[llvm] r288636 - [AVX-512] Teach fast isel to use masked compare and movss for handling scalar cmp and select sequence when AVX-512 is enabled. This matches the behavior of normal isel.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Sun Dec 4 20:51:32 PST 2016
Author: ctopper
Date: Sun Dec 4 22:51:31 2016
New Revision: 288636
URL: http://llvm.org/viewvc/llvm-project?rev=288636&view=rev
Log:
[AVX-512] Teach fast isel to use masked compare and movss for handling scalar cmp and select sequence when AVX-512 is enabled. This matches the behavior of normal isel.
Modified:
llvm/trunk/lib/Target/X86/X86FastISel.cpp
llvm/trunk/test/CodeGen/X86/fast-isel-select-sse.ll
Modified: llvm/trunk/lib/Target/X86/X86FastISel.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86FastISel.cpp?rev=288636&r1=288635&r2=288636&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86FastISel.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86FastISel.cpp Sun Dec 4 22:51:31 2016
@@ -170,6 +170,12 @@ private:
const MachineInstrBuilder &addFullAddress(const MachineInstrBuilder &MIB,
X86AddressMode &AM);
+
+ unsigned fastEmitInst_rrrr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC, unsigned Op0,
+ bool Op0IsKill, unsigned Op1, bool Op1IsKill,
+ unsigned Op2, bool Op2IsKill, unsigned Op3,
+ bool Op3IsKill);
};
} // end anonymous namespace.
@@ -2180,9 +2186,36 @@ bool X86FastISel::X86FastEmitSSESelect(M
const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
unsigned ResultReg;
-
- if (Subtarget->hasAVX()) {
- const TargetRegisterClass *FR32 = &X86::FR32RegClass;
+
+ if (Subtarget->hasAVX512()) {
+ // If we have AVX512 we can use a mask compare and masked movss/sd.
+ const TargetRegisterClass *VR128X = &X86::VR128XRegClass;
+ const TargetRegisterClass *VK1 = &X86::VK1RegClass;
+
+ unsigned CmpOpcode =
+ (RetVT.SimpleTy == MVT::f32) ? X86::VCMPSSZrr : X86::VCMPSDZrr;
+ unsigned CmpReg = fastEmitInst_rri(CmpOpcode, VK1, CmpLHSReg, CmpLHSIsKill,
+ CmpRHSReg, CmpRHSIsKill, CC);
+
+ // Need an IMPLICIT_DEF for the input that is used to generate the upper
+ // bits of the result register since its not based on any of the inputs.
+ unsigned ImplicitDefReg = createResultReg(VR128X);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
+
+ // Place RHSReg is the passthru of the masked movss/sd operation and put
+ // LHS in the input. The mask input comes from the compare.
+ unsigned MovOpcode =
+ (RetVT.SimpleTy == MVT::f32) ? X86::VMOVSSZrrk : X86::VMOVSDZrrk;
+ unsigned MovReg = fastEmitInst_rrrr(MovOpcode, VR128X, RHSReg, RHSIsKill,
+ CmpReg, true, ImplicitDefReg, true,
+ LHSReg, LHSIsKill);
+
+ ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(MovReg);
+
+ } else if (Subtarget->hasAVX()) {
const TargetRegisterClass *VR128 = &X86::VR128RegClass;
// If we have AVX, create 1 blendv instead of 3 logic instructions.
@@ -2195,7 +2228,7 @@ bool X86FastISel::X86FastEmitSSESelect(M
unsigned BlendOpcode =
(RetVT.SimpleTy == MVT::f32) ? X86::VBLENDVPSrr : X86::VBLENDVPDrr;
- unsigned CmpReg = fastEmitInst_rri(CmpOpcode, FR32, CmpLHSReg, CmpLHSIsKill,
+ unsigned CmpReg = fastEmitInst_rri(CmpOpcode, RC, CmpLHSReg, CmpLHSIsKill,
CmpRHSReg, CmpRHSIsKill, CC);
unsigned VBlendReg = fastEmitInst_rrr(BlendOpcode, VR128, RHSReg, RHSIsKill,
LHSReg, LHSIsKill, CmpReg, true);
@@ -3849,6 +3882,38 @@ bool X86FastISel::tryToFoldLoadIntoMI(Ma
return true;
}
+unsigned X86FastISel::fastEmitInst_rrrr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ unsigned Op2, bool Op2IsKill,
+ unsigned Op3, bool Op3IsKill) {
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
+
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+ Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
+ Op2 = constrainOperandRegClass(II, Op2, II.getNumDefs() + 2);
+ Op2 = constrainOperandRegClass(II, Op2, II.getNumDefs() + 3);
+
+ if (II.getNumDefs() >= 1)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
+ .addReg(Op0, getKillRegState(Op0IsKill))
+ .addReg(Op1, getKillRegState(Op1IsKill))
+ .addReg(Op2, getKillRegState(Op2IsKill))
+ .addReg(Op3, getKillRegState(Op3IsKill));
+ else {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
+ .addReg(Op0, getKillRegState(Op0IsKill))
+ .addReg(Op1, getKillRegState(Op1IsKill))
+ .addReg(Op2, getKillRegState(Op2IsKill))
+ .addReg(Op3, getKillRegState(Op3IsKill));
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
+ }
+ return ResultReg;
+}
+
namespace llvm {
FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo,
Modified: llvm/trunk/test/CodeGen/X86/fast-isel-select-sse.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/fast-isel-select-sse.ll?rev=288636&r1=288635&r2=288636&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/fast-isel-select-sse.ll (original)
+++ llvm/trunk/test/CodeGen/X86/fast-isel-select-sse.ll Sun Dec 4 22:51:31 2016
@@ -3,8 +3,8 @@
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefix=SSE
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx | FileCheck %s --check-prefix=AVX
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -fast-isel -fast-isel-abort=1 -mattr=avx | FileCheck %s --check-prefix=AVX
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512SLOW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -fast-isel -fast-isel-abort=1 -mattr=avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512FAST
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefix=AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -fast-isel -fast-isel-abort=1 -mattr=avx512f | FileCheck %s --check-prefix=AVX512
; Test all cmp predicates that can be used with SSE.
@@ -23,18 +23,12 @@ define float @select_fcmp_oeq_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_oeq_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpeqss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_oeq_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpeqss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_oeq_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpeqss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp oeq float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -56,18 +50,12 @@ define double @select_fcmp_oeq_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_oeq_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpeqsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_oeq_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpeqsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_oeq_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpeqsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp oeq double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -90,18 +78,12 @@ define float @select_fcmp_ogt_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ogt_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpltss %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ogt_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpltss %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ogt_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpltss %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ogt float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -124,18 +106,12 @@ define double @select_fcmp_ogt_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ogt_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpltsd %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ogt_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpltsd %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ogt_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpltsd %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ogt double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -158,18 +134,12 @@ define float @select_fcmp_oge_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_oge_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpless %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_oge_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpless %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_oge_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpless %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp oge float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -192,18 +162,12 @@ define double @select_fcmp_oge_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_oge_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmplesd %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_oge_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmplesd %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_oge_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmplesd %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp oge double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -225,18 +189,12 @@ define float @select_fcmp_olt_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_olt_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpltss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_olt_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpltss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_olt_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpltss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp olt float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -258,18 +216,12 @@ define double @select_fcmp_olt_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_olt_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpltsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_olt_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpltsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_olt_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpltsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp olt double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -291,18 +243,12 @@ define float @select_fcmp_ole_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ole_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpless %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ole_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpless %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ole_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpless %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ole float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -324,18 +270,12 @@ define double @select_fcmp_ole_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ole_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmplesd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ole_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmplesd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ole_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmplesd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ole double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -357,18 +297,12 @@ define float @select_fcmp_ord_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ord_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpordss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ord_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpordss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ord_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpordss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ord float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -390,18 +324,12 @@ define double @select_fcmp_ord_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ord_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpordsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ord_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpordsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ord_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpordsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ord double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -423,18 +351,12 @@ define float @select_fcmp_uno_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_uno_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpunordss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_uno_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpunordss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_uno_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpunordss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp uno float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -456,18 +378,12 @@ define double @select_fcmp_uno_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_uno_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpunordsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_uno_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpunordsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_uno_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpunordsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp uno double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -489,18 +405,12 @@ define float @select_fcmp_ugt_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ugt_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnless %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ugt_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnless %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ugt_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnless %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ugt float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -522,18 +432,12 @@ define double @select_fcmp_ugt_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ugt_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnlesd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ugt_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnlesd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ugt_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnlesd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ugt double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -555,18 +459,12 @@ define float @select_fcmp_uge_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_uge_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnltss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_uge_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnltss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_uge_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnltss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp uge float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -588,18 +486,12 @@ define double @select_fcmp_uge_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_uge_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnltsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_uge_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnltsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_uge_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnltsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp uge double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -622,18 +514,12 @@ define float @select_fcmp_ult_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ult_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnless %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ult_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnless %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ult_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnless %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ult float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -656,18 +542,12 @@ define double @select_fcmp_ult_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ult_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnlesd %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ult_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnlesd %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ult_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnlesd %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ult double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -690,18 +570,12 @@ define float @select_fcmp_ule_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ule_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnltss %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ule_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnltss %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ule_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnltss %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ule float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -724,18 +598,12 @@ define double @select_fcmp_ule_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_ule_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpnltsd %xmm0, %xmm1, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_ule_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpnltsd %xmm0, %xmm1, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_ule_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpnltsd %xmm0, %xmm1, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp ule double %a, %b
%2 = select i1 %1, double %c, double %d
@@ -757,18 +625,12 @@ define float @select_fcmp_une_f32(float
; AVX-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_une_f32:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpneqss %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovaps %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_une_f32:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpneqss %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_une_f32:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpneqss %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovss %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovaps %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp une float %a, %b
%2 = select i1 %1, float %c, float %d
@@ -790,18 +652,12 @@ define double @select_fcmp_une_f64(doubl
; AVX-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
; AVX-NEXT: retq
;
-; AVX512SLOW-LABEL: select_fcmp_une_f64:
-; AVX512SLOW: # BB#0:
-; AVX512SLOW-NEXT: vcmpneqsd %xmm1, %xmm0, %k1
-; AVX512SLOW-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
-; AVX512SLOW-NEXT: vmovapd %xmm3, %xmm0
-; AVX512SLOW-NEXT: retq
-;
-; AVX512FAST-LABEL: select_fcmp_une_f64:
-; AVX512FAST: # BB#0:
-; AVX512FAST-NEXT: vcmpneqsd %xmm1, %xmm0, %xmm0
-; AVX512FAST-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
-; AVX512FAST-NEXT: retq
+; AVX512-LABEL: select_fcmp_une_f64:
+; AVX512: # BB#0:
+; AVX512-NEXT: vcmpneqsd %xmm1, %xmm0, %k1
+; AVX512-NEXT: vmovsd %xmm2, %xmm0, %xmm3 {%k1}
+; AVX512-NEXT: vmovapd %xmm3, %xmm0
+; AVX512-NEXT: retq
;
%1 = fcmp une double %a, %b
%2 = select i1 %1, double %c, double %d
More information about the llvm-commits
mailing list