera  <rkumpera@novell.com>
        This patch squeese one more byte from the SimdIntrinsc struct.
 

diff --git a/mono/mini/cpu-x86.md b/mono/mini/cpu-x86.md
index b1a846fe5c8d097828f036e11e4a9b51d6609367..37983f8e45737a1f63de1278106029461d34b703 100644 (file)
--- a/mono/mini/cpu-x86.md
+++ b/mono/mini/cpu-x86.md
@@ -414,6 +414,11 @@ mulps: dest:x src1:x src2:x len:3 clob:1
 subps: dest:x src1:x src2:x len:3 clob:1
 maxps: dest:x src1:x src2:x len:3 clob:1
 minps: dest:x src1:x src2:x len:3 clob:1
+compps: dest:x src1:x src2:x len:4 clob:1
+andps: dest:x src1:x src2:x len:3 clob:1
+andnps: dest:x src1:x src2:x len:3 clob:1
+orps: dest:x src1:x src2:x len:3 clob:1
+xorps: dest:x src1:x src2:x len:3 clob:1
 
 haddps: dest:x src1:x src2:x len:4 clob:1
 hsubps: dest:x src1:x src2:x len:4 clob:1
@@ -425,6 +430,7 @@ pxor: dest:x src1:x src2:x len:4 clob:1
 
 sqrtps: dest:x src1:x len:4
 rsqrtps: dest:x src1:x len:4
+rcpps: dest:x src1:x len:4
 shuffleps: dest:x src1:x len:5
 
 paddb: dest:x src1:x src2:x len:4 clob:1
@@ -439,11 +445,13 @@ unpack_lowb: dest:x src1:x src2:x len:4 clob:1
 unpack_loww: dest:x src1:x src2:x len:4 clob:1
 unpack_lowd: dest:x src1:x src2:x len:4 clob:1
 unpack_lowq: dest:x src1:x src2:x len:4 clob:1
+unpack_lowps: dest:x src1:x src2:x len:3 clob:1
 
 unpack_highb: dest:x src1:x src2:x len:4 clob:1
 unpack_highw: dest:x src1:x src2:x len:4 clob:1
 unpack_highd: dest:x src1:x src2:x len:4 clob:1
 unpack_highq: dest:x src1:x src2:x len:4 clob:1
+unpack_highps: dest:x src1:x src2:x len:3 clob:1
 
 packw: dest:x src1:x src2:x len:4 clob:1 
 packd: dest:x src1:x src2:x len:5 clob:1 

diff --git a/mono/mini/mini-ops.h b/mono/mini/mini-ops.h
index 7310f55ea32ccec7461c25e5e88f6cba571a5099..99ac82fe869a3780501554b1b292becff85324af 100644 (file)
--- a/mono/mini/mini-ops.h
+++ b/mono/mini/mini-ops.h
@@ -638,12 +638,18 @@ MINI_OP(OP_MULPS, "mulps", XREG, XREG, XREG)
 MINI_OP(OP_SUBPS, "subps", XREG, XREG, XREG)
 MINI_OP(OP_MAXPS, "maxps", XREG, XREG, XREG)
 MINI_OP(OP_MINPS, "minps", XREG, XREG, XREG)
+MINI_OP(OP_COMPPS, "compps", XREG, XREG, XREG)
+MINI_OP(OP_ANDPS, "andps", XREG, XREG, XREG)
+MINI_OP(OP_ANDNPS, "andnps", XREG, XREG, XREG)
+MINI_OP(OP_ORPS, "orps", XREG, XREG, XREG)
+MINI_OP(OP_XORPS, "xorps", XREG, XREG, XREG)
 MINI_OP(OP_HADDPS, "haddps", XREG, XREG, XREG)
 MINI_OP(OP_HSUBPS, "hsubps", XREG, XREG, XREG)
 MINI_OP(OP_ADDSUBPS, "addsubps", XREG, XREG, XREG)
 
 MINI_OP(OP_RSQRTPS, "rsqrtps", XREG, XREG, NONE)
 MINI_OP(OP_SQRTPS, "sqrtps", XREG, XREG, NONE)
+MINI_OP(OP_RCPPS, "rcpps", XREG, XREG, NONE)
 MINI_OP(OP_SHUFLEPS, "shuffleps", XREG, XREG, NONE)
 
