2008-10-15 Rodrigo Kumpera <rkumpera@novell.com>

	* mini-ops.h: Add ops for packed shuffle/max/avg and
	extract mask.

	* cpu-x86.md: Same.

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

	* simd-intrinsics.c (vector8us_intrinsics): Add avg, shuffle and
	extract mask.

	* simd-intrinsics.c (simd_intrinsic_emit_extract_mask): New function
	to emit extract mask op.

	* simd-intrinsics.c (simd_intrinsic_emit_shuffle): Retrofic this function
	to emit word shuffles.

svn path=/trunk/mono/; revision=115920

diff --git a/mono/mini/ChangeLog b/mono/mini/ChangeLog
index 3196eae35ecd61132de6273ed85af150165fcfb4..2c11df2c45d0d7320796884f8574df7ce42f6028 100644 (file)
--- a/mono/mini/ChangeLog
+++ b/mono/mini/ChangeLog
@@ -1,3 +1,21 @@
+2008-10-15  Rodrigo Kumpera  <rkumpera@novell.com>
+
+       * mini-ops.h: Add ops for packed shuffle/max/avg and
+       extract mask.
+       
+       * cpu-x86.md: Same.
+
+       * mini-x86.c (mono_arch_output_basic_block): Same.
+
+       * simd-intrinsics.c (vector8us_intrinsics): Add avg, shuffle and
+       extract mask.
+
+       * simd-intrinsics.c (simd_intrinsic_emit_extract_mask): New function
+       to emit extract mask op.
+
+       * simd-intrinsics.c (simd_intrinsic_emit_shuffle): Retrofic this function
+       to emit word shuffles.
+
 2008-10-15  Mark Probst  <mark.probst@gmail.com>
 
        * mini.c (mono_allocate_stack_slots_full): Align stack frame to

diff --git a/mono/mini/cpu-x86.md b/mono/mini/cpu-x86.md
index 3a737499845d78e6090b2441e6f41a0d6d0901aa..b0f0f22ffd66d0991c141c26c81eac9542c9afea 100644 (file)
--- a/mono/mini/cpu-x86.md
+++ b/mono/mini/cpu-x86.md
@@ -435,6 +435,10 @@ rsqrtps: dest:x src1:x len:4
 rcpps: dest:x src1:x len:4
 shuffleps: dest:x src1:x len:5
 
+pshufflew_high: dest:x src1:x len:5
+pshufflew_low: dest:x src1:x len:5
+extract_mask: dest:i src1:x len:4
+
 paddb: dest:x src1:x src2:x len:4 clob:1
 paddw: dest:x src1:x src2:x len:4 clob:1
 paddd: dest:x src1:x src2:x len:4 clob:1
@@ -443,6 +447,13 @@ psubb: dest:x src1:x src2:x len:4 clob:1
 psubw: dest:x src1:x src2:x len:4 clob:1
 psubd: dest:x src1:x src2:x len:4 clob:1
 
+pmaxb_un: dest:x src1:x src2:x len:4 clob:1
+pmaxw_un: dest:x src1:x src2:x len:5 clob:1
+pmaxd_un: dest:x src1:x src2:x len:5 clob:1
+
+pavgb_un: dest:x src1:x src2:x len:4 clob:1
+pavgw_un: dest:x src1:x src2:x len:4 clob:1
+
 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

diff --git a/mono/mini/mini-ops.h b/mono/mini/mini-ops.h
index f79d4294ea8dff17361ffb5717316883531d547d..92d97610ecec699cb027aac0f5d8714872a503bf 100644 (file)
--- a/mono/mini/mini-ops.h
+++ b/mono/mini/mini-ops.h
@@ -654,6 +654,11 @@ 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_PSHUFLEW_HIGH, "pshufflew_high", XREG, XREG, NONE)
+MINI_OP(OP_PSHUFLEW_LOW, "pshufflew_low", XREG, XREG, NONE)
+
+MINI_OP(OP_EXTRACT_MASK, "extract_mask", IREG, XREG, NONE)
+
 MINI_OP(OP_PAND, "pand", XREG, XREG, XREG)
 MINI_OP(OP_POR, "por", XREG, XREG, XREG)
 MINI_OP(OP_PXOR, "pxor", XREG, XREG, XREG)
@@ -666,6 +671,13 @@ MINI_OP(OP_PSUBB, "psubb", XREG, XREG, XREG)
 MINI_OP(OP_PSUBW, "psubw", XREG, XREG, XREG)
 MINI_OP(OP_PSUBD, "psubd", XREG, XREG, XREG)
 
+MINI_OP(OP_PMAXB_UN, "pmaxb_un", XREG, XREG, XREG)
+MINI_OP(OP_PMAXW_UN, "pmaxw_un", XREG, XREG, XREG)
+MINI_OP(OP_PMAXD_UN, "pmaxd_un", XREG, XREG, XREG)
+
+MINI_OP(OP_PAVGB_UN, "pavgb_un", XREG, XREG, XREG)
+MINI_OP(OP_PAVGW_UN, "pavgw_un", XREG, XREG, XREG)
+
 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)

diff --git a/mono/mini/mini-x86.c b/mono/mini/mini-x86.c
index 62d776a5163ac5ecf550195358f81271528950ed..6955b31b1903bbc2f8e2606d7400a6dfad410075 100644 (file)
--- a/mono/mini/mini-x86.c
+++ b/mono/mini/mini-x86.c
@@ -3873,6 +3873,20 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
                        x86_pshufd_reg_reg (code, ins->dreg, ins->sreg1, ins->inst_c0);
                        break;
+
+               case OP_PSHUFLEW_HIGH:
+                       g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
+                       x86_pshufw_reg_reg (code, ins->dreg, ins->sreg1, ins->inst_c0, 1);
+                       break;
+               case OP_PSHUFLEW_LOW:
+                       g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
+                       x86_pshufw_reg_reg (code, ins->dreg, ins->sreg1, ins->inst_c0, 0);
+                       break;
+
+               case OP_EXTRACT_MASK:
+                       x86_sse_alu_pd_reg_reg (code, X86_SSE_PMOVMSKB, ins->dreg, ins->sreg1);
+                       break;
+       
                case OP_PAND:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PAND, ins->sreg1, ins->sreg2);
                        break;
@@ -3903,6 +3917,23 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PSUBD, ins->sreg1, ins->sreg2);
                        break;
 
+               case OP_PMAXB_UN:
+                       x86_sse_alu_pd_reg_reg (code, X86_SSE_PMAXUB, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_PMAXW_UN:
+                       x86_sse_alu_sse41_reg_reg (code, X86_SSE_PMAXUW, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_PMAXD_UN:
+                       x86_sse_alu_sse41_reg_reg (code, X86_SSE_PMAXUD, ins->sreg1, ins->sreg2);
+                       break;
+
+               case OP_PAVGB_UN:
+                       x86_sse_alu_pd_reg_reg (code, X86_SSE_PAVGB, ins->sreg1, ins->sreg2);
+                       break;
+               case OP_PAVGW_UN:
+                       x86_sse_alu_pd_reg_reg (code, X86_SSE_PAVGW, ins->sreg1, ins->sreg2);
+                       break;
+
                case OP_UNPACK_LOWB:
                        x86_sse_alu_pd_reg_reg (code, X86_SSE_PUNPCKLBW, ins->sreg1, ins->sreg2);
                        break;

diff --git a/mono/mini/simd-intrinsics.c b/mono/mini/simd-intrinsics.c
index 4c129640d80d4aab543fb3f21486b28aaed94ea7..5f41c07c0b5b54e2527abb8c42ffd46868902b7f 100644 (file)
--- a/mono/mini/simd-intrinsics.c
+++ b/mono/mini/simd-intrinsics.c
@@ -71,7 +71,8 @@ enum {
        SIMD_EMIT_SHUFFLE,
        SIMD_EMIT_SHIFT,
        SIMD_EMIT_LOAD_ALIGNED,
-       SIMD_EMIT_STORE_ALIGNED
+       SIMD_EMIT_STORE_ALIGNED,
+       SIMD_EMIT_EXTRACT_MASK
 };
 
