Add AVX SSE4.1 binop (some forms of packed max,min,mul,pack,cmp) instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107558 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 206e3979534b9d44494bab0dbb4027d257d7be10..b5e54e6bebd6facd1dada8408b4cf61b8bca7e05 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -4123,6 +4123,32 @@ multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
           (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
 }
 
+let isAsmParserOnly = 1, Predicates = [HasAVX, HasSSE41] in {
+  let isCommutable = 0 in
+  defm VPACKUSDW : SS41I_binop_rm_int<0x2B, "vpackusdw", int_x86_sse41_packusdw,
+                                                         0>, VEX_4V;
+  defm VPCMPEQQ  : SS41I_binop_rm_int<0x29, "vpcmpeqq",  int_x86_sse41_pcmpeqq,
+                                                         0>, VEX_4V;
+  defm VPMINSB   : SS41I_binop_rm_int<0x38, "vpminsb",   int_x86_sse41_pminsb,
+                                                         0>, VEX_4V;
+  defm VPMINSD   : SS41I_binop_rm_int<0x39, "vpminsd",   int_x86_sse41_pminsd,
+                                                         0>, VEX_4V;
+  defm VPMINUD   : SS41I_binop_rm_int<0x3B, "vpminud",   int_x86_sse41_pminud,
+                                                         0>, VEX_4V;
+  defm VPMINUW   : SS41I_binop_rm_int<0x3A, "vpminuw",   int_x86_sse41_pminuw,
+                                                         0>, VEX_4V;
+  defm VPMAXSB   : SS41I_binop_rm_int<0x3C, "vpmaxsb",   int_x86_sse41_pmaxsb,
+                                                         0>, VEX_4V;
+  defm VPMAXSD   : SS41I_binop_rm_int<0x3D, "vpmaxsd",   int_x86_sse41_pmaxsd,
+                                                         0>, VEX_4V;
+  defm VPMAXUD   : SS41I_binop_rm_int<0x3F, "vpmaxud",   int_x86_sse41_pmaxud,
+                                                         0>, VEX_4V;
+  defm VPMAXUW   : SS41I_binop_rm_int<0x3E, "vpmaxuw",   int_x86_sse41_pmaxuw,
+                                                         0>, VEX_4V;
+  defm VPMULDQ   : SS41I_binop_rm_int<0x28, "vpmuldq",   int_x86_sse41_pmuldq,
+                                                         0>, VEX_4V;
+}
+
 let Constraints = "$src1 = $dst" in {
   let isCommutable = 0 in
   defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw", int_x86_sse41_packusdw>;

diff --git a/test/MC/AsmParser/X86/x86_32-encoding.s b/test/MC/AsmParser/X86/x86_32-encoding.s
index e1085a3074053b554c36cb73cf4d10e66c720494..28273ee4d28694268fa86e1f4dcd92b8176ecd12 100644 (file)
--- a/test/MC/AsmParser/X86/x86_32-encoding.s
+++ b/test/MC/AsmParser/X86/x86_32-encoding.s
@@ -11782,3 +11782,91 @@
 // CHECK: encoding: [0xc4,0xe2,0x79,0x41,0x10]
           vphminposuw  (%eax), %xmm2
 
