//===---------------------------------------------------------------------===//
+[LOOP DELETION]
+
+We don't delete this output free loop, because trip count analysis doesn't
+realize that it is finite (if it were infinite, it would be undefined). Not
+having this blocks Loop Idiom from matching strlen and friends.
+
+void foo(char *C) {
+ int x = 0;
+ while (*C)
+ ++x,++C;
+}
+
+//===---------------------------------------------------------------------===//
+
[LOOP RECOGNITION]
These idioms should be recognized as popcount (see PR1488):
return count;
}
+This should be recognized as CLZ: rdar://8459039
+
+unsigned clz_a(unsigned a) {
+ int i;
+ for (i=0;i<32;i++)
+ if (a & (1<<(31-i)))
+ return i;
+ return 32;
+}
+
This sort of thing should be added to the loop idiom pass.
//===---------------------------------------------------------------------===//
if ((a >> 2) > 5)
bar ();
}
+
+void neg_eq_cst(unsigned int a) {
+if (-a == 123)
+bar();
+}
+
All should simplify to a single comparison. All of these are
currently not optimized with "clang -emit-llvm-bc | opt
-std-compile-opts".
//===---------------------------------------------------------------------===//
+simplifylibcalls should turn these snprintf idioms into memcpy (GCC PR47917)
+
+char buf1[6], buf2[6], buf3[4], buf4[4];
+int i;
+
+int foo (void) {
+ int ret = snprintf (buf1, sizeof buf1, "abcde");
+ ret += snprintf (buf2, sizeof buf2, "abcdef") * 16;
+ ret += snprintf (buf3, sizeof buf3, "%s", i++ < 6 ? "abc" : "def") * 256;
+ ret += snprintf (buf4, sizeof buf4, "%s", i++ > 10 ? "abcde" : "defgh")*4096;
+ return ret;
+}
+
+//===---------------------------------------------------------------------===//
+
"gas" uses this idiom:
else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
..
//===---------------------------------------------------------------------===//
-Take the following testcase on x86-64 (similar testcases exist for all targets
-with addc/adde):
-
-define void @a(i64* nocapture %s, i64* nocapture %t, i64 %a, i64 %b,
-i64 %c) nounwind {
-entry:
- %0 = zext i64 %a to i128 ; <i128> [#uses=1]
- %1 = zext i64 %b to i128 ; <i128> [#uses=1]
- %2 = add i128 %1, %0 ; <i128> [#uses=2]
- %3 = zext i64 %c to i128 ; <i128> [#uses=1]
- %4 = shl i128 %3, 64 ; <i128> [#uses=1]
- %5 = add i128 %4, %2 ; <i128> [#uses=1]
- %6 = lshr i128 %5, 64 ; <i128> [#uses=1]
- %7 = trunc i128 %6 to i64 ; <i64> [#uses=1]
- store i64 %7, i64* %s, align 8
- %8 = trunc i128 %2 to i64 ; <i64> [#uses=1]
- store i64 %8, i64* %t, align 8
- ret void
-}
-
-Generated code:
- addq %rcx, %rdx
- sbbq %rax, %rax
- subq %rax, %r8
- movq %r8, (%rdi)
- movq %rdx, (%rsi)
- ret
-
-Expected code:
- addq %rcx, %rdx
- adcq $0, %r8
- movq %r8, (%rdi)
- movq %rdx, (%rsi)
- ret
-
-//===---------------------------------------------------------------------===//
-
Switch lowering generates less than ideal code for the following switch:
define void @a(i32 %x) nounwind {
entry:
//===---------------------------------------------------------------------===//
-Some missed instcombine xforms (from GCC PR14753):
-
-void bar (void);
-
-void mask_gt (unsigned int a) {
-/* This is equivalent to a > 15. */
-if ((a & ~7) > 8)
-bar();
-}
-
-void neg_eq_cst(unsigned int a) {
-if (-a == 123)
-bar();
-}
-
-void minus_cst(unsigned int a) {
-if (20 - a == 5)
-bar();
-}
-
-void rotate_cst (unsigned a) {
-a = (a << 10) | (a >> 22);
-if (a == 123)
-bar ();
-}
+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";
+ }
+ }
//===---------------------------------------------------------------------===//
+
projects / oota-llvm.git / blobdiff