Teach codegen to lower llvm.powi to an efficient (but not optimal)

multiply sequence when the power is a constant integer.  Before, our
codegen for std::pow(.., int) always turned into a libcall, which was
really inefficient.

This should also make many gfortran programs happier I'd imagine.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92388 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 29 insertions, 35 deletions

diff --git a/test/CodeGen/X86/2007-09-27-LDIntrinsics.ll b/test/CodeGen/X86/2007-09-27-LDIntrinsics.ll
index 4a56ee446a..4d6971586c 100644
--- a/test/CodeGen/X86/2007-09-27-LDIntrinsics.ll
+++ b/test/CodeGen/X86/2007-09-27-LDIntrinsics.ll
@@ -1,47 +1,30 @@
-; RUN: llc < %s | grep powixf2
-; RUN: llc < %s | grep fsqrt
-; ModuleID = 'yyy.c'
+; RUN: llc < %s | FileCheck %s
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
 target triple = "i686-apple-darwin8"
 
-define x86_fp80 @foo(x86_fp80 %x) {
+define x86_fp80 @foo(x86_fp80 %x) nounwind{
 entry:
-	%x_addr = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%retval = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%tmp = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%"alloca point" = bitcast i32 0 to i32		; <i32> [#uses=0]
-	store x86_fp80 %x, x86_fp80* %x_addr
-	%tmp1 = load x86_fp80* %x_addr, align 16		; <x86_fp80> [#uses=1]
-	%tmp2 = call x86_fp80 @llvm.sqrt.f80( x86_fp80 %tmp1 )		; <x86_fp80> [#uses=1]
-	store x86_fp80 %tmp2, x86_fp80* %tmp, align 16
-	%tmp3 = load x86_fp80* %tmp, align 16		; <x86_fp80> [#uses=1]
-	store x86_fp80 %tmp3, x86_fp80* %retval, align 16
-	br label %return
-
-return:		; preds = %entry
-	%retval4 = load x86_fp80* %retval		; <x86_fp80> [#uses=1]
-	ret x86_fp80 %retval4
+	%tmp2 = call x86_fp80 @llvm.sqrt.f80( x86_fp80 %x )
+	ret x86_fp80 %tmp2
+        
+; CHECK: foo:
+; CHECK: fldt 4(%esp)
+; CHECK-NEXT: fsqrt
+; CHECK-NEXT: ret
 }
 
 declare x86_fp80 @llvm.sqrt.f80(x86_fp80)
 
-define x86_fp80 @bar(x86_fp80 %x) {
+define x86_fp80 @bar(x86_fp80 %x) nounwind {
 entry:
-	%x_addr = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%retval = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%tmp = alloca x86_fp80		; <x86_fp80*> [#uses=2]
-	%"alloca point" = bitcast i32 0 to i32		; <i32> [#uses=0]
-	store x86_fp80 %x, x86_fp80* %x_addr
-	%tmp1 = load x86_fp80* %x_addr, align 16		; <x86_fp80> [#uses=1]
-	%tmp2 = call x86_fp80 @llvm.powi.f80( x86_fp80 %tmp1, i32 3 )		; <x86_fp80> [#uses=1]
-	store x86_fp80 %tmp2, x86_fp80* %tmp, align 16
-	%tmp3 = load x86_fp80* %tmp, align 16		; <x86_fp80> [#uses=1]
-	store x86_fp80 %tmp3, x86_fp80* %retval, align 16
-	br label %return
-
-return:		; preds = %entry
-	%retval4 = load x86_fp80* %retval		; <x86_fp80> [#uses=1]
-	ret x86_fp80 %retval4
+	%tmp2 = call x86_fp80 @llvm.powi.f80( x86_fp80 %x, i32 3 )
+	ret x86_fp80 %tmp2
+; CHECK: bar:
+; CHECK: fldt 4(%esp)
+; CHECK-NEXT: fld	%st(0)
+; CHECK-NEXT: fmul	%st(1)
+; CHECK-NEXT: fmulp	%st(1)
+; CHECK-NEXT: ret
 }
 
 declare x86_fp80 @llvm.powi.f80(x86_fp80, i32)
diff --git a/test/CodeGen/X86/powi.ll b/test/CodeGen/X86/powi.ll
new file mode 100644
index 0000000000..c3d68312ce
--- /dev/null
+++ b/test/CodeGen/X86/powi.ll
@@ -0,0 +1,11 @@
+; RUN: llc %s -march=x86 -mcpu=yonah -o - | grep mulsd | count 6
+; Ideally this would compile to 5 multiplies.
+
+define double @_Z3f10d(double %a) nounwind readonly ssp noredzone {
+entry:
+  %0 = tail call double @llvm.powi.f64(double %a, i32 15) nounwind ; <double> [#uses=1]
+  ret double %0
+}
+
+declare double @llvm.powi.f64(double, i32) nounwind readonly
+