summaryrefslogtreecommitdiff
path: root/lib/Target/README.txt
diff options
context:
space:
mode:
authorChris Lattner <sabre@nondot.org>2011-04-14 04:21:42 +0000
committerChris Lattner <sabre@nondot.org>2011-04-14 04:21:42 +0000
commitdf2ef9015f8a02eee25c45a9be648540641610c2 (patch)
treeed2b9c3afffbd2d5fa4c45d351b662cd58b1d6b4 /lib/Target/README.txt
parentc0c7fca2fedd2259569b7b84338259c62d38802d (diff)
add a minor missed dag combine that is blocking mid-level optimization
improvements, that will lead to fixing PR6627. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@129504 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/README.txt')
-rw-r--r--lib/Target/README.txt76
1 files changed, 76 insertions, 0 deletions
diff --git a/lib/Target/README.txt b/lib/Target/README.txt
index 99616b47c84..5032b2bbbb5 100644
--- a/lib/Target/README.txt
+++ b/lib/Target/README.txt
@@ -2258,3 +2258,79 @@ SimplifyDemandedBits shrinks the "and" constant to 2 but instcombine misses the
icmp transform.
//===---------------------------------------------------------------------===//
+
+These functions:
+int foo(int *X) {
+ if ((*X & 255) == 47)
+ bar();
+}
+int foo2(int X) {
+ if ((X & 255) == 47)
+ bar();
+}
+
+codegen to:
+
+ movzbl (%rdi), %eax
+ cmpl $47, %eax
+ jne LBB0_2
+
+and:
+ movzbl %dil, %eax
+ cmpl $47, %eax
+ jne LBB1_2
+
+If a dag combine shrunk the compare to a byte compare, then we'd fold the load
+in the first example, and eliminate the movzbl in the second, saving a register.
+This can be a target independent dag combine that works on ISD::SETCC, it would
+catch this before the legalize ops pass.
+
+//===---------------------------------------------------------------------===//
+
+We should optimize this:
+
+ %tmp = load i16* %arrayidx, align 4, !tbaa !0
+ %A = trunc i16 %tmp to i8
+ %cmp = icmp eq i8 %A, 127
+ %B.mask = and i16 %tmp, -256
+ %cmp7 = icmp eq i16 %B.mask, 17664
+ %or.cond = and i1 %cmp, %cmp7
+ br i1 %or.cond, label %land.lhs.true9, label %if.end
+
+into:
+
+ %tmp = load i16* %arrayidx, align 4, !tbaa !0
+ %0 = icmp eq i16 %tmp, 17791
+ br i1 %0, label %land.lhs.true9, label %if.end
+
+with this patch:
+Index: InstCombine/InstCombineCompares.cpp
+===================================================================
+--- InstCombine/InstCombineCompares.cpp (revision 129500)
++++ InstCombine/InstCombineCompares.cpp (working copy)
+@@ -2506,6 +2506,18 @@
+ return &I;
+ }
+ }
++
++ // Transform "icmp eq (trunc X), cst" to "icmp (and X, mask), cst"
++ if (Op0->hasOneUse() && match(Op0, m_Trunc(m_Value(A))) &&
++ isa<ConstantInt>(Op1)) {
++ APInt MaskV = APInt::getLowBitsSet(A->getType()->getPrimitiveSizeInBits(),
++ Op0->getType()->getPrimitiveSizeInBits());
++ Value *Mask =
++ Builder->CreateAnd(A, ConstantInt::get(A->getContext(), MaskV));
++ return new ICmpInst(I.getPredicate(), Mask,
++ ConstantExpr::getZExt(cast<ConstantInt>(Op1),
++ Mask->getType()));
++ }
+ }
+
+ {
+
+
+but we can't do that until the dag combine above is added. Not having this
+is blocking resolving PR6627.
+
+//===---------------------------------------------------------------------===//
+