RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2025-01-31 12:41:49 +01:00
Code Issues Projects Releases Wiki Activity

[InstCombine] use m_APInt to allow icmp (trunc X, Y), C folds for splat constant vectors

Browse Source 
This is a sibling of:
https://reviews.llvm.org/rL278859
https://reviews.llvm.org/rL278935
https://reviews.llvm.org/rL278945
https://reviews.llvm.org/rL279066
https://reviews.llvm.org/rL279077
https://reviews.llvm.org/rL279101

llvm-svn: 279133
This commit is contained in:
Sanjay Patel 2016-08-18 20:28:54 +00:00
parent da28ab95fe
commit f315654cc5
3 changed files with 8 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
lib/Transforms/InstCombine/InstCombineCompares.cpp
View File

@ -1538,14 +1538,9 @@ Instruction *InstCombiner::foldICmpCstShlConst(ICmpInst &I, Value *Op, Value *A,
Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
Instruction *Trunc,
const APInt *C) {
// FIXME: This check restricts all folds under here to scalar types.
ConstantInt *RHS = dyn_cast<ConstantInt>(Cmp.getOperand(1));
if (!RHS)
return nullptr;
ICmpInst::Predicate Pred = Cmp.getPredicate();
Value *X = Trunc->getOperand(0);
if (RHS->isOne() && C->getBitWidth() > 1) {
if (*C == 1 && C->getBitWidth() > 1) {
// icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
Value *V = nullptr;
if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
@ -1556,8 +1551,8 @@ Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
if (Cmp.isEquality() && Trunc->hasOneUse()) {
// Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
// of the high bits truncated out of x are known.
unsigned DstBits = Trunc->getType()->getPrimitiveSizeInBits(),
SrcBits = X->getType()->getPrimitiveSizeInBits();
unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
SrcBits = X->getType()->getScalarSizeInBits();
APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
computeKnownBits(X, KnownZero, KnownOne, 0, &Cmp);
@ -1566,7 +1561,7 @@ Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
// Pull in the high bits from known-ones set.
APInt NewRHS = C->zext(SrcBits);
NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
return new ICmpInst(Pred, X, Builder->getInt(NewRHS));
return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
}
}

7
test/Transforms/InstCombine/cast.ll
View File

@ -356,7 +356,7 @@ define i1 @test31(i64 %A) {
ret i1 %D
}
; FIXME: Vectors should fold too...or not?
; FIXME: Vectors should fold too...or not?
; Does this depend on the whether the source/dest types of the trunc are legal in the data layout?
define <2 x i1> @test31vec(<2 x i64> %A) {
; CHECK-LABEL: @test31vec(
@ -414,12 +414,11 @@ define i1 @test36(i32 %a) {
ret i1 %d
}
; FIXME: Vectors should fold too.
; FIXME: The trunc is removed, but the icmp+lshr fold is missing.
define <2 x i1> @test36vec(<2 x i32> %a) {
; CHECK-LABEL: @test36vec(
; CHECK-NEXT: [[B:%.*]] = lshr <2 x i32> %a, <i32 31, i32 31>
; CHECK-NEXT: [[C:%.*]] = trunc <2 x i32> [[B]] to <2 x i8>
; CHECK-NEXT: [[D:%.*]] = icmp eq <2 x i8> [[C]], zeroinitializer
; CHECK-NEXT: [[D:%.*]] = icmp eq <2 x i32> [[B]], zeroinitializer
; CHECK-NEXT: ret <2 x i1> [[D]]
;
%b = lshr <2 x i32> %a, <i32 31, i32 31>

8
test/Transforms/InstCombine/compare-signs.ll
View File

@ -107,15 +107,9 @@ define i1 @test4c(i64 %a) {
ret i1 %c
}
; FIXME: Vectors should fold too.
define <2 x i1> @test4c_vec(<2 x i64> %a) {
; CHECK-LABEL: @test4c_vec(
; CHECK-NEXT: [[L:%.*]] = ashr <2 x i64> %a, <i64 63, i64 63>
; CHECK-NEXT: [[NA:%.*]] = sub <2 x i64> zeroinitializer, %a
; CHECK-NEXT: [[R:%.*]] = lshr <2 x i64> [[NA]], <i64 63, i64 63>
; CHECK-NEXT: [[SIGNUM:%.*]] = or <2 x i64> [[L]], [[R]]
; CHECK-NEXT: [[SIGNUM_TRUNC:%.*]] = trunc <2 x i64> [[SIGNUM]] to <2 x i32>
; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i32> [[SIGNUM_TRUNC]], <i32 1, i32 1>
; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i64> %a, <i64 1, i64 1>
; CHECK-NEXT: ret <2 x i1> [[C]]
;
%l = ashr <2 x i64> %a, <i64 63, i64 63>