+
+We don't fold (icmp (add) (add)) unless the two adds only have a single use.
+There are a lot of cases that we're refusing to fold in (e.g.) 256.bzip2, for
+example:
+
+ %indvar.next90 = add i64 %indvar89, 1 ;; Has 2 uses
+ %tmp96 = add i64 %tmp95, 1 ;; Has 1 use
+ %exitcond97 = icmp eq i64 %indvar.next90, %tmp96
+
+We don't fold this because we don't want to introduce an overlapped live range
+of the ivar. However if we can make this more aggressive without causing
+performance issues in two ways:
+
+1. If *either* the LHS or RHS has a single use, we can definitely do the
+ transformation. In the overlapping liverange case we're trading one register
+ use for one fewer operation, which is a reasonable trade. Before doing this
+ we should verify that the llc output actually shrinks for some benchmarks.
+2. If both ops have multiple uses, we can still fold it if the operations are
+ both sinkable to *after* the icmp (e.g. in a subsequent block) which doesn't
+ increase register pressure.
+
+There are a ton of icmp's we aren't simplifying because of the reg pressure
+concern. Care is warranted here though because many of these are induction
+variables and other cases that matter a lot to performance, like the above.
+Here's a blob of code that you can drop into the bottom of visitICmp to see some
+missed cases:
+
+ { Value *A, *B, *C, *D;
+ if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
+ match(Op1, m_Add(m_Value(C), m_Value(D))) &&
+ (A == C || A == D || B == C || B == D)) {
+ errs() << "OP0 = " << *Op0 << " U=" << Op0->getNumUses() << "\n";
+ errs() << "OP1 = " << *Op1 << " U=" << Op1->getNumUses() << "\n";
+ errs() << "CMP = " << I << "\n\n";
+ }
+ }
+
+//===---------------------------------------------------------------------===//
+
+define i1 @test1(i32 %x) nounwind {
+ %and = and i32 %x, 3
+ %cmp = icmp ult i32 %and, 2
+ ret i1 %cmp
+}
+
+Can be folded to (x & 2) == 0.
+
+define i1 @test2(i32 %x) nounwind {
+ %and = and i32 %x, 3
+ %cmp = icmp ugt i32 %and, 1
+ ret i1 %cmp
+}
+
+Can be folded to (x & 2) != 0.
+
+SimplifyDemandedBits shrinks the "and" constant to 2 but instcombine misses the
+icmp transform.
+
+//===---------------------------------------------------------------------===//
+
+This code:
+
+typedef struct {
+int f1:1;
+int f2:1;
+int f3:1;
+int f4:29;
+} t1;
+
+typedef struct {
+int f1:1;
+int f2:1;
+int f3:30;
+} t2;
+
+t1 s1;
+t2 s2;
+
+void func1(void)
+{
+s1.f1 = s2.f1;
+s1.f2 = s2.f2;
+}
+
+Compiles into this IR (on x86-64 at least):
+
+%struct.t1 = type { i8, [3 x i8] }
+@s2 = global %struct.t1 zeroinitializer, align 4
+@s1 = global %struct.t1 zeroinitializer, align 4
+define void @func1() nounwind ssp noredzone {
+entry:
+ %0 = load i32* bitcast (%struct.t1* @s2 to i32*), align 4
+ %bf.val.sext5 = and i32 %0, 1
+ %1 = load i32* bitcast (%struct.t1* @s1 to i32*), align 4
+ %2 = and i32 %1, -4
+ %3 = or i32 %2, %bf.val.sext5
+ %bf.val.sext26 = and i32 %0, 2
+ %4 = or i32 %3, %bf.val.sext26
+ store i32 %4, i32* bitcast (%struct.t1* @s1 to i32*), align 4
+ ret void
+}
+
+The two or/and's should be merged into one each.
+
+//===---------------------------------------------------------------------===//
+
+Machine level code hoisting can be useful in some cases. For example, PR9408
+is about:
+
+typedef union {
+ void (*f1)(int);
+ void (*f2)(long);
+} funcs;
+
+void foo(funcs f, int which) {
+ int a = 5;
+ if (which) {
+ f.f1(a);
+ } else {
+ f.f2(a);
+ }
+}
+
+which we compile to:
+
+foo: # @foo
+# BB#0: # %entry
+ pushq %rbp
+ movq %rsp, %rbp
+ testl %esi, %esi
+ movq %rdi, %rax
+ je .LBB0_2
+# BB#1: # %if.then
+ movl $5, %edi
+ callq *%rax
+ popq %rbp
+ ret
+.LBB0_2: # %if.else
+ movl $5, %edi
+ callq *%rax
+ popq %rbp
+ ret
+
+Note that bb1 and bb2 are the same. This doesn't happen at the IR level
+because one call is passing an i32 and the other is passing an i64.
+
+//===---------------------------------------------------------------------===//
+
+I see this sort of pattern in 176.gcc in a few places (e.g. the start of
+store_bit_field). The rem should be replaced with a multiply and subtract:
+
+ %3 = sdiv i32 %A, %B
+ %4 = srem i32 %A, %B
+
+Similarly for udiv/urem. Note that this shouldn't be done on X86 or ARM,
+which can do this in a single operation (instruction or libcall). It is
+probably best to do this in the code generator.
+
+//===---------------------------------------------------------------------===//
+
+unsigned foo(unsigned x, unsigned y) { return (x & y) == 0 || x == 0; }
+should fold to (x & y) == 0.
+
+//===---------------------------------------------------------------------===//
+
+unsigned foo(unsigned x, unsigned y) { return x > y && x != 0; }
+should fold to x > y.
+
+//===---------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff