InstCombine: Turn (zext A) == (B & (1<<X)-1) into A == (trunc B), narrowing the compare.

This saves a cast, and zext is more expensive on platforms with subreg support
than trunc is. This occurs in the BSD implementation of memchr(3), see PR12750.
On the synthetic benchmark from that bug stupid_memchr and bsd_memchr have the
same performance now when not inlining either function.

stupid_memchr: 323.0us
bsd_memchr: 321.0us
memchr: 479.0us

where memchr is the llvm-gcc compiled bsd_memchr from osx lion's libc. When
inlining is enabled bsd_memchr still regresses down to llvm-gcc memchr time,
I haven't fully understood the issue yet, something is grossly mangling the
loop after inlining.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@158297 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCompares.cpp b/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 7bbdd450f111bf2fe33c3545f62e7117b61f3a08..3fa0aba012945162701119c9bceca488aed1c0a1 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -2580,10 +2580,32 @@ Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
       }
     }
 
+    // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
+    ConstantInt *Cst1;
+    if (Op0->hasOneUse() &&
+        match(Op0, m_ZExt(m_Value(A))) &&
+        match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) {
+      APInt Pow2 = Cst1->getValue() + 1;
+      if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
+          Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
+        return new ICmpInst(I.getPredicate(), A,
+                            Builder->CreateTrunc(B, A->getType()));
+    }
+
+    // Transform (B & (1<<X)-1) == (zext A) --> A == (trunc B)
+    if (Op1->hasOneUse() &&
+        match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
+        match(Op1, m_ZExt(m_Value(A)))) {
+      APInt Pow2 = Cst1->getValue() + 1;
+      if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
+          Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
+        return new ICmpInst(I.getPredicate(), A,
+                            Builder->CreateTrunc(B, A->getType()));
+    }
+
     // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
     // "icmp (and X, mask), cst"
     uint64_t ShAmt = 0;
-    ConstantInt *Cst1;
     if (Op0->hasOneUse() &&
         match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
                                            m_ConstantInt(ShAmt))))) &&

diff --git a/test/Transforms/InstCombine/icmp.ll b/test/Transforms/InstCombine/icmp.ll
index a9ae221d8f96473d27279b9d0e0d856e7a49695a..eaff87d695ed96e6e828c0f5a8b6981061c76aa3 100644 (file)
--- a/test/Transforms/InstCombine/icmp.ll
+++ b/test/Transforms/InstCombine/icmp.ll
@@ -637,3 +637,25 @@ define i1 @test62(i8* %a) {
 ; CHECK: @test62
 ; CHECK-NEXT: ret i1 true
 }
+
+define i1 @test63(i8 %a, i32 %b) nounwind {
+  %z = zext i8 %a to i32
+  %t = and i32 %b, 255
+  %c = icmp eq i32 %z, %t
+  ret i1 %c
+; CHECK: @test63
+; CHECK-NEXT: %1 = trunc i32 %b to i8
+; CHECK-NEXT: %c = icmp eq i8 %1, %a
+; CHECK-NEXT: ret i1 %c
+}
+
+define i1 @test64(i8 %a, i32 %b) nounwind {
+  %t = and i32 %b, 255
+  %z = zext i8 %a to i32
+  %c = icmp eq i32 %t, %z
+  ret i1 %c
+; CHECK: @test64
+; CHECK-NEXT: %1 = trunc i32 %b to i8
+; CHECK-NEXT: %c = icmp eq i8 %1, %a
+; CHECK-NEXT: ret i1 %c
+}