 /*This is the size of the largest method name + 1 (to fit the ending \0). Align to 4 as well.*/
@@ -112,7 +113,7 @@ static const SimdIntrinsc vector4f_intrinsics[] = {
        { "Max", OP_MAXPS, SIMD_EMIT_BINARY },
        { "Min", OP_MINPS, SIMD_EMIT_BINARY },
        { "Reciprocal", OP_RCPPS, SIMD_EMIT_UNARY },
-       { "Shuffle", 0, SIMD_EMIT_SHUFFLE },
+       { "Shuffle", OP_SHUFLEPS, SIMD_EMIT_SHUFFLE },
        { "Sqrt", OP_SQRTPS, SIMD_EMIT_UNARY },
        { "StoreAligned", 0, SIMD_EMIT_STORE_ALIGNED },
        { "get_W", 3, SIMD_EMIT_GETTER },
@@ -162,8 +163,12 @@ setters
  */
 static const SimdIntrinsc vector8us_intrinsics[] = {
        { "AddWithSaturation", OP_PADDW_SAT_UN, SIMD_EMIT_BINARY },
+       { "Average", OP_PAVGW_UN, SIMD_EMIT_BINARY },
+       { "ExtractByteMask", 0, SIMD_EMIT_EXTRACT_MASK },
        { "LoadAligned", 0, SIMD_EMIT_LOAD_ALIGNED },
        { "ShiftRightArithmetic", OP_PSARW, SIMD_EMIT_SHIFT },
+       { "ShuffleHigh", OP_PSHUFLEW_HIGH, SIMD_EMIT_SHUFFLE },
+       { "ShuffleLow", OP_PSHUFLEW_LOW, SIMD_EMIT_SHUFFLE },
        { "StoreAligned", 0, SIMD_EMIT_STORE_ALIGNED },
        { "SubWithSaturation", OP_PSUBW_SAT_UN, SIMD_EMIT_BINARY },
        { "UnpackHigh", OP_UNPACK_HIGHW, SIMD_EMIT_BINARY },
@@ -623,13 +628,13 @@ simd_intrinsic_emit_shuffle (const SimdIntrinsc *intrinsic, MonoCompile *cfg, Mo
        /*TODO Exposing shuffle is not a good thing as it's non obvious. We should come up with better abstractions*/
 
        if (args [1]->opcode != OP_ICONST) {
-               g_warning ("Vector4f:Shuffle with non literals is not yet supported");
+               g_warning ("Shuffle with non literals is not yet supported");
                g_assert_not_reached ();
        }
        vreg = get_simd_vreg (cfg, cmethod, args [0]);
        NULLIFY_INS (args [1]);
 
-       MONO_INST_NEW (cfg, ins, OP_SHUFLEPS);
+       MONO_INST_NEW (cfg, ins, intrinsic->opcode);
        ins->klass = cmethod->klass;
        ins->sreg1 = vreg;
        ins->inst_c0 = args [1]->inst_c0;
@@ -672,6 +677,24 @@ simd_intrinsic_emit_store_aligned (const SimdIntrinsc *intrinsic, MonoCompile *c
        return ins;
 }
 
+static MonoInst*
+simd_intrinsic_emit_extract_mask (const SimdIntrinsc *intrinsic, MonoCompile *cfg, MonoMethod *cmethod, MonoInst **args)
+{
+       MonoInst *ins;
+       int vreg;
+       
+       vreg = get_simd_vreg (cfg, cmethod, args [0]);
+
+       MONO_INST_NEW (cfg, ins, OP_EXTRACT_MASK);
+       ins->klass = cmethod->klass;
+       ins->sreg1 = vreg;
+       ins->type = STACK_I4;
+       ins->dreg = alloc_ireg (cfg);
+       MONO_ADD_INS (cfg->cbb, ins);
+
+       return ins;
+}
+
 static const char *
 simd_version_name (guint32 version)
 {
@@ -736,6 +759,8 @@ emit_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
                return simd_intrinsic_emit_load_aligned (result, cfg, cmethod, args);
        case SIMD_EMIT_STORE_ALIGNED:
                return simd_intrinsic_emit_store_aligned (result, cfg, cmethod, args);
+       case SIMD_EMIT_EXTRACT_MASK:
+               return simd_intrinsic_emit_extract_mask (result, cfg, cmethod, args);
        }
        g_assert_not_reached ();
 }