This requires reassociating to forms of expressions that are already available,
something that reassoc doesn't think about yet.
+
+//===---------------------------------------------------------------------===//
+
+This function: (derived from GCC PR19988)
+double foo(double x, double y) {
+ return ((x + 0.1234 * y) * (x + -0.1234 * y));
+}
+
+compiles to:
+_foo:
+ movapd %xmm1, %xmm2
+ mulsd LCPI1_1(%rip), %xmm1
+ mulsd LCPI1_0(%rip), %xmm2
+ addsd %xmm0, %xmm1
+ addsd %xmm0, %xmm2
+ movapd %xmm1, %xmm0
+ mulsd %xmm2, %xmm0
+ ret
+
+Reassociate should be able to turn it into:
+
+double foo(double x, double y) {
+ return ((x + 0.1234 * y) * (x - 0.1234 * y));
+}
+
+Which allows the multiply by constant to be CSE'd, producing:
+
+_foo:
+ mulsd LCPI1_0(%rip), %xmm1
+ movapd %xmm1, %xmm2
+ addsd %xmm0, %xmm2
+ subsd %xmm1, %xmm0
+ mulsd %xmm2, %xmm0
+ ret
+
+This doesn't need -ffast-math support at all. This is particularly bad because
+the llvm-gcc frontend is canonicalizing the later into the former, but clang
+doesn't have this problem.
+
//===---------------------------------------------------------------------===//
These two functions should generate the same code on big-endian systems:
//===---------------------------------------------------------------------===//
-Turn this into a single byte store with no load (the other 3 bytes are
-unmodified):
-
-define void @test(i32* %P) {
- %tmp = load i32* %P
- %tmp14 = or i32 %tmp, 3305111552
- %tmp15 = and i32 %tmp14, 3321888767
- store i32 %tmp15, i32* %P
- ret void
-}
-
-//===---------------------------------------------------------------------===//
-
quantum_sigma_x in 462.libquantum contains the following loop:
for(i=0; i<reg->size; i++)
Missed instcombine transformation:
define i32 @a(i32 %x) nounwind readnone {
entry:
- %shr = lshr i32 %x, 5 ; <i32> [#uses=1]
- %xor = xor i32 %shr, 67108864 ; <i32> [#uses=1]
- %sub = add i32 %xor, -67108864 ; <i32> [#uses=1]
+ %rem = srem i32 %x, 32
+ %shl = shl i32 1, %rem
+ ret i32 %shl
+}
+
+The srem can be transformed to an and because if x is negative, the shift is
+undefined. Testcase derived from gcc.
+
+//===---------------------------------------------------------------------===//
+
+Missed instcombine/dagcombine transformation:
+define i32 @a(i32 %x, i32 %y) nounwind readnone {
+entry:
+ %mul = mul i32 %y, -8
+ %sub = sub i32 %x, %mul
ret i32 %sub
}
-This function is equivalent to "ashr i32 %x, 5". Testcase derived from gcc.
+Should compile to something like x+y*8, but currently compiles to an
+inefficient result. Testcase derived from gcc.
+
+//===---------------------------------------------------------------------===//
+
+Missed instcombine/dagcombine transformation:
+define void @lshift_lt(i8 zeroext %a) nounwind {
+entry:
+ %conv = zext i8 %a to i32
+ %shl = shl i32 %conv, 3
+ %cmp = icmp ult i32 %shl, 33
+ br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+ tail call void @bar() nounwind
+ ret void
+
+if.end:
+ ret void
+}
+declare void @bar() nounwind
+
+The shift should be eliminated. Testcase derived from gcc.
+
+//===---------------------------------------------------------------------===//
+
+These compile into different code, one gets recognized as a switch and the
+other doesn't due to phase ordering issues (PR6212):
+
+int test1(int mainType, int subType) {
+ if (mainType == 7)
+ subType = 4;
+ else if (mainType == 9)
+ subType = 6;
+ else if (mainType == 11)
+ subType = 9;
+ return subType;
+}
+
+int test2(int mainType, int subType) {
+ if (mainType == 7)
+ subType = 4;
+ if (mainType == 9)
+ subType = 6;
+ if (mainType == 11)
+ subType = 9;
+ return subType;
+}
//===---------------------------------------------------------------------===//
-isSafeToLoadUnconditionally should allow a GEP of a global/alloca with constant
-indicies within the bounds of the allocated object. Reduced example:
+The following test case (from PR6576):
+
+define i32 @mul(i32 %a, i32 %b) nounwind readnone {
+entry:
+ %cond1 = icmp eq i32 %b, 0 ; <i1> [#uses=1]
+ br i1 %cond1, label %exit, label %bb.nph
+bb.nph: ; preds = %entry
+ %tmp = mul i32 %b, %a ; <i32> [#uses=1]
+ ret i32 %tmp
+exit: ; preds = %entry
+ ret i32 0
+}
+
+could be reduced to:
+
+define i32 @mul(i32 %a, i32 %b) nounwind readnone {
+entry:
+ %tmp = mul i32 %b, %a
+ ret i32 %tmp
+}
+
+//===---------------------------------------------------------------------===//
+
+We should use DSE + llvm.lifetime.end to delete dead vtable pointer updates.
+See GCC PR34949
+
+Another interesting case is that something related could be used for variables
+that go const after their ctor has finished. In these cases, globalopt (which
+can statically run the constructor) could mark the global const (so it gets put
+in the readonly section). A testcase would be:
+
+#include <complex>
+using namespace std;
+const complex<char> should_be_in_rodata (42,-42);
+complex<char> should_be_in_data (42,-42);
+complex<char> should_be_in_bss;
+
+Where we currently evaluate the ctors but the globals don't become const because
+the optimizer doesn't know they "become const" after the ctor is done. See
+GCC PR4131 for more examples.
+
+//===---------------------------------------------------------------------===//
+
+In this code:
+
+long foo(long x) {
+ return x > 1 ? x : 1;
+}
-const int a[] = {3,6};
-int b(int y) { int* x = y ? &a[0] : &a[1]; return *x; }
+LLVM emits a comparison with 1 instead of 0. 0 would be equivalent
+and cheaper on most targets.
-All the loads should be eliminated. Testcase derived from gcc.
+LLVM prefers comparisons with zero over non-zero in general, but in this
+case it choses instead to keep the max operation obvious.
//===---------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff