Add 'cmp' SSE builtins and get rid of a bunch of other builtins.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@72032 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/AST/X86Builtins.def b/include/clang/AST/X86Builtins.def
index 73a4d1e0d021edb946fe7e1382c5fbd6bce9731c..fdfead5945cc21f9943042d638ac3283bd7ab60c 100644 (file)
--- a/include/clang/AST/X86Builtins.def
+++ b/include/clang/AST/X86Builtins.def
@@ -62,28 +62,8 @@ BUILTIN(__builtin_ia32_addss, "V4fV4fV4f", "")
 BUILTIN(__builtin_ia32_subss, "V4fV4fV4f", "")
 BUILTIN(__builtin_ia32_mulss, "V4fV4fV4f", "")
 BUILTIN(__builtin_ia32_divss, "V4fV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpeqps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpltps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpleps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpgtps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpgeps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpunordps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpneqps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpnltps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpnleps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpngtps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpngeps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpordps, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpeqss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpltss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpless, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpunordss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpneqss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpnltss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpnless, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpngtss, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpngess, "V4iV4fV4f", "")
-BUILTIN(__builtin_ia32_cmpordss, "V4iV4fV4f", "")
+BUILTIN(__builtin_ia32_cmpps, "V4fV4fV4fc", "")
+BUILTIN(__builtin_ia32_cmpss, "V4fV4fV4fc", "")
 BUILTIN(__builtin_ia32_minps, "V4fV4fV4f", "")
 BUILTIN(__builtin_ia32_maxps, "V4fV4fV4f", "")
 BUILTIN(__builtin_ia32_minss, "V4fV4fV4f", "")
@@ -146,26 +126,8 @@ BUILTIN(__builtin_ia32_addsd, "V2dV2dV2d", "")
 BUILTIN(__builtin_ia32_subsd, "V2dV2dV2d", "")
 BUILTIN(__builtin_ia32_mulsd, "V2dV2dV2d", "")
 BUILTIN(__builtin_ia32_divsd, "V2dV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpeqpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpltpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmplepd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpgtpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpgepd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpunordpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpneqpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpnltpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpnlepd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpngtpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpngepd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpordpd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpeqsd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpltsd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmplesd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpunordsd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpneqsd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpnltsd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpnlesd, "V4iV2dV2d", "")
-BUILTIN(__builtin_ia32_cmpordsd, "V4iV2dV2d", "")
+BUILTIN(__builtin_ia32_cmppd, "V2dV2dV2dc", "")
+BUILTIN(__builtin_ia32_cmpsd, "V2dV2dV2dc", "")
 BUILTIN(__builtin_ia32_minpd, "V2dV2dV2d", "")
 BUILTIN(__builtin_ia32_maxpd, "V2dV2dV2d", "")
 BUILTIN(__builtin_ia32_minsd, "V2dV2dV2d", "")

diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
index 93901fd5359de3967ef2f2ba5ea24efc21651b18..bb0d59acf690fc6e1d5f0aee0ac908692a8cf6fd 100644 (file)
--- a/lib/CodeGen/CGBuiltin.cpp
+++ b/lib/CodeGen/CGBuiltin.cpp
@@ -816,77 +816,13 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
   case X86::BI__builtin_ia32_vec_ext_v4hi:
   case X86::BI__builtin_ia32_vec_ext_v2df:
     return Builder.CreateExtractElement(Ops[0], Ops[1], "result");
-  case X86::BI__builtin_ia32_cmpordss:
-  case X86::BI__builtin_ia32_cmpordsd:
-  case X86::BI__builtin_ia32_cmpunordss:
-  case X86::BI__builtin_ia32_cmpunordsd:
-  case X86::BI__builtin_ia32_cmpeqss:
-  case X86::BI__builtin_ia32_cmpeqsd:
-  case X86::BI__builtin_ia32_cmpltss:
-  case X86::BI__builtin_ia32_cmpltsd:
-  case X86::BI__builtin_ia32_cmpless:
-  case X86::BI__builtin_ia32_cmplesd:
-  case X86::BI__builtin_ia32_cmpneqss:
-  case X86::BI__builtin_ia32_cmpneqsd:
-  case X86::BI__builtin_ia32_cmpnltss:
-  case X86::BI__builtin_ia32_cmpnltsd:
-  case X86::BI__builtin_ia32_cmpnless:
-  case X86::BI__builtin_ia32_cmpnlesd: {
-    unsigned i = 0;
-    const char *name = 0;
-    switch (BuiltinID) {
-    default: assert(0 && "Unknown compare builtin!");
-    case X86::BI__builtin_ia32_cmpeqss:
-    case X86::BI__builtin_ia32_cmpeqsd:
-      i = 0;
-      name = "cmpeq";
-      break;
-    case X86::BI__builtin_ia32_cmpltss:
-    case X86::BI__builtin_ia32_cmpltsd:
-      i = 1;
-      name = "cmplt";
-      break;
-    case X86::BI__builtin_ia32_cmpless:
-    case X86::BI__builtin_ia32_cmplesd:
-      i = 2;
-      name = "cmple";
-      break;
-    case X86::BI__builtin_ia32_cmpunordss:
-    case X86::BI__builtin_ia32_cmpunordsd:
-      i = 3;
-      name = "cmpunord";
-      break;
-    case X86::BI__builtin_ia32_cmpneqss:
-    case X86::BI__builtin_ia32_cmpneqsd:
-      i = 4;
-      name = "cmpneq";
-      break;
-    case X86::BI__builtin_ia32_cmpnltss:
-    case X86::BI__builtin_ia32_cmpnltsd:
-      i = 5;
-      name = "cmpntl";
-      break;
-    case X86::BI__builtin_ia32_cmpnless:
-    case X86::BI__builtin_ia32_cmpnlesd:
-      i = 6;
-      name = "cmpnle";
-      break;
-    case X86::BI__builtin_ia32_cmpordss:
-    case X86::BI__builtin_ia32_cmpordsd:
-      i = 7;
-      name = "cmpord";
-      break;
-    }
-
-    llvm::Function *F;
-    if (cast<llvm::VectorType>(Ops[0]->getType())->getElementType() ==
-        llvm::Type::FloatTy)
-      F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss);
-    else
-      F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd);
-
-    Ops.push_back(llvm::ConstantInt::get(llvm::Type::Int8Ty, i));
-    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name);
+  case X86::BI__builtin_ia32_cmpps: {
+    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps);
+    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), "cmpps");
+  }
+  case X86::BI__builtin_ia32_cmpss: {
+    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss);
+    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), "cmpss");
   }
   case X86::BI__builtin_ia32_ldmxcsr: {
     llvm::Type *PtrTy = llvm::PointerType::getUnqual(llvm::Type::Int8Ty);
@@ -904,89 +840,13 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
                              Builder.CreateBitCast(Tmp, PtrTy));
     return Builder.CreateLoad(Tmp, "stmxcsr");
   }
-  case X86::BI__builtin_ia32_cmpordps:
-  case X86::BI__builtin_ia32_cmpordpd:
-  case X86::BI__builtin_ia32_cmpunordps:
-  case X86::BI__builtin_ia32_cmpunordpd:
-  case X86::BI__builtin_ia32_cmpeqps: 
-  case X86::BI__builtin_ia32_cmpeqpd: 
-  case X86::BI__builtin_ia32_cmpltps: 
-  case X86::BI__builtin_ia32_cmpltpd: 
-  case X86::BI__builtin_ia32_cmpleps:
-  case X86::BI__builtin_ia32_cmplepd:
-  case X86::BI__builtin_ia32_cmpneqps:
-  case X86::BI__builtin_ia32_cmpneqpd:
-  case X86::BI__builtin_ia32_cmpngtps:
-  case X86::BI__builtin_ia32_cmpngtpd:
-  case X86::BI__builtin_ia32_cmpnltps: 
-  case X86::BI__builtin_ia32_cmpnltpd: 
-  case X86::BI__builtin_ia32_cmpgtps:
-  case X86::BI__builtin_ia32_cmpgtpd:
-  case X86::BI__builtin_ia32_cmpgeps:
-  case X86::BI__builtin_ia32_cmpgepd:
-  case X86::BI__builtin_ia32_cmpngeps:
-  case X86::BI__builtin_ia32_cmpngepd:
-  case X86::BI__builtin_ia32_cmpnleps: 
-  case X86::BI__builtin_ia32_cmpnlepd: {
-    unsigned i = 0;
-    const char *name = 0;
-    bool ShouldSwap = false;
-    switch (BuiltinID) {
-    default: assert(0 && "Unknown compare builtin!");
-    case X86::BI__builtin_ia32_cmpeqps:
-    case X86::BI__builtin_ia32_cmpeqpd:    i = 0; name = "cmpeq"; break;
-    case X86::BI__builtin_ia32_cmpltps:
-    case X86::BI__builtin_ia32_cmpltpd:    i = 1; name = "cmplt"; break;
-    case X86::BI__builtin_ia32_cmpleps:
-    case X86::BI__builtin_ia32_cmplepd:    i = 2; name = "cmple"; break;
-    case X86::BI__builtin_ia32_cmpunordps:
-    case X86::BI__builtin_ia32_cmpunordpd: i = 3; name = "cmpunord"; break;
-    case X86::BI__builtin_ia32_cmpneqps:
-    case X86::BI__builtin_ia32_cmpneqpd:   i = 4; name = "cmpneq"; break;
-    case X86::BI__builtin_ia32_cmpnltps:
-    case X86::BI__builtin_ia32_cmpnltpd:   i = 5; name = "cmpntl"; break;
-    case X86::BI__builtin_ia32_cmpnleps:
-    case X86::BI__builtin_ia32_cmpnlepd:   i = 6; name = "cmpnle"; break;
-    case X86::BI__builtin_ia32_cmpordps:
-    case X86::BI__builtin_ia32_cmpordpd:   i = 7; name = "cmpord"; break;
-    case X86::BI__builtin_ia32_cmpgtps:
-    case X86::BI__builtin_ia32_cmpgtpd:
-      ShouldSwap = true;
-      i = 1;
-      name = "cmpgt";
-      break;
-    case X86::BI__builtin_ia32_cmpgeps:
-    case X86::BI__builtin_ia32_cmpgepd:
-      i = 2;
-      name = "cmpge";
-      ShouldSwap = true;
-      break;
-    case X86::BI__builtin_ia32_cmpngtps:
-    case X86::BI__builtin_ia32_cmpngtpd:
-      i = 5;
-      name = "cmpngt";
-      ShouldSwap = true;
-      break;
-    case X86::BI__builtin_ia32_cmpngeps:
-    case X86::BI__builtin_ia32_cmpngepd:
-      i = 6;
-      name = "cmpnge";
-      ShouldSwap = true;
-      break;
-    }
-
-    if (ShouldSwap)
-      std::swap(Ops[0], Ops[1]);
-
-    llvm::Function *F;
-    if (cast<llvm::VectorType>(Ops[0]->getType())->getElementType() ==
-        llvm::Type::FloatTy)
-      F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps);
-    else
-      F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd);
-    
-    Ops.push_back(llvm::ConstantInt::get(llvm::Type::Int8Ty, i));
-    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name);
+  case X86::BI__builtin_ia32_cmppd: {
+    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd);
+    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), "cmppd");
+  }
+  case X86::BI__builtin_ia32_cmpsd: {
+    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd);
+    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), "cmpsd");
   }
   case X86::BI__builtin_ia32_movss:
     return EmitShuffleVector(Ops[0], Ops[1], 4, 1, 2, 3, "movss");

