When forming SSE min and max nodes for UGE and ULE comparisons, it's
necessary to swap the operands to handle NaN and negative zero properly.
Also, reintroduce logic for checking for NaN conditions when forming
SSE min and max instructions, fixed to take into consideration NaNs and
negative zeros. This allows forming min and max instructions in more
cases.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@97025 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/llvm/CodeGen/SelectionDAG.h b/include/llvm/CodeGen/SelectionDAG.h
index e628603..09651e6 100644
--- a/include/llvm/CodeGen/SelectionDAG.h
+++ b/include/llvm/CodeGen/SelectionDAG.h
@@ -898,6 +898,15 @@
/// isKnownNeverNan - Test whether the given SDValue is known to never be NaN.
bool isKnownNeverNaN(SDValue Op) const;
+ /// isKnownNeverZero - Test whether the given SDValue is known to never be
+ /// positive or negative Zero.
+ bool isKnownNeverZero(SDValue Op) const;
+
+ /// isEqualTo - Test whether two SDValues are known to compare equal. This
+ /// is true if they are the same value, or if one is negative zero and the
+ /// other positive zero.
+ bool isEqualTo(SDValue A, SDValue B) const;
+
/// isVerifiedDebugInfoDesc - Returns true if the specified SDValue has
/// been verified as a debug information descriptor.
bool isVerifiedDebugInfoDesc(SDValue Op) const;
diff --git a/include/llvm/CodeGen/SelectionDAGNodes.h b/include/llvm/CodeGen/SelectionDAGNodes.h
index 9066913..950963e 100644
--- a/include/llvm/CodeGen/SelectionDAGNodes.h
+++ b/include/llvm/CodeGen/SelectionDAGNodes.h
@@ -1815,6 +1815,12 @@
const APFloat& getValueAPF() const { return Value->getValueAPF(); }
const ConstantFP *getConstantFPValue() const { return Value; }
+ /// isZero - Return true if the value is positive or negative zero.
+ bool isZero() const { return Value->isZero(); }
+
+ /// isNaN - Return true if the value is a NaN.
+ bool isNaN() const { return Value->isNaN(); }
+
/// isExactlyValue - We don't rely on operator== working on double values, as
/// it returns true for things that are clearly not equal, like -0.0 and 0.0.
/// As such, this method can be used to do an exact bit-for-bit comparison of
diff --git a/include/llvm/Constants.h b/include/llvm/Constants.h
index bd14303..c706329 100644
--- a/include/llvm/Constants.h
+++ b/include/llvm/Constants.h
@@ -276,6 +276,12 @@
return Val.isZero() && Val.isNegative();
}
+ /// isZero - Return true if the value is positive or negative zero.
+ bool isZero() const { return Val.isZero(); }
+
+ /// isNaN - Return true if the value is a NaN.
+ bool isNaN() const { return Val.isNaN(); }
+
/// isExactlyValue - We don't rely on operator== working on double values, as
/// it returns true for things that are clearly not equal, like -0.0 and 0.0.
/// As such, this method can be used to do an exact bit-for-bit comparison of
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index f8930b8..4105021 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2235,6 +2235,29 @@
return false;
}
+bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
+ // If the value is a constant, we can obviously see if it is a zero or not.
+ if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
+ return !C->isZero();
+
+ // TODO: Recognize more cases here.
+
+ return false;
+}
+
+bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
+ // Check the obvious case.
+ if (A == B) return true;
+
+ // For for negative and positive zero.
+ if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A))
+ if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B))
+ if (CA->isZero() && CB->isZero()) return true;
+
+ // Otherwise they may not be equal.
+ return false;
+}
+
bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const {
GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
if (!GA) return false;
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 3bbe5fb..d2d76da 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -8833,18 +8833,35 @@
unsigned Opcode = 0;
// Check for x CC y ? x : y.
