X86: Implement the multiply and add instructions.

3 files changed, 80 insertions, 33 deletions

diff --git a/src/arch/x86/isa/decoder/two_byte_opcodes.isa b/src/arch/x86/isa/decoder/two_byte_opcodes.isa
index cb57bf9f0..d136fcbce 100644
--- a/src/arch/x86/isa/decoder/two_byte_opcodes.isa
+++ b/src/arch/x86/isa/decoder/two_byte_opcodes.isa
@@ -1062,38 +1062,36 @@
                     }
                     default: UD2();
                 }
-            }
-            0x1E: decode LEGACY_DECODEVAL {
-                // no prefix
-                0x0: decode OPCODE_OP_BOTTOM3 {
-                    0x1: Inst::PSLLW(Pq,Qq);
-                    0x2: Inst::PSLLD(Pq,Qq);
-                    0x3: Inst::PSLLQ(Pq,Qq);
-                    0x4: Inst::PMULUDQ(Pq,Qq);
-                    0x5: pmaddwd_Pq_Qq();
-                    0x6: Inst::PSADBW(Pq,Qq);
-                    0x7: maskmovq_Pq_PRq();
-                    default: Inst::UD2();
-                }
-                // operand size (0x66)
-                0x1: decode OPCODE_OP_BOTTOM3 {
-                    0x1: Inst::PSLLW(Vo,Wo);
-                    0x2: Inst::PSLLD(Vo,Wo);
-                    0x3: Inst::PSLLQ(Vo,Wo);
-                    0x4: Inst::PMULUDQ(Vo,Wo);
-                    0x5: pmaddwd_Vo_Wo();
-                    0x6: Inst::PSADBW(Vo,Wo);
-                    0x7: maskmovdqu_Vo_VRo();
-                    default: Inst::UD2();
-                }
-                // repne (0xF2)
-                0x8: decode OPCODE_OP_BOTTOM3 {
-                    0x0: lddqu_Vo_Mo();
-                    default: Inst::UD2();
+                0x1E: decode LEGACY_DECODEVAL {
+                    // no prefix
+                    0x0: decode OPCODE_OP_BOTTOM3 {
+                        0x1: PSLLW(Pq,Qq);
+                        0x2: PSLLD(Pq,Qq);
+                        0x3: PSLLQ(Pq,Qq);
+                        0x4: PMULUDQ(Pq,Qq);
+                        0x5: PMADDWD(Pq,Qq);
+                        0x6: PSADBW(Pq,Qq);
+                        0x7: WarnUnimpl::maskmovq_Pq_PRq();
+                        default: UD2();
+                    }
+                    // operand size (0x66)
+                    0x1: decode OPCODE_OP_BOTTOM3 {
+                        0x1: PSLLW(Vo,Wo);
+                        0x2: PSLLD(Vo,Wo);
+                        0x3: PSLLQ(Vo,Wo);
+                        0x4: PMULUDQ(Vo,Wo);
+                        0x5: PMADDWD(Vo,Wo);
+                        0x6: PSADBW(Vo,Wo);
+                        0x7: WarnUnimpl::maskmovdqu_Vo_VRo();
+                        default: UD2();
+                    }
+                    // repne (0xF2)
+                    0x8: decode OPCODE_OP_BOTTOM3 {
+                        0x0: WarnUnimpl::lddqu_Vo_Mo();
+                        default: UD2();
+                    }
+                    default: UD2();
                 }
-                default: Inst::UD2();
-            }
-            format Inst {
                 0x1F: decode LEGACY_DECODEVAL {
                     // no prefix
                     0x0: decode OPCODE_OP_BOTTOM3 {
diff --git a/src/arch/x86/isa/insts/simd128/integer/arithmetic/multiply_add.py b/src/arch/x86/isa/insts/simd128/integer/arithmetic/multiply_add.py
index 40b38867b..f157d165f 100644
--- a/src/arch/x86/isa/insts/simd128/integer/arithmetic/multiply_add.py
+++ b/src/arch/x86/isa/insts/simd128/integer/arithmetic/multiply_add.py
@@ -54,5 +54,35 @@
 # Authors: Gabe Black
 
 microcode = '''
-# PMADDWD
+def macroop PMADDWD_XMM_XMM {
+    mmuli ufp3, xmml, xmmlm, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmml, xmmlm, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmml, ufp3, ufp4, size=4, ext=0
+    mmuli ufp3, xmmh, xmmhm, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmmh, xmmhm, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmmh, ufp3, ufp4, size=4, ext=0
+};
+
+def macroop PMADDWD_XMM_M {
+    ldfp ufp1, seg, sib, "DISPLACEMENT", dataSize=8
+    ldfp ufp2, seg, sib, "DISPLACEMENT + 8", dataSize=8
+    mmuli ufp3, xmml, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmml, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmml, ufp3, ufp4, size=4, ext=0
+    mmuli ufp3, xmmh, ufp2, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmmh, ufp2, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmmh, ufp3, ufp4, size=4, ext=0
+};
+
+def macroop PMADDWD_XMM_P {
+    rdip t7
+    ldfp ufp1, seg, riprel, "DISPLACEMENT", dataSize=8
+    ldfp ufp2, seg, riprel, "DISPLACEMENT + 8", dataSize=8
+    mmuli ufp3, xmml, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmml, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmml, ufp3, ufp4, size=4, ext=0
+    mmuli ufp3, xmmh, ufp2, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, xmmh, ufp2, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi xmmh, ufp3, ufp4, size=4, ext=0
+};
 '''
diff --git a/src/arch/x86/isa/insts/simd64/integer/arithmetic/multiply_add.py b/src/arch/x86/isa/insts/simd64/integer/arithmetic/multiply_add.py
index 40b38867b..f6940d159 100644
--- a/src/arch/x86/isa/insts/simd64/integer/arithmetic/multiply_add.py
+++ b/src/arch/x86/isa/insts/simd64/integer/arithmetic/multiply_add.py
@@ -54,5 +54,24 @@
 # Authors: Gabe Black
 
 microcode = '''
-# PMADDWD
+def macroop PMADDWD_MMX_MMX {
+    mmuli ufp3, mmx, mmxm, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, mmx, mmxm, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi mmx, ufp3, ufp4, size=4, ext=0
+};
+
+def macroop PMADDWD_MMX_M {
+    ldfp ufp1, seg, sib, "DISPLACEMENT", dataSize=8
+    mmuli ufp3, mmx, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, mmx, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi mmx, ufp3, ufp4, size=4, ext=0
+};
+
+def macroop PMADDWD_MMX_P {
+    rdip t7
+    ldfp ufp1, seg, riprel, "DISPLACEMENT", dataSize=8
+    mmuli ufp3, mmx, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10 | 0x20)
+    mmuli ufp4, mmx, ufp1, srcSize=2, destSize=4, ext=(0x2 | 0x10)
+    maddi mmx, ufp3, ufp4, size=4, ext=0
+};
 '''