 MINI_OP(OP_PAND, "pand", XREG, XREG, XREG)
@@ -662,11 +668,13 @@ MINI_OP(OP_UNPACK_LOWB, "unpack_lowb", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_LOWW, "unpack_loww", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_LOWD, "unpack_lowd", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_LOWQ, "unpack_lowq", XREG, XREG, XREG)
+MINI_OP(OP_UNPACK_LOWPS, "unpack_lowps", XREG, XREG, XREG)
 
 MINI_OP(OP_UNPACK_HIGHB, "unpack_highb", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_HIGHW, "unpack_highw", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_HIGHD, "unpack_highd", XREG, XREG, XREG)
 MINI_OP(OP_UNPACK_HIGHQ, "unpack_highq", XREG, XREG, XREG)
+MINI_OP(OP_UNPACK_HIGHPS, "unpack_highps", XREG, XREG, XREG)
 
 MINI_OP(OP_PACKW, "packw", XREG, XREG, XREG)
 MINI_OP(OP_PACKD, "packd", XREG, XREG, XREG)

diff --git a/mono/mini/mini-x86.c b/mono/mini/mini-x86.c
index e4367d1951878417ea6e12e078be8e8c047f5d99..be9746a73560c8be7ea0f711809ca0bdf6fac060 100644 (file)
--- a/mono/mini/mini-x86.c
+++ b/mono/mini/mini-x86.c
@@ -3829,12 +3829,31 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_MINPS:
                        x86_sse_alu_ps_reg_reg (code, X86_SSE_MIN, ins->sreg1, ins->sreg2);
                        break;
+               case OP_COMPPS:
+                       g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
+                       x86_sse_alu_ps_reg_reg_imm (code, X86_SSE_COMP, ins->sreg1, ins->sreg2, ins->inst_c0);
+                       break;
+               case OP_ANDPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_AND, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_ANDNPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_ANDN, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_ORPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_OR, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_XORPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_XOR, ins->sreg1, ins->sreg2);
+                       break;
                case OP_SQRTPS:
                        x86_sse_alu_ps_reg_reg (code, X86_SSE_SQRT, ins->dreg, ins->sreg1);
                        break;
                case OP_RSQRTPS:
                        x86_sse_alu_ps_reg_reg (code, X86_SSE_RSQRT, ins->dreg, ins->sreg1);
                        break;
+               case OP_RCPPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_RCP, ins->dreg, ins->sreg1);
+                       break;
                case OP_ADDSUBPS:
                        x86_sse_alu_sd_reg_reg (code, X86_SSE_ADDSUB, ins->sreg1, ins->sreg2);
                        break;
@@ -3887,6 +3906,9 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_UNPACK_LOWD:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PUNPCKLDQ, ins->sreg1, ins->sreg2);
                        break;
+               case OP_UNPACK_LOWPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_UNPCKL, ins->sreg1, ins->sreg2);
+                       break;
 
                case OP_UNPACK_HIGHB:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PUNPCKHBW, ins->sreg1, ins->sreg2);
@@ -3897,6 +3919,9 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_UNPACK_HIGHD:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PUNPCKHDQ, ins->sreg1, ins->sreg2);
                        break;
+               case OP_UNPACK_HIGHPS:
+                       x86_sse_alu_ps_reg_reg (code, X86_SSE_UNPCKH, ins->sreg1, ins->sreg2);
+                       break;
 
                case OP_PACKW:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PACKUSWB, ins->sreg1, ins->sreg2);

diff --git a/mono/mini/mini.h b/mono/mini/mini.h
index 423f325fe4f703fff7a3291ebeb2c87ebd55ae6c..307f00b7159ca96197ae23123d4a522087d96dee 100644 (file)
--- a/mono/mini/mini.h
+++ b/mono/mini/mini.h
@@ -1706,6 +1706,17 @@ enum {
        SIMD_VERSION_SSE4a      = 6,
 };
 