+// CHECK: vpackusdw  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x2b,0xca]
+          vpackusdw  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpackusdw  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x2b,0x18]
+          vpackusdw  (%eax), %xmm2, %xmm3
+
+// CHECK: vpcmpeqq  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x29,0xca]
+          vpcmpeqq  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpcmpeqq  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x29,0x18]
+          vpcmpeqq  (%eax), %xmm2, %xmm3
+
+// CHECK: vpminsb  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x38,0xca]
+          vpminsb  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpminsb  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x38,0x18]
+          vpminsb  (%eax), %xmm2, %xmm3
+
+// CHECK: vpminsd  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x39,0xca]
+          vpminsd  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpminsd  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x39,0x18]
+          vpminsd  (%eax), %xmm2, %xmm3
+
+// CHECK: vpminud  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3b,0xca]
+          vpminud  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpminud  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3b,0x18]
+          vpminud  (%eax), %xmm2, %xmm3
+
+// CHECK: vpminuw  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3a,0xca]
+          vpminuw  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpminuw  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3a,0x18]
+          vpminuw  (%eax), %xmm2, %xmm3
+
+// CHECK: vpmaxsb  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3c,0xca]
+          vpmaxsb  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpmaxsb  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3c,0x18]
+          vpmaxsb  (%eax), %xmm2, %xmm3
+
+// CHECK: vpmaxsd  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3d,0xca]
+          vpmaxsd  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpmaxsd  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3d,0x18]
+          vpmaxsd  (%eax), %xmm2, %xmm3
+
+// CHECK: vpmaxud  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3f,0xca]
+          vpmaxud  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpmaxud  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3f,0x18]
+          vpmaxud  (%eax), %xmm2, %xmm3
+
+// CHECK: vpmaxuw  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x3e,0xca]
+          vpmaxuw  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpmaxuw  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x3e,0x18]
+          vpmaxuw  (%eax), %xmm2, %xmm3
+
+// CHECK: vpmuldq  %xmm2, %xmm3, %xmm1
+// CHECK: encoding: [0xc4,0xe2,0x61,0x28,0xca]
+          vpmuldq  %xmm2, %xmm3, %xmm1
+
+// CHECK: vpmuldq  (%eax), %xmm2, %xmm3
+// CHECK: encoding: [0xc4,0xe2,0x69,0x28,0x18]
+          vpmuldq  (%eax), %xmm2, %xmm3
+

diff --git a/test/MC/AsmParser/X86/x86_64-encoding.s b/test/MC/AsmParser/X86/x86_64-encoding.s
index ad52ffb36a63d4c17bad162ac42b74a82294c75f..b676eab740a20826e1b6afb7f40ab180c7e7afc1 100644 (file)
--- a/test/MC/AsmParser/X86/x86_64-encoding.s
+++ b/test/MC/AsmParser/X86/x86_64-encoding.s
@@ -1830,3 +1830,91 @@ pshufb   CPI1_0(%rip), %xmm1
 // CHECK: encoding: [0xc4,0x62,0x79,0x41,0x20]
           vphminposuw  (%rax), %xmm12
 
+// CHECK: vpackusdw  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x2b,0xdc]
+          vpackusdw  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpackusdw  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x2b,0x28]
+          vpackusdw  (%rax), %xmm12, %xmm13
+
+// CHECK: vpcmpeqq  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x29,0xdc]
+          vpcmpeqq  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpcmpeqq  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x29,0x28]
+          vpcmpeqq  (%rax), %xmm12, %xmm13
+
+// CHECK: vpminsb  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x38,0xdc]
+          vpminsb  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpminsb  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x38,0x28]
+          vpminsb  (%rax), %xmm12, %xmm13
+
+// CHECK: vpminsd  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x39,0xdc]
+          vpminsd  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpminsd  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x39,0x28]
+          vpminsd  (%rax), %xmm12, %xmm13
+
+// CHECK: vpminud  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3b,0xdc]
+          vpminud  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpminud  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3b,0x28]
+          vpminud  (%rax), %xmm12, %xmm13
+
+// CHECK: vpminuw  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3a,0xdc]
+          vpminuw  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpminuw  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3a,0x28]
+          vpminuw  (%rax), %xmm12, %xmm13
+
+// CHECK: vpmaxsb  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3c,0xdc]
+          vpmaxsb  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpmaxsb  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3c,0x28]
+          vpmaxsb  (%rax), %xmm12, %xmm13
+
+// CHECK: vpmaxsd  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3d,0xdc]
+          vpmaxsd  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpmaxsd  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3d,0x28]
+          vpmaxsd  (%rax), %xmm12, %xmm13
+
+// CHECK: vpmaxud  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3f,0xdc]
+          vpmaxud  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpmaxud  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3f,0x28]
+          vpmaxud  (%rax), %xmm12, %xmm13
+
+// CHECK: vpmaxuw  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x3e,0xdc]
+          vpmaxuw  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpmaxuw  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x3e,0x28]
+          vpmaxuw  (%rax), %xmm12, %xmm13
+
+// CHECK: vpmuldq  %xmm12, %xmm13, %xmm11
+// CHECK: encoding: [0xc4,0x42,0x11,0x28,0xdc]
+          vpmuldq  %xmm12, %xmm13, %xmm11
+
+// CHECK: vpmuldq  (%rax), %xmm12, %xmm13
+// CHECK: encoding: [0xc4,0x62,0x19,0x28,0x28]
+          vpmuldq  (%rax), %xmm12, %xmm13
+