Lower unsigned vsetcc to psubus in certain cases

The current approach to lower a vsetult is to flip the sign bit of the
operands, swap the operands and then use a (signed) pcmpgt.  psubus (unsigned
saturating subtract) can be used to emulate a vsetult more efficiently:

+    case ISD::SETULT: {
+      // If the comparison is against a constant we can turn this into a
+      // setule.  With psubus, setule does not require a swap.  This is
+      // beneficial because the constant in the register is no longer
+      // destructed as the destination so it can be hoisted out of a loop.

I also enable lowering via psubus in a few other cases where it's clearly
beneficial: setule and setuge if minu/maxu cannot be used.
    
rdar://problem/14338765

Patch by Adam Nemet <anemet@apple.com>.

llvm-svn: 202301
This commit is contained in:
Quentin Colombet 2014-02-26 21:39:12 +00:00
parent 8e9791f62c
commit 85c9e16291
3 changed files with 171 additions and 12 deletions

View File

@ -10009,6 +10009,37 @@ static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
DAG.getConstant(SSECC, MVT::i8));
}
/// \brief Try to turn a VSETULT into a VSETULE by modifying its second
/// operand \p Op1. If non-trivial (for example because it's not constant)
/// return an empty value.
static SDValue ChangeVSETULTtoVSETULE(SDValue Op1, SelectionDAG &DAG)
{
BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
if (!BV)
return SDValue();
MVT VT = Op1.getSimpleValueType();
MVT EVT = VT.getVectorElementType();
unsigned n = VT.getVectorNumElements();
SmallVector<SDValue, 8> ULTOp1;
for (unsigned i = 0; i < n; ++i) {
ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
return SDValue();
// Avoid underflow.
APInt Val = Elt->getAPIntValue();
if (Val == 0)
return SDValue();
ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
}
return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op1), VT, ULTOp1.data(),
ULTOp1.size());
}
static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
SDValue Op0 = Op.getOperand(0);
@ -10080,6 +10111,7 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
// operations may be required for some comparisons.
unsigned Opc;
bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
bool Subus = false;
switch (SetCCOpcode) {
default: llvm_unreachable("Unexpected SETCC condition");
@ -10114,6 +10146,40 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
}
bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
if (!MinMax && hasSubus) {
// As another special case, use PSUBUS[BW] when it's profitable. E.g. for
// Op0 u<= Op1:
// t = psubus Op0, Op1
// pcmpeq t, <0..0>
switch (SetCCOpcode) {
default: break;
case ISD::SETULT: {
// If the comparison is against a constant we can turn this into a
// setule. With psubus, setule does not require a swap. This is
// beneficial because the constant in the register is no longer
// destructed as the destination so it can be hoisted out of a loop.
// Only do this pre-AVX since vpcmp* is no longer destructive.
if (Subtarget->hasAVX())
break;
SDValue ULEOp1 = ChangeVSETULTtoVSETULE(Op1, DAG);
if (ULEOp1.getNode()) {
Op1 = ULEOp1;
Subus = true; Invert = false; Swap = false;
}
break;
}
// Psubus is better than flip-sign because it requires no inversion.
case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
}
if (Subus) {
Opc = X86ISD::SUBUS;
FlipSigns = false;
}
}
if (Swap)
std::swap(Op0, Op1);
@ -10204,6 +10270,10 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
if (MinMax)
Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
if (Subus)
Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
getZeroVector(VT, Subtarget, DAG, dl));
return Result;
}