+enum {
+       SIMD_COMP_EQ,
+       SIMD_COMP_LT,
+       SIMD_COMP_LE,
+       SIMD_COMP_UNORD,
+       SIMD_COMP_NEQ,
+       SIMD_COMP_NLT,
+       SIMD_COMP_NLE,
+       SIMD_COMP_ORD
+};
+
 const char *mono_arch_xregname (int reg) MONO_INTERNAL;
 void mono_simd_simplify_indirection (MonoCompile *cfg) MONO_INTERNAL;
 MonoInst* mono_emit_simd_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args) MONO_INTERNAL;

diff --git a/mono/mini/simd-intrinsics.c b/mono/mini/simd-intrinsics.c
index 0bdd8a7e8e5c5e5e07b380a5aba6f5a74cd780c4..a59819cac69f8d05a147a4b5db73c197b065cebb 100644 (file)
--- a/mono/mini/simd-intrinsics.c
+++ b/mono/mini/simd-intrinsics.c
@@ -91,13 +91,25 @@ setters
  */
 static const SimdIntrinsc vector4f_intrinsics[] = {
        { ".ctor", 0, SIMD_EMIT_CTOR },
+       { "AndNot", OP_ANDNPS, SIMD_EMIT_BINARY },
        { "AddSub", OP_ADDSUBPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE3 },
+       { "CompareEquals", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_EQ },
+       { "CompareLessEqual", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_LE },
+       { "CompareLessThan", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_LT },
+       { "CompareNotEqual", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_NEQ },
+       { "CompareNotLessEqual", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_NLE },
+       { "CompareNotLessThan", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_NLT },
+       { "CompareOrdered", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_ORD },
+       { "CompareUnordered", OP_COMPPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE1, SIMD_COMP_UNORD },
        { "HorizontalAdd", OP_HADDPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE3 },
        { "HorizontalSub", OP_HSUBPS, SIMD_EMIT_BINARY, SIMD_VERSION_SSE3 },    
+       { "InterleaveHigh", OP_UNPACK_HIGHPS, SIMD_EMIT_BINARY },
+       { "InterleaveLow", OP_UNPACK_LOWPS, SIMD_EMIT_BINARY },
        { "InvSqrt", OP_RSQRTPS, SIMD_EMIT_UNARY },
        { "LoadAligned", 0, SIMD_EMIT_LOAD_ALIGNED },
        { "Max", OP_MAXPS, SIMD_EMIT_BINARY },
        { "Min", OP_MINPS, SIMD_EMIT_BINARY },
+       { "Reciprocal", OP_RCPPS, SIMD_EMIT_UNARY },
        { "Shuffle", 0, SIMD_EMIT_SHUFFLE },
        { "Sqrt", OP_SQRTPS, SIMD_EMIT_UNARY },
        { "StoreAligned", 0, SIMD_EMIT_STORE_ALIGNED },
@@ -106,7 +118,10 @@ static const SimdIntrinsc vector4f_intrinsics[] = {
        { "get_Y", 1, SIMD_EMIT_GETTER },
        { "get_Z", 2, SIMD_EMIT_GETTER },
        { "op_Addition", OP_ADDPS, SIMD_EMIT_BINARY },
+       { "op_BitwiseAnd", OP_ANDPS, SIMD_EMIT_BINARY },
+       { "op_BitwiseOr", OP_ORPS, SIMD_EMIT_BINARY },
        { "op_Division", OP_DIVPS, SIMD_EMIT_BINARY },
+       { "op_ExclusiveOr", OP_XORPS, SIMD_EMIT_BINARY },
        { "op_Explicit", 0, SIMD_EMIT_CAST }, 
        { "op_Multiply", OP_MULPS, SIMD_EMIT_BINARY },
        { "op_Subtraction", OP_SUBPS, SIMD_EMIT_BINARY },
@@ -414,6 +429,7 @@ simd_intrinsic_emit_binary (const SimdIntrinsc *intrinsic, MonoCompile *cfg, Mon
        ins->type = STACK_VTYPE;
        ins->klass = cmethod->klass;
        ins->dreg = alloc_ireg (cfg);
+       ins->inst_c0 = intrinsic->flags;
        MONO_ADD_INS (cfg->cbb, ins);
        return ins;
 }