Use logical/arithmetic operations instead of builtins in tbmintrin.h. This way we can remove the intrinsic support from the backend.

llvm-svn: 192036

3 files changed, 93 insertions, 103 deletions

diff --git a/clang/include/clang/Basic/BuiltinsX86.def b/clang/include/clang/Basic/BuiltinsX86.def
index dea1cc6b16d..51397fa45df 100644
--- a/clang/include/clang/Basic/BuiltinsX86.def
+++ b/clang/include/clang/Basic/BuiltinsX86.def
@@ -645,24 +645,6 @@ BUILTIN(__builtin_ia32_pext_di, "ULLiULLiULLi", "")
 // TBM
 BUILTIN(__builtin_ia32_bextri_u32, "UiUiIUi", "")
 BUILTIN(__builtin_ia32_bextri_u64, "ULLiULLiIULLi", "")
-BUILTIN(__builtin_ia32_blcfill_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blcfill_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blci_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blci_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blcic_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blcic_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blcmsk_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blcmsk_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blcs_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blcs_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blsfill_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blsfill_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_blsic_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_blsic_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_t1mskc_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_t1mskc_u64, "ULLiULLi", "")
-BUILTIN(__builtin_ia32_tzmsk_u32, "UiUi", "")
-BUILTIN(__builtin_ia32_tzmsk_u64, "ULLiULLi", "")
 
 // SHA
 BUILTIN(__builtin_ia32_sha1rnds4, "V4iV4iV4iIc", "")
diff --git a/clang/lib/Headers/tbmintrin.h b/clang/lib/Headers/tbmintrin.h
index 56e13be9cff..f95e34fbc18 100644
--- a/clang/lib/Headers/tbmintrin.h
+++ b/clang/lib/Headers/tbmintrin.h
@@ -34,142 +34,124 @@
 
 #define __bextri_u32(a, b) (__builtin_ia32_bextri_u32((a), (b)))
 
-#ifdef __x86_64__
-#define __bextri_u64(a, b) (__builtin_ia32_bextri_u64((a), (int)(b)))
-#endif
-
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
 __blcfill_u32(unsigned int a)
 {
-  return __builtin_ia32_blcfill_u32(a);
+  return a & (a + 1);
 }
 
-#ifdef __x86_64__
-static __inline__ unsigned long long __attribute__((__always_inline__,
-                                                    __nodebug__))
-__blcfill_u64(unsigned long long a)
+static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
+__blci_u32(unsigned int a)
 {
-  return __builtin_ia32_blcfill_u64(a);
+  return a | ~(a + 1);
 }
-#endif
 
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blci_u32(unsigned int a)
+__blcic_u32(unsigned int a)
 {
-  return __builtin_ia32_blci_u32(a);
+  return ~a & (a + 1);
 }
 
-#ifdef __x86_64__
-static __inline__ unsigned long long __attribute__((__always_inline__,
-                                                    __nodebug__))
-__blci_u64(unsigned long long a)
+static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
+__blcmsk_u32(unsigned int a)
 {
-  return __builtin_ia32_blci_u64(a);
+  return a ^ (a + 1);
 }
-#endif
 
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blcic_u32(unsigned int a)
+__blcs_u32(unsigned int a)
 {
-  return __builtin_ia32_blcic_u32(a);
+  return a | (a + 1);
 }
 
-#ifdef __x86_64__
-static __inline__ unsigned long long __attribute__((__always_inline__,
-                                                    __nodebug__))
-__blcic_u64(unsigned long long a)
+static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
+__blsfill_u32(unsigned int a)
 {
-  return __builtin_ia32_blcic_u64(a);
+  return a | (a - 1);
 }
-#endif
 
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blcmsk_u32(unsigned int a)
+__blsic_u32(unsigned int a)
 {
-  return __builtin_ia32_blcmsk_u32(a);
+  return ~a | (a - 1);
 }
 
-#ifdef __x86_64__
-static __inline__ unsigned long long __attribute__((__always_inline__,
-                                                    __nodebug__))
-__blcmsk_u64(unsigned long long a)
+static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
+__t1mskc_u32(unsigned int a)
 {
-  return __builtin_ia32_blcmsk_u64(a);
+  return ~a | (a + 1);
 }
-#endif
 
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blcs_u32(unsigned int a)
+__tzmsk_u32(unsigned int a)
 {
-  return __builtin_ia32_blcs_u32(a);
+  return ~a & (a - 1);
 }
 
 #ifdef __x86_64__
+#define __bextri_u64(a, b) (__builtin_ia32_bextri_u64((a), (int)(b)))
+
 static __inline__ unsigned long long __attribute__((__always_inline__,
                                                     __nodebug__))
-__blcs_u64(unsigned long long a)
+__blcfill_u64(unsigned long long a)
 {
-  return __builtin_ia32_blcs_u64(a);
+  return a & (a + 1);
 }
-#endif
 
-static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blsfill_u32(unsigned int a)
+static __inline__ unsigned long long __attribute__((__always_inline__,
+                                                    __nodebug__))
+__blci_u64(unsigned long long a)
 {
-  return __builtin_ia32_blsfill_u32(a);
+  return a | ~(a + 1);
 }
 
-#ifdef __x86_64__
 static __inline__ unsigned long long __attribute__((__always_inline__,
                                                     __nodebug__))
-__blsfill_u64(unsigned long long a)
+__blcic_u64(unsigned long long a)
 {
-  return __builtin_ia32_blsfill_u64(a);
+  return ~a & (a + 1);
 }
-#endif
 
-static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__blsic_u32(unsigned int a)
+static __inline__ unsigned long long __attribute__((__always_inline__,
+                                                    __nodebug__))
+__blcmsk_u64(unsigned long long a)
 {
-  return __builtin_ia32_blsic_u32(a);
+  return a ^ (a + 1);
 }
 
-#ifdef __x86_64__
 static __inline__ unsigned long long __attribute__((__always_inline__,
                                                     __nodebug__))
-__blsic_u64(unsigned long long a)
+__blcs_u64(unsigned long long a)
 {
-  return __builtin_ia32_blsic_u64(a);
+  return a | (a + 1);
 }
-#endif
 
-static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__t1mskc_u32(unsigned int a)
+static __inline__ unsigned long long __attribute__((__always_inline__,
+                                                    __nodebug__))
+__blsfill_u64(unsigned long long a)
 {
-  return __builtin_ia32_t1mskc_u32(a);
+  return a | (a - 1);
 }
 
-#ifdef __x86_64__
 static __inline__ unsigned long long __attribute__((__always_inline__,
                                                     __nodebug__))
-__t1mskc_u64(unsigned long long a)
+__blsic_u64(unsigned long long a)
 {
-  return __builtin_ia32_t1mskc_u64(a);
+  return ~a | (a - 1);
 }
-#endif
 
-static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
-__tzmsk_u32(unsigned int a)
+static __inline__ unsigned long long __attribute__((__always_inline__,
+                                                    __nodebug__))
+__t1mskc_u64(unsigned long long a)
 {
-  return __builtin_ia32_tzmsk_u32(a);
+  return ~a | (a + 1);
 }
 
-#ifdef __x86_64__
 static __inline__ unsigned long long __attribute__((__always_inline__,
                                                     __nodebug__))
 __tzmsk_u64(unsigned long long a)
 {
-  return __builtin_ia32_tzmsk_u64(a);
+  return ~a & (a - 1);
 }
 #endif
 
diff --git a/clang/test/CodeGen/tbm-builtins.c b/clang/test/CodeGen/tbm-builtins.c
index 109c7f897a2..e3a702161ec 100644
--- a/clang/test/CodeGen/tbm-builtins.c
+++ b/clang/test/CodeGen/tbm-builtins.c
@@ -1,4 +1,4 @@
-// RUN: %clang_cc1 %s -triple=x86_64-unknown-unknown -target-feature +tbm -emit-llvm -o - | FileCheck %s
+// RUN: %clang_cc1 %s -O3 -triple=x86_64-unknown-unknown -target-feature +tbm -emit-llvm -o - | FileCheck %s
 
 // Don't include mm_malloc.h, it's system specific.
 #define __MM_MALLOC_H
@@ -21,91 +21,117 @@ unsigned long long test__bextri_u64_bigint(unsigned long long a) {
 }
 
 unsigned int test__blcfill_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blcfill.u32
+  // CHECK: [[TMP:%.*]] = add i32 [[SRC:%.*]], 1
+  // CHECK-NEXT: %{{.*}} = and i32 [[TMP]], [[SRC]]
   return __blcfill_u32(a);
 }
 
 unsigned long long test__blcfill_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blcfill.u64
+  // CHECK: [[TMPT:%.*]] = add i64 [[SRC:%.*]], 1
+  // CHECK-NEXT: %{{.*}} = and i64 [[TMP]], [[SRC]]
   return __blcfill_u64(a);
 }
 
 unsigned int test__blci_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blci.u32
+  // CHECK: [[TMP:%.*]] = sub i32 -2, [[SRC:%.*]]
+  // CHECK-NEXT: %{{.*}} = or i32 [[TMP]], [[SRC]]
   return __blci_u32(a);
 }
 
 unsigned long long test__blci_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blci.u64
+  // CHECK: [[TMP:%.*]] = sub i64 -2, [[SRC:%.*]]
+  // CHECK-NEXT: %{{.*}} = or i64 [[TMP]], [[SRC]]
   return __blci_u64(a);
 }
 
 unsigned int test__blcic_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blcic.u32
+  // CHECK: [[TMP1:%.*]] = xor i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i32 [[SRC]], 1
+  // CHECK-NEXT: {{.*}} = and i32 [[TMP2]], [[TMP1]]
   return __blcic_u32(a);
 }
 
 unsigned long long test__blcic_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blcic.u64
+  // CHECK: [[TMP1:%.*]] = xor i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i64 [[SRC]], 1
+  // CHECK-NEXT: {{.*}} = and i64 [[TMP2]], [[TMP1]]
   return __blcic_u64(a);
 }
 
 unsigned int test__blcmsk_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blcmsk.u32
+  // CHECK: [[TMP:%.*]] = add i32 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = xor i32 [[TMP]], [[SRC]]
   return __blcmsk_u32(a);
 }
 
 unsigned long long test__blcmsk_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blcmsk.u64
+  // CHECK: [[TMP:%.*]] = add i64 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = xor i64 [[TMP]], [[SRC]]
   return __blcmsk_u64(a);
 }
 
 unsigned int test__blcs_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blcs.u32
+  // CHECK: [[TMP:%.*]] = add i32 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = or i32 [[TMP]], [[SRC]]
   return __blcs_u32(a);
 }
 
 unsigned long long test__blcs_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blcs.u64
+  // CHECK: [[TMP:%.*]] = add i64 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = or i64 [[TMP]], [[SRC]]
   return __blcs_u64(a);
 }
 
 unsigned int test__blsfill_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blsfill.u32
+  // CHECK: [[TMP:%.*]] = add i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = or i32 [[TMP]], [[SRC]]
   return __blsfill_u32(a);
 }
 
 unsigned long long test__blsfill_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blsfill.u64
+  // CHECK: [[TMP:%.*]] = add i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = or i64 [[TMP]], [[SRC]]
   return __blsfill_u64(a);
 }
 
 unsigned int test__blsic_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.blsic.u32
+  // CHECK: [[TMP1:%.*]] = xor i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = or i32 [[TMP2]], [[TMP1]]
   return __blsic_u32(a);
 }
 
 unsigned long long test__blsic_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.blsic.u64
+  // CHECK: [[TMP1:%.*]] = xor i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = or i64 [[TMP2]], [[TMP1]]
   return __blsic_u64(a);
 }
 
 unsigned int test__t1mskc_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.t1mskc.u32
+  // CHECK: [[TMP1:%.*]] = xor i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i32 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = or i32 [[TMP2]], [[TMP1]]
   return __t1mskc_u32(a);
 }
 
 unsigned long long test__t1mskc_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.t1mskc.u64
+  // CHECK: [[TMP1:%.*]] = xor i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i64 [[SRC:%.*]], 1
+  // CHECK-NEXT: {{.*}} = or i64 [[TMP2]], [[TMP1]]
   return __t1mskc_u64(a);
 }
 
 unsigned int test__tzmsk_u32(unsigned int a) {
-  // CHECK: call i32 @llvm.x86.tbm.tzmsk.u32
+  // CHECK: [[TMP1:%.*]] = xor i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i32 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = and i32 [[TMP2]], [[TMP1]]
   return __tzmsk_u32(a);
 }
 
 unsigned long long test__tzmsk_u64(unsigned long long a) {
-  // CHECK: call i64 @llvm.x86.tbm.tzmsk.u64
+  // CHECK: [[TMP1:%.*]] = xor i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: [[TMP2:%.*]] = add i64 [[SRC:%.*]], -1
+  // CHECK-NEXT: {{.*}} = and i64 [[TMP2]], [[TMP1]]
   return __tzmsk_u64(a);
 }