- if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
+ if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(1))) {
switch (CC) {
default: break;
case ISD::SETULT:
- if (!UnsafeFPMath) break;
+ // Converting this to a min would handle NaNs incorrectly, and swapping
+ // the operands would cause it to handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
+ break;
+ std::swap(LHS, RHS);
+ }
Opcode = X86ISD::FMIN;
break;
case ISD::SETOLE:
- if (!UnsafeFPMath) break;
+ // Converting this to a min would handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
+ break;
Opcode = X86ISD::FMIN;
break;
case ISD::SETULE:
+ // Converting this to a min would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
+ std::swap(LHS, RHS);
case ISD::SETOLT:
case ISD::SETLT:
case ISD::SETLE:
@@ -8852,14 +8869,30 @@
break;
case ISD::SETOGE:
- if (!UnsafeFPMath) break;
+ // Converting this to a max would handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
+ break;
Opcode = X86ISD::FMAX;
break;
case ISD::SETUGT:
- if (!UnsafeFPMath) break;
+ // Converting this to a max would handle NaNs incorrectly, and swapping
+ // the operands would cause it to handle comparisons between positive
+ // and negative zero incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
+ break;
+ std::swap(LHS, RHS);
+ }
Opcode = X86ISD::FMAX;
break;
case ISD::SETUGE:
+ // Converting this to a max would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
+ std::swap(LHS, RHS);
case ISD::SETOGT:
case ISD::SETGT:
case ISD::SETGE:
@@ -8867,18 +8900,34 @@
break;
}
// Check for x CC y ? y : x -- a min/max with reversed arms.
- } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
+ } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(0))) {
switch (CC) {
default: break;
case ISD::SETOGE:
- if (!UnsafeFPMath) break;
+ // Converting this to a min would handle comparisons between positive
+ // and negative zero incorrectly, and swapping the operands would
+ // cause it to handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
+ std::swap(LHS, RHS);
+ }
Opcode = X86ISD::FMIN;
break;
case ISD::SETUGT:
- if (!UnsafeFPMath) break;
+ // Converting this to a min would handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
Opcode = X86ISD::FMIN;
break;
case ISD::SETUGE:
+ // Converting this to a min would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
+ std::swap(LHS, RHS);
case ISD::SETOGT:
case ISD::SETGT:
case ISD::SETGE:
@@ -8886,14 +8935,29 @@
break;
case ISD::SETULT:
- if (!UnsafeFPMath) break;
+ // Converting this to a max would handle NaNs incorrectly.
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
Opcode = X86ISD::FMAX;
break;
case ISD::SETOLE:
- if (!UnsafeFPMath) break;
+ // Converting this to a max would handle comparisons between positive
+ // and negative zero incorrectly, and swapping the operands would
+ // cause it to handle NaNs incorrectly.
+ if (!UnsafeFPMath &&
+ !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
+ if (!FiniteOnlyFPMath() &&
+ (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
+ break;
+ std::swap(LHS, RHS);
+ }
Opcode = X86ISD::FMAX;
break;
case ISD::SETULE:
+ // Converting this to a max would handle both negative zeros and NaNs
+ // incorrectly, but we can swap the operands to fix both.
+ std::swap(LHS, RHS);
case ISD::SETOLT:
case ISD::SETLT:
case ISD::SETLE:
diff --git a/test/CodeGen/X86/sse-minmax.ll b/test/CodeGen/X86/sse-minmax.ll
index f8590fd..19fbed0 100644
--- a/test/CodeGen/X86/sse-minmax.ll
+++ b/test/CodeGen/X86/sse-minmax.ll
@@ -1,13 +1,15 @@
; RUN: llc < %s -march=x86-64 -asm-verbose=false | FileCheck %s
; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-unsafe-fp-math | FileCheck -check-prefix=UNSAFE %s
+; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-finite-only-fp-math | FileCheck -check-prefix=FINITE %s
; Some of these patterns can be matched as SSE min or max. Some of
; then can be matched provided that the operands are swapped.
; Some of them can't be matched at all and require a comparison
; and a conditional branch.