diff --git a/lib/Headers/emmintrin.h b/lib/Headers/emmintrin.h
index 84ce06a92fc8e0f06109714b40f0fb527367466d..12b548db7d3c3fd2ac3f14cae7ad2d5a59d2ddb0 100644 (file)
--- a/lib/Headers/emmintrin.h
+++ b/lib/Headers/emmintrin.h
@@ -149,145 +149,145 @@ _mm_xor_pd(__m128d a, __m128d b)
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpeqpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 0);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpltpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 1);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmplepd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 2);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpltpd(b, a);
+  return (__m128d)__builtin_ia32_cmppd(b, a, 1);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmplepd(b, a);
+  return (__m128d)__builtin_ia32_cmppd(b, a, 2);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpordpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 7);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpunordpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 3);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpneqpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 4);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnltpd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 5);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnlepd(a, b);
+  return (__m128d)__builtin_ia32_cmppd(a, b, 6);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnltpd(b, a);
+  return (__m128d)__builtin_ia32_cmppd(b, a, 5);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_pd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnlepd(b, a);
+  return (__m128d)__builtin_ia32_cmppd(b, a, 6);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpeqsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 0);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpltsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 1);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmplesd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 2);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpltsd(b, a);
+  return (__m128d)__builtin_ia32_cmpsd(b, a, 1);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmplesd(b, a);
+  return (__m128d)__builtin_ia32_cmpsd(b, a, 2);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpordsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 7);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpunordsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 3);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpneqsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 4);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnltsd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 5);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnlesd(a, b);
+  return (__m128d)__builtin_ia32_cmpsd(a, b, 6);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnltsd(b, a);
+  return (__m128d)__builtin_ia32_cmpsd(b, a, 5);
 }
 
 static inline __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_sd(__m128d a, __m128d b)
 {
-  return (__m128d)__builtin_ia32_cmpnlesd(b, a);
+  return (__m128d)__builtin_ia32_cmpsd(b, a, 6);
 }
 
 static inline int __attribute__((__always_inline__, __nodebug__))

diff --git a/lib/Headers/xmmintrin.h b/lib/Headers/xmmintrin.h
index c863144d60877b201780cc1011309cf1396e6744..264d2d68fde9ea47b9dc83ba3c5edd1b0e06beb7 100644 (file)
--- a/lib/Headers/xmmintrin.h
+++ b/lib/Headers/xmmintrin.h
@@ -170,145 +170,145 @@ _mm_xor_ps(__m128 a, __m128 b)
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpeqss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 0);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpeqps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 0);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpltss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 1);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpltps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 1);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpless(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 2);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpleps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 2);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpltss(b, a);
+  return (__m128)__builtin_ia32_cmpss(b, a, 1);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpltps(b, a);
+  return (__m128)__builtin_ia32_cmpps(b, a, 1);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpless(b, a);
+  return (__m128)__builtin_ia32_cmpss(b, a, 2);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpleps(b, a);
+  return (__m128)__builtin_ia32_cmpps(b, a, 2);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpneqss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 4);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpneqps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 4);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnltss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 5);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnltps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 5);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnless(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 6);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnleps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 6);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnltss(b, a);
+  return (__m128)__builtin_ia32_cmpss(b, a, 5);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnltps(b, a);
+  return (__m128)__builtin_ia32_cmpps(b, a, 5);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnless(b, a);
+  return (__m128)__builtin_ia32_cmpss(b, a, 6);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpnleps(b, a);
+  return (__m128)__builtin_ia32_cmpps(b, a, 6);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpordss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 7);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpordps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 7);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_ss(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpunordss(a, b);
+  return (__m128)__builtin_ia32_cmpss(a, b, 3);
 }
 
 static inline __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_ps(__m128 a, __m128 b)
 {
-  return (__m128)__builtin_ia32_cmpunordps(a, b);
+  return (__m128)__builtin_ia32_cmpps(a, b, 3);
 }
 
 static inline int __attribute__((__always_inline__, __nodebug__))

diff --git a/test/CodeGen/builtins-x86.c b/test/CodeGen/builtins-x86.c
index 17c2840a49a4f0e05d992d49dd62f45ffd053094..d710b07deacd864f939bb05da4516cc91d2b2cb0 100644 (file)
--- a/test/CodeGen/builtins-x86.c
+++ b/test/CodeGen/builtins-x86.c
@@ -109,30 +109,22 @@ void f0() {
   tmp_V4f = __builtin_ia32_subss(tmp_V4f, tmp_V4f);
   tmp_V4f = __builtin_ia32_mulss(tmp_V4f, tmp_V4f);
   tmp_V4f = __builtin_ia32_divss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpeqps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpltps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpleps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpgtps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpgeps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpunordps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpneqps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpnltps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpnleps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpngtps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpngeps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpordps(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpeqss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpltss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpless(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpunordss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpneqss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpnltss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpnless(tmp_V4f, tmp_V4f);
-#ifdef USE_ALL
-  tmp_V4i = __builtin_ia32_cmpngtss(tmp_V4f, tmp_V4f);
-  tmp_V4i = __builtin_ia32_cmpngess(tmp_V4f, tmp_V4f);
-#endif
-  tmp_V4i = __builtin_ia32_cmpordss(tmp_V4f, tmp_V4f);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 0);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 1);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 2);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 3);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 4);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 5);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 6);
+  tmp_V4f = __builtin_ia32_cmpps(tmp_V4f, tmp_V4f, 7);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 0);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 1);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 2);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 3);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 4);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 5);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 6);
+  tmp_V4f = __builtin_ia32_cmpss(tmp_V4f, tmp_V4f, 7);
   tmp_V4f = __builtin_ia32_minps(tmp_V4f, tmp_V4f);
   tmp_V4f = __builtin_ia32_maxps(tmp_V4f, tmp_V4f);
   tmp_V4f = __builtin_ia32_minss(tmp_V4f, tmp_V4f);
@@ -196,26 +188,22 @@ void f0() {
   tmp_V2d = __builtin_ia32_subsd(tmp_V2d, tmp_V2d);
   tmp_V2d = __builtin_ia32_mulsd(tmp_V2d, tmp_V2d);
   tmp_V2d = __builtin_ia32_divsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpeqpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpltpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmplepd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpgtpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpgepd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpunordpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpneqpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpnltpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpnlepd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpngtpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpngepd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpordpd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpeqsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpltsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmplesd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpunordsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpneqsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpnltsd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpnlesd(tmp_V2d, tmp_V2d);
-  tmp_V4i = __builtin_ia32_cmpordsd(tmp_V2d, tmp_V2d);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 0);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 1);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 2);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 3);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 4);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 5);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 6);
+  tmp_V2d = __builtin_ia32_cmppd(tmp_V2d, tmp_V2d, 7);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 0);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 1);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 2);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 3);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 4);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 5);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 6);
+  tmp_V2d = __builtin_ia32_cmpsd(tmp_V2d, tmp_V2d, 7);
   tmp_V2d = __builtin_ia32_minpd(tmp_V2d, tmp_V2d);
   tmp_V2d = __builtin_ia32_maxpd(tmp_V2d, tmp_V2d);
   tmp_V2d = __builtin_ia32_minsd(tmp_V2d, tmp_V2d);