View File

@ -0,0 +1,95 @@
; RUN: llc < %s -o - -mcpu=generic -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s
; RUN: llc < %s -o - -mcpu=generic -mtriple=x86_64-apple-darwin -mattr=+sse4.2 | FileCheck %s
; For a setult against a constant, turn it into a setule and lower via psubusw.
define void @loop_no_const_reload(<2 x i64>* %in, <2 x i64>* %out, i32 %n) {
; CHECK: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-NEXT: .short 25
; CHECK-LABEL: loop_no_const_reload:
; CHECK: psubusw
; Constant is no longer clobbered so no need to reload it in the loop.
; CHECK-NOT: movdqa {{%xmm[0-9]+}}, {{%xmm[0-9]+}}
entry:
%cmp9 = icmp eq i32 %n, 0
br i1 %cmp9, label %for.end, label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
%arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
%arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
%0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
%cmp.i.i = icmp ult <8 x i16> %0, <i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26>
%sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
%1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
%arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
%exitcond = icmp eq i32 %lftr.wideiv, %n
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body, %entry
ret void
}
; Be careful if decrementing the constant would undeflow.
define void @loop_const_folding_underflow(<2 x i64>* %in, <2 x i64>* %out, i32 %n) {
; CHECK-NOT: .short 25
; CHECK-LABEL: loop_const_folding_underflow:
; CHECK-NOT: psubusw
entry:
%cmp9 = icmp eq i32 %n, 0
br i1 %cmp9, label %for.end, label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
%arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
%arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
%0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
%cmp.i.i = icmp ult <8 x i16> %0, <i16 0, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26>
%sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
%1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
%arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
%exitcond = icmp eq i32 %lftr.wideiv, %n
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body, %entry
ret void
}
; Test for PSUBUSB
define <16 x i8> @test_ult_byte(<16 x i8> %a) {
; CHECK: .space 16,10
; CHECK-LABEL: test_ult_byte:
; CHECK: psubus
entry:
%icmp = icmp ult <16 x i8> %a, <i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11>
%sext = sext <16 x i1> %icmp to <16 x i8>
ret <16 x i8> %sext
}
; Only do this when we can turn the comparison into a setule. I.e. not for
; register operands.
define <8 x i16> @test_ult_register(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_ult_register:
; CHECK-NOT: psubus
entry:
%icmp = icmp ult <8 x i16> %a, %b
%sext = sext <8 x i1> %icmp to <8 x i16>
ret <8 x i16> %sext
}

View File

@ -42,12 +42,9 @@ define <8 x i16> @v8i16_icmp_uge(<8 x i16> %a, <8 x i16> %b) nounwind readnone s
%2 = sext <8 x i1> %1 to <8 x i16>
ret <8 x i16> %2
; SSE2-LABEL: v8i16_icmp_uge:
; SSE2: movdqa {{.*}}(%rip), %xmm2
; SEE2: pxor %xmm2, %xmm0
; SSE2: pxor %xmm1, %xmm2
; SSE2: pcmpgtw %xmm0, %xmm2
; SSE2: pcmpeqd %xmm0, %xmm0
; SSE2: pxor %xmm2, %xmm0
; SSE2: psubusw %xmm0, %xmm1
; SEE2: pxor %xmm0, %xmm0
; SSE2: pcmpeqw %xmm1, %xmm0
; SSE41-LABEL: v8i16_icmp_uge:
; SSE41: pmaxuw %xmm0, %xmm1
@ -63,12 +60,9 @@ define <8 x i16> @v8i16_icmp_ule(<8 x i16> %a, <8 x i16> %b) nounwind readnone s
%2 = sext <8 x i1> %1 to <8 x i16>
ret <8 x i16> %2
; SSE2-LABEL: v8i16_icmp_ule:
; SSE2: movdqa {{.*}}(%rip), %xmm2
; SSE2: pxor %xmm2, %xmm1
; SSE2: pxor %xmm2, %xmm0
; SSE2: pcmpgtw %xmm1, %xmm0
; SSE2: pcmpeqd %xmm1, %xmm1
; SSE2: pxor %xmm0, %xmm1
; SSE2: psubusw %xmm1, %xmm0
; SSE2: pxor %xmm1, %xmm1
; SSE2: pcmpeqw %xmm0, %xmm1
; SSE2: movdqa %xmm1, %xmm0
; SSE41-LABEL: v8i16_icmp_ule: