From beb41281f8355caa05700d0a77539defbdf428f8 Mon Sep 17 00:00:00 2001
From: John McCall <rjmccall@apple.com>
Date: Wed, 7 Apr 2010 08:20:20 +0000
Subject: [PATCH] @llvm.sqrt isn't really close enough to C's sqrt to justify
 emitting calls to the intrinsic, even when math-errno is off.

Fixes rdar://problem/7828230 by falling back on the library function.



git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@100613 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/CGBuiltin.cpp | 13 +++++-----
 test/CodeGen/libcalls.c   | 50 +++++++++++++++++++++++++++++++--------
 2 files changed, 46 insertions(+), 17 deletions(-)

diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
index 38c40ed489..a46cc39efa 100644
--- a/lib/CodeGen/CGBuiltin.cpp
+++ b/lib/CodeGen/CGBuiltin.cpp
@@ -682,13 +682,12 @@ RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
   case Builtin::BIsqrt:
   case Builtin::BIsqrtf:
   case Builtin::BIsqrtl: {
-    // Rewrite sqrt to intrinsic if allowed.
-    if (!FD->hasAttr<ConstAttr>())
-      break;
-    Value *Arg0 = EmitScalarExpr(E->getArg(0));
-    const llvm::Type *ArgType = Arg0->getType();
-    Value *F = CGM.getIntrinsic(Intrinsic::sqrt, &ArgType, 1);
-    return RValue::get(Builder.CreateCall(F, Arg0, "tmp"));
+    // TODO: there is currently no set of optimizer flags
+    // sufficient for us to rewrite sqrt to @llvm.sqrt.
+    // -fmath-errno=0 is not good enough; we need finiteness.
+    // We could probably precondition the call with an ult
+    // against 0, but is that worth the complexity?
+    break;
   }
 
   case Builtin::BIpow:
diff --git a/test/CodeGen/libcalls.c b/test/CodeGen/libcalls.c
index a96176afe5..828d7de6cb 100644
--- a/test/CodeGen/libcalls.c
+++ b/test/CodeGen/libcalls.c
@@ -1,22 +1,52 @@
-// RUN: %clang_cc1 -fmath-errno -emit-llvm -o %t %s -triple i386-unknown-unknown
-// RUN: grep "declare " %t | count 6
-// RUN: grep "declare " %t | grep "@llvm." | count 1
-// RUN: %clang_cc1 -emit-llvm -o %t %s -triple i386-unknown-unknown
-// RUN: grep "declare " %t | count 6
-// RUN: grep "declare " %t | grep -v "@llvm." | count 0
-
-// IRgen only pays attention to const; it should always call llvm for
-// this.
-float sqrtf(float) __attribute__((const));
+// RUN: %clang_cc1 -fmath-errno -emit-llvm -o - %s -triple i386-unknown-unknown | FileCheck -check-prefix YES %s
+// RUN: %clang_cc1 -emit-llvm -o - %s -triple i386-unknown-unknown | FileCheck -check-prefix NO %s
 
+// CHECK-YES: define void @test_sqrt
+// CHECK-NO: define void @test_sqrt
 void test_sqrt(float a0, double a1, long double a2) {
+  // Following llvm-gcc's lead, we never emit these as intrinsics;
+  // no-math-errno isn't good enough.  We could probably use intrinsics
+  // with appropriate guards if it proves worthwhile.
+
+  // CHECK-YES: call float @sqrtf
+  // CHECK-NO: call float @sqrtf
   float l0 = sqrtf(a0);
+
+  // CHECK-YES: call double @sqrt
+  // CHECK-NO: call double @sqrt
   double l1 = sqrt(a1);
+
+  // CHECK-YES: call x86_fp80 @sqrtl
+  // CHECK-NO: call x86_fp80 @sqrtl
   long double l2 = sqrtl(a2);
 }
 
+// CHECK-YES: declare float @sqrtf(float)
+// CHECK-YES: declare double @sqrt(double)
+// CHECK-YES: declare x86_fp80 @sqrtl(x86_fp80)
+// CHECK-NO: declare float @sqrtf(float) readnone
+// CHECK-NO: declare double @sqrt(double) readnone
+// CHECK-NO: declare x86_fp80 @sqrtl(x86_fp80) readnone
+
+// CHECK-YES: define void @test_pow
+// CHECK-NO: define void @test_pow
 void test_pow(float a0, double a1, long double a2) {
+  // CHECK-YES: call float @powf
+  // CHECK-NO: call float @llvm.pow.f32
   float l0 = powf(a0, a0);
+
+  // CHECK-YES: call double @pow
+  // CHECK-NO: call double @llvm.pow.f64
   double l1 = pow(a1, a1);
+
+  // CHECK-YES: call x86_fp80 @powl
+  // CHECK-NO: call x86_fp80 @llvm.pow.f80
   long double l2 = powl(a2, a2);
 }
+
+// CHECK-YES: declare float @powf(float, float)
+// CHECK-YES: declare double @pow(double, double)
+// CHECK-YES: declare x86_fp80 @powl(x86_fp80, x86_fp80)
+// CHECK-NO: declare float @llvm.pow.f32(float, float) nounwind readonly
+// CHECK-NO: declare double @llvm.pow.f64(double, double) nounwind readonly
+// CHECK-NO: declare x86_fp80 @llvm.pow.f80(x86_fp80, x86_fp80) nounwind readonly