-; The naming convention is {,x_}{o,u}{gt,lt,ge,le}{,_inverse}
+; The naming convention is {,x_,y_}{o,u}{gt,lt,ge,le}{,_inverse}
; x_ : use 0.0 instead of %y
+; y_ : use -0.0 instead of %y
; _inverse : swap the arms of the select.
; CHECK: ogt:
@@ -16,6 +18,9 @@
; UNSAFE: ogt:
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ogt:
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ogt(double %x, double %y) nounwind {
%c = fcmp ogt double %x, %y
%d = select i1 %c, double %x, double %y
@@ -28,6 +33,9 @@
; UNSAFE: olt:
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: olt:
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @olt(double %x, double %y) nounwind {
%c = fcmp olt double %x, %y
%d = select i1 %c, double %x, double %y
@@ -42,6 +50,10 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ogt_inverse:
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ogt_inverse(double %x, double %y) nounwind {
%c = fcmp ogt double %x, %y
%d = select i1 %c, double %y, double %x
@@ -56,6 +68,10 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: olt_inverse:
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @olt_inverse(double %x, double %y) nounwind {
%c = fcmp olt double %x, %y
%d = select i1 %c, double %y, double %x
@@ -67,6 +83,9 @@
; UNSAFE: oge:
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: oge:
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @oge(double %x, double %y) nounwind {
%c = fcmp oge double %x, %y
%d = select i1 %c, double %x, double %y
@@ -77,6 +96,8 @@
; CHECK-NEXT: ucomisd %xmm0, %xmm1
; UNSAFE: ole:
; UNSAFE-NEXT: minsd %xmm1, %xmm0
+; FINITE: ole:
+; FINITE-NEXT: minsd %xmm1, %xmm0
define double @ole(double %x, double %y) nounwind {
%c = fcmp ole double %x, %y
%d = select i1 %c, double %x, double %y
@@ -89,6 +110,10 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: oge_inverse:
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @oge_inverse(double %x, double %y) nounwind {
%c = fcmp oge double %x, %y
%d = select i1 %c, double %y, double %x
@@ -101,6 +126,10 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ole_inverse:
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ole_inverse(double %x, double %y) nounwind {
%c = fcmp ole double %x, %y
%d = select i1 %c, double %y, double %x
@@ -115,6 +144,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ogt:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ogt(double %x) nounwind {
%c = fcmp ogt double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -129,6 +162,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_olt:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_olt(double %x) nounwind {
%c = fcmp olt double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -145,6 +182,11 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ogt_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ogt_inverse(double %x) nounwind {
%c = fcmp ogt double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -161,6 +203,11 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_olt_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_olt_inverse(double %x) nounwind {
%c = fcmp olt double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -173,6 +220,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_oge:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_oge(double %x) nounwind {
%c = fcmp oge double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -185,6 +236,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ole:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ole(double %x) nounwind {
%c = fcmp ole double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -198,6 +253,11 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_oge_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_oge_inverse(double %x) nounwind {
%c = fcmp oge double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -211,6 +271,11 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ole_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ole_inverse(double %x) nounwind {
%c = fcmp ole double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -222,6 +287,9 @@
; UNSAFE: ugt:
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ugt:
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ugt(double %x, double %y) nounwind {
%c = fcmp ugt double %x, %y
%d = select i1 %c, double %x, double %y
@@ -233,6 +301,9 @@
; UNSAFE: ult:
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ult:
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ult(double %x, double %y) nounwind {
%c = fcmp ult double %x, %y
%d = select i1 %c, double %x, double %y
@@ -245,6 +316,10 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ugt_inverse:
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ugt_inverse(double %x, double %y) nounwind {
%c = fcmp ugt double %x, %y
%d = select i1 %c, double %y, double %x
@@ -257,6 +332,10 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ult_inverse:
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ult_inverse(double %x, double %y) nounwind {
%c = fcmp ult double %x, %y
%d = select i1 %c, double %y, double %x
@@ -264,11 +343,15 @@
}
; CHECK: uge:
-; CHECK-NEXT: maxsd %xmm1, %xmm0
+; CHECK-NEXT: maxsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: uge:
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: uge:
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @uge(double %x, double %y) nounwind {
%c = fcmp uge double %x, %y
%d = select i1 %c, double %x, double %y
@@ -276,11 +359,15 @@
}
; CHECK: ule:
-; CHECK-NEXT: minsd %xmm1, %xmm0
+; CHECK-NEXT: minsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: ule:
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ule:
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ule(double %x, double %y) nounwind {
%c = fcmp ule double %x, %y
%d = select i1 %c, double %x, double %y
@@ -288,13 +375,16 @@
}
; CHECK: uge_inverse:
-; CHECK-NEXT: minsd %xmm0, %xmm1
-; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: minsd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: uge_inverse:
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: uge_inverse:
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @uge_inverse(double %x, double %y) nounwind {
%c = fcmp uge double %x, %y
%d = select i1 %c, double %y, double %x
@@ -302,13 +392,16 @@
}
; CHECK: ule_inverse:
-; CHECK-NEXT: maxsd %xmm0, %xmm1
-; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: maxsd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: ule_inverse:
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: ule_inverse:
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @ule_inverse(double %x, double %y) nounwind {
%c = fcmp ule double %x, %y
%d = select i1 %c, double %y, double %x
@@ -321,6 +414,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ugt:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ugt(double %x) nounwind {
%c = fcmp ugt double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -333,6 +430,10 @@
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ult:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ult(double %x) nounwind {
%c = fcmp ult double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -346,6 +447,11 @@
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ugt_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ugt_inverse(double %x) nounwind {
%c = fcmp ugt double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -359,6 +465,11 @@
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ult_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ult_inverse(double %x) nounwind {
%c = fcmp ult double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -367,12 +478,17 @@
; CHECK: x_uge:
; CHECK-NEXT: pxor %xmm1, %xmm1
-; CHECK-NEXT: maxsd %xmm1, %xmm0
+; CHECK-NEXT: maxsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: x_uge:
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: maxsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_uge:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_uge(double %x) nounwind {
%c = fcmp uge double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -381,12 +497,17 @@
; CHECK: x_ule:
; CHECK-NEXT: pxor %xmm1, %xmm1
-; CHECK-NEXT: minsd %xmm1, %xmm0
+; CHECK-NEXT: minsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: x_ule:
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: minsd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ule:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ule(double %x) nounwind {
%c = fcmp ule double %x, 0.000000e+00
%d = select i1 %c, double %x, double 0.000000e+00
@@ -395,14 +516,18 @@
; CHECK: x_uge_inverse:
; CHECK-NEXT: pxor %xmm1, %xmm1
-; CHECK-NEXT: minsd %xmm0, %xmm1
-; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: minsd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: x_uge_inverse:
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: minsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_uge_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_uge_inverse(double %x) nounwind {
%c = fcmp uge double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
@@ -411,26 +536,303 @@
; CHECK: x_ule_inverse:
; CHECK-NEXT: pxor %xmm1, %xmm1
-; CHECK-NEXT: maxsd %xmm0, %xmm1
-; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: maxsd %xmm1, %xmm0
; CHECK-NEXT: ret
; UNSAFE: x_ule_inverse:
; UNSAFE-NEXT: pxor %xmm1, %xmm1
; UNSAFE-NEXT: maxsd %xmm0, %xmm1
; UNSAFE-NEXT: movapd %xmm1, %xmm0
; UNSAFE-NEXT: ret
+; FINITE: x_ule_inverse:
+; FINITE-NEXT: pxor %xmm1, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
define double @x_ule_inverse(double %x) nounwind {
%c = fcmp ule double %x, 0.000000e+00
%d = select i1 %c, double 0.000000e+00, double %x
ret double %d
}
+; CHECK: y_ogt:
+; CHECK-NEXT: maxsd {{[^,]*}}, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_ogt:
+; UNSAFE-NEXT: maxsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ogt:
+; FINITE-NEXT: maxsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_ogt(double %x) nounwind {
+ %c = fcmp ogt double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_olt:
+; CHECK-NEXT: minsd {{[^,]*}}, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_olt:
+; UNSAFE-NEXT: minsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_olt:
+; FINITE-NEXT: minsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_olt(double %x) nounwind {
+ %c = fcmp olt double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_ogt_inverse:
+; CHECK-NEXT: movsd {{[^,]*}}, %xmm1
+; CHECK-NEXT: minsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_ogt_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: minsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ogt_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_ogt_inverse(double %x) nounwind {
+ %c = fcmp ogt double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_olt_inverse:
+; CHECK-NEXT: movsd {{[^,]*}}, %xmm1
+; CHECK-NEXT: maxsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_olt_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: maxsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_olt_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_olt_inverse(double %x) nounwind {
+ %c = fcmp olt double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_oge:
+; CHECK: ucomisd %xmm1, %xmm0
+; UNSAFE: y_oge:
+; UNSAFE-NEXT: maxsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_oge:
+; FINITE-NEXT: maxsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_oge(double %x) nounwind {
+ %c = fcmp oge double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_ole:
+; CHECK: ucomisd %xmm0, %xmm1
+; UNSAFE: y_ole:
+; UNSAFE-NEXT: minsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ole:
+; FINITE-NEXT: minsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_ole(double %x) nounwind {
+ %c = fcmp ole double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_oge_inverse:
+; CHECK: ucomisd %xmm1, %xmm0
+; UNSAFE: y_oge_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: minsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_oge_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_oge_inverse(double %x) nounwind {
+ %c = fcmp oge double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_ole_inverse:
+; CHECK: ucomisd %xmm0, %xmm1
+; UNSAFE: y_ole_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: maxsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ole_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_ole_inverse(double %x) nounwind {
+ %c = fcmp ole double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_ugt:
+; CHECK: ucomisd %xmm0, %xmm1
+; UNSAFE: y_ugt:
+; UNSAFE-NEXT: maxsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ugt:
+; FINITE-NEXT: maxsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_ugt(double %x) nounwind {
+ %c = fcmp ugt double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_ult:
+; CHECK: ucomisd %xmm1, %xmm0
+; UNSAFE: y_ult:
+; UNSAFE-NEXT: minsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ult:
+; FINITE-NEXT: minsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_ult(double %x) nounwind {
+ %c = fcmp ult double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_ugt_inverse:
+; CHECK: ucomisd %xmm0, %xmm1
+; UNSAFE: y_ugt_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: minsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ugt_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_ugt_inverse(double %x) nounwind {
+ %c = fcmp ugt double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_ult_inverse:
+; CHECK: ucomisd %xmm1, %xmm0
+; UNSAFE: y_ult_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: maxsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ult_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_ult_inverse(double %x) nounwind {
+ %c = fcmp ult double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_uge:
+; CHECK-NEXT: movsd {{[^,]*}}, %xmm1
+; CHECK-NEXT: maxsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_uge:
+; UNSAFE-NEXT: maxsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_uge:
+; FINITE-NEXT: maxsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_uge(double %x) nounwind {
+ %c = fcmp uge double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_ule:
+; CHECK-NEXT: movsd {{[^,]*}}, %xmm1
+; CHECK-NEXT: minsd %xmm0, %xmm1
+; CHECK-NEXT: movapd %xmm1, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_ule:
+; UNSAFE-NEXT: minsd {{[^,]*}}, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ule:
+; FINITE-NEXT: minsd {{[^,]*}}, %xmm0
+; FINITE-NEXT: ret
+define double @y_ule(double %x) nounwind {
+ %c = fcmp ule double %x, -0.000000e+00
+ %d = select i1 %c, double %x, double -0.000000e+00
+ ret double %d
+}
+
+; CHECK: y_uge_inverse:
+; CHECK-NEXT: minsd {{[^,]*}}, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_uge_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: minsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_uge_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: minsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_uge_inverse(double %x) nounwind {
+ %c = fcmp uge double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
+
+; CHECK: y_ule_inverse:
+; CHECK-NEXT: maxsd {{[^,]*}}, %xmm0
+; CHECK-NEXT: ret
+; UNSAFE: y_ule_inverse:
+; UNSAFE-NEXT: movsd {{[^,]*}}, %xmm1
+; UNSAFE-NEXT: maxsd %xmm0, %xmm1
+; UNSAFE-NEXT: movapd %xmm1, %xmm0
+; UNSAFE-NEXT: ret
+; FINITE: y_ule_inverse:
+; FINITE-NEXT: movsd {{[^,]*}}, %xmm1
+; FINITE-NEXT: maxsd %xmm0, %xmm1
+; FINITE-NEXT: movapd %xmm1, %xmm0
+; FINITE-NEXT: ret
+define double @y_ule_inverse(double %x) nounwind {
+ %c = fcmp ule double %x, -0.000000e+00
+ %d = select i1 %c, double -0.000000e+00, double %x
+ ret double %d
+}
; Test a few more misc. cases.
; CHECK: clampTo3k_a:
; CHECK: minsd
; UNSAFE: clampTo3k_a:
; UNSAFE: minsd
+; FINITE: clampTo3k_a:
+; FINITE: minsd
define double @clampTo3k_a(double %x) nounwind readnone {
entry:
%0 = fcmp ogt double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -442,6 +844,8 @@
; CHECK: minsd
; UNSAFE: clampTo3k_b:
; UNSAFE: minsd
+; FINITE: clampTo3k_b:
+; FINITE: minsd
define double @clampTo3k_b(double %x) nounwind readnone {
entry:
%0 = fcmp uge double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -453,6 +857,8 @@
; CHECK: maxsd
; UNSAFE: clampTo3k_c:
; UNSAFE: maxsd
+; FINITE: clampTo3k_c:
+; FINITE: maxsd
define double @clampTo3k_c(double %x) nounwind readnone {
entry:
%0 = fcmp olt double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -464,6 +870,8 @@
; CHECK: maxsd
; UNSAFE: clampTo3k_d:
; UNSAFE: maxsd
+; FINITE: clampTo3k_d:
+; FINITE: maxsd
define double @clampTo3k_d(double %x) nounwind readnone {
entry:
%0 = fcmp ule double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -475,6 +883,8 @@
; CHECK: maxsd
; UNSAFE: clampTo3k_e:
; UNSAFE: maxsd
+; FINITE: clampTo3k_e:
+; FINITE: maxsd
define double @clampTo3k_e(double %x) nounwind readnone {
entry:
%0 = fcmp olt double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -486,6 +896,8 @@
; CHECK: maxsd
; UNSAFE: clampTo3k_f:
; UNSAFE: maxsd
+; FINITE: clampTo3k_f:
+; FINITE: maxsd
define double @clampTo3k_f(double %x) nounwind readnone {
entry:
%0 = fcmp ule double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -497,6 +909,8 @@
; CHECK: minsd
; UNSAFE: clampTo3k_g:
; UNSAFE: minsd
+; FINITE: clampTo3k_g:
+; FINITE: minsd
define double @clampTo3k_g(double %x) nounwind readnone {
entry:
%0 = fcmp ogt double %x, 3.000000e+03 ; <i1> [#uses=1]
@@ -508,6 +922,8 @@
; CHECK: minsd
; UNSAFE: clampTo3k_h:
; UNSAFE: minsd
+; FINITE: clampTo3k_h:
+; FINITE: minsd
define double @clampTo3k_h(double %x) nounwind readnone {
entry:
%0 = fcmp uge double %x, 3.000000e+03 ; <i1> [#uses=1]