The signed version of our "magic number" computation for the integer approximation

of a constant had a minor typo introduced when copying it from the book, which
caused it to favor negative approximations over positive approximations in many
cases. Positive approximations require fewer operations beyond the multiplication.

In the case of division by 3, we still generate code that is a single instruction
larger than GCC's code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@126097 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Support/APInt.cpp b/lib/Support/APInt.cpp
index 7703342..08f36d2 100644 (file)
--- a/lib/Support/APInt.cpp
+++ b/lib/Support/APInt.cpp
@@ -1505,7 +1505,7 @@ APInt::ms APInt::magic() const {
       r2 = r2 - ad;
     }
     delta = ad - r2;
-  } while (q1.ule(delta) || (q1 == delta && r1 == 0));
+  } while (q1.ult(delta) || (q1 == delta && r1 == 0));
 
   mag.m = q2 + 1;
   if (d.isNegative()) mag.m = -mag.m;   // resulting magic number

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index c10e170..ed3bff1 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -1888,12 +1888,10 @@ Compiles to: $clang t.c -S -o - -O3 -mkernel -fomit-frame-pointer
 
 _t:                                     ## @t
        movslq  %edi, %rax
-       imulq   $-1431655765, %rax, %rcx ## imm = 0xFFFFFFFFAAAAAAAB
-       shrq    $32, %rcx
-       addl    %ecx, %eax
-       movl    %eax, %ecx
-       shrl    $31, %ecx
-       shrl    %eax
+       imulq   $1431655766, %rax, %rax ## imm = 0x55555556
+       movq    %rax, %rcx
+       shrq    $63, %rcx
+       shrq    $32, %rax
        addl    %ecx, %eax
        movsbl  %al, %eax
        ret

diff --git a/test/CodeGen/X86/divide-by-constant.ll b/test/CodeGen/X86/divide-by-constant.ll
index 7ceb972..fe335b9 100644 (file)
--- a/test/CodeGen/X86/divide-by-constant.ll
+++ b/test/CodeGen/X86/divide-by-constant.ll
@@ -40,7 +40,7 @@ entry:
        %div = sdiv i16 %x, 33          ; <i32> [#uses=1]
        ret i16 %div
 ; CHECK: test4:
-; CHECK: imull $-1985, %ecx, %ecx 
+; CHECK: imull $1986, %eax, %eax 
 }
 
 define i32 @test5(i32 %A) nounwind {

diff --git a/unittests/ADT/APIntTest.cpp b/unittests/ADT/APIntTest.cpp
index 557d835..e05bdbf 100644 (file)
--- a/unittests/ADT/APIntTest.cpp
+++ b/unittests/ADT/APIntTest.cpp
@@ -332,6 +332,24 @@ TEST(APIntTest, Log2) {
   EXPECT_EQ(APInt(15, 9).exactLogBase2(), -1);
 }
 
+TEST(APIntTest, magic) {
+  EXPECT_EQ(APInt(32, 3).magic().m, APInt(32, "55555556", 16));
+  EXPECT_EQ(APInt(32, 3).magic().s, 0U);
+  EXPECT_EQ(APInt(32, 5).magic().m, APInt(32, "66666667", 16));
+  EXPECT_EQ(APInt(32, 5).magic().s, 1U);
+  EXPECT_EQ(APInt(32, 7).magic().m, APInt(32, "92492493", 16));
+  EXPECT_EQ(APInt(32, 7).magic().s, 2U);
+}
+
+TEST(APIntTest, magicu) {
+  EXPECT_EQ(APInt(32, 3).magicu().m, APInt(32, "AAAAAAAB", 16));
+  EXPECT_EQ(APInt(32, 3).magicu().s, 1U);
+  EXPECT_EQ(APInt(32, 5).magicu().m, APInt(32, "CCCCCCCD", 16));
+  EXPECT_EQ(APInt(32, 5).magicu().s, 2U);
+  EXPECT_EQ(APInt(32, 7).magicu().m, APInt(32, "24924925", 16));
+  EXPECT_EQ(APInt(32, 7).magicu().s, 3U);
+}
+
 #ifdef GTEST_HAS_DEATH_TEST
 #ifndef NDEBUG
 TEST(APIntTest, StringDeath) {