AMDGPU/GlobalISel: Implement select() for 32-bit @llvm.minnun and @llvm.maxnum

[tstellar]

Rebase patch.

[arsenm]

Also f64 tests?

[tstellar]

Add f64 patterns

[tstellar]

Rebase patch on ToT.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUGISel.td

class GISelVop3Pat2CommutePat <
   Instruction inst,
   ValueType dst_vt,
   ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt>   : GCNPat <
 
   (dst_vt (node (src0_vt (sd_vcsrc src0_vt:$src0)), (src1_vt (sd_vcsrc src1_vt:$src1)))),
   (inst src0_vt:$src1, src1_vt:$src0)
 >;
 
+class GISelVop3Pat2ModsPat <
+  SDPatternOperator node,
+  Instruction inst,
+  ValueType dst_vt,
+  ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
+
+  (dst_vt (node (src0_vt (VOP3Mods0 src0_vt:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omods)),
+                (src1_vt (VOP3Mods src1_vt:$src1, i32:$src1_modifiers)))),
+  (inst i32:$src0_modifiers, src0_vt:$src0,
+        i32:$src1_modifiers, src1_vt:$src1, $clamp, $omods)
+>;
+
 multiclass GISelVop2IntrPat <
   SDPatternOperator node, Instruction inst,
   ValueType dst_vt, ValueType src_vt = dst_vt> {
 
   def : GISelVop2Pat <node, inst, dst_vt, src_vt>;
 
   // FIXME: Intrinsics aren't marked as commutable, so we need to add an explcit
   // pattern to handle commuting.  This is another reason why legalizing to a
   // generic machine instruction may be better that matching the intrinsic
   // directly.
   def : GISelVop2CommutePat <node, inst, dst_vt, src_vt>;
 }
 
+
 def : GISelSop2Pat <or, S_OR_B32, i32>;
 def : GISelVop2Pat <or, V_OR_B32_e32, i32>;
 
 def : GISelSop2Pat <sra, S_ASHR_I32, i32>;
 let AddedComplexity = 100 in {
 let SubtargetPredicate = isSICI in {
 def : GISelVop2Pat <sra, V_ASHR_I32_e32, i32>;
 }
 def : GISelVop2CommutePat <sra, V_ASHRREV_I32_e32, i32>;
 }
 def : GISelVop3Pat2CommutePat <sra, V_ASHRREV_I32_e64, i32>;
 
 // FIXME: Select directly to _e32 so we don't need to deal with modifiers.
 // FIXME: We can't re-use SelectionDAG patterns here because they match
 // against a custom SDNode and we would need to create a generic machine
 // instruction that is equivalent to the custom SDNode.  This would also require
 // us to custom legalize the intrinsic to the new generic machine instruction,
 // but I can't get custom legalizing of intrinsic to work and I'm not sure if
 // this is even supported yet.
 defm : GISelVop2IntrPat <
   int_amdgcn_cvt_pkrtz, V_CVT_PKRTZ_F16_F32_e32, v2f16, f32>;
+
+defm : GISelVop2IntrPat <int_maxnum, V_MAX_F32_e32, f32>;
+def : GISelVop3Pat2ModsPat <int_maxnum, V_MAX_F64, f64>;
+defm : GISelVop2IntrPat <int_minnum, V_MIN_F32_e32, f32>;
+def : GISelVop3Pat2ModsPat <int_minnum, V_MIN_F64, f64>;

lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

 
 bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I,
                                           CodeGenCoverage &CoverageInfo) const {
   unsigned IntrinsicID =  I.getOperand(1).getIntrinsicID();
 
   switch (IntrinsicID) {
   default:
     break;
+  case Intrinsic::maxnum:
+  case Intrinsic::minnum:
   case Intrinsic::amdgcn_cvt_pkrtz:
     return selectImpl(I, CoverageInfo);
   }
   return false;
 }
 
 static MachineInstr *
 buildEXP(const TargetInstrInfo &TII, MachineInstr *Insert, unsigned Tgt,

test/CodeGen/AMDGPU/GlobalISel/inst-select-maxnum.mir

+# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefixes=GCN
+
+--- |
+  define void @maxnum(i32 addrspace(1)* %global0) { ret void }
+...
+---
+
+name:            maxnum
+legalized:       true
+regBankSelected: true
+
+# GCN-LABEL: name: maxnum
+body: |
+  bb.0:
+    liveins: $sgpr0, $vgpr0, $vgpr1, $vgpr3_vgpr4, $sgpr10_sgpr11, $vgpr10_vgpr11, $vgpr12_vgpr13
+    ; GCN: [[SGPR0:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: [[VGPR0:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GCN: [[VGPR1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    %0:sgpr(s32) = COPY $sgpr0
+    %1:vgpr(s32) = COPY $vgpr0
+    %2:vgpr(s32) = COPY $vgpr1
+    %3:vgpr(s64) = COPY $vgpr3_vgpr4
+
+    ; GCN: [[SGPR64_0:%[0-9]+]]:sreg_64_xexec = COPY $sgpr10_sgpr11
+    ; GCN: [[VGPR64_0:%[0-9]+]]:vreg_64 = COPY $vgpr10_vgpr11
+    ; GCN: [[VGPR64_1:%[0-9]+]]:vreg_64 = COPY $vgpr12_vgpr13
+    %10:sgpr(s64) = COPY $sgpr10_sgpr11
+    %11:vgpr(s64) = COPY $vgpr10_vgpr11
+    %12:vgpr(s64) = COPY $vgpr12_vgpr13
+
+    ; maxnum vs
+    ; GCN: V_MAX_F32_e32 [[SGPR0]], [[VGPR0]]
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.maxnum.f32), %1, %0
+
+    ; maxnum sv
+    ; GCN: V_MAX_F32_e32 [[SGPR0]], [[VGPR0]]
+    %5:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.maxnum.f32), %0, %1
+
+    ; maxnum vv
+    ; GCN: V_MAX_F32_e32 [[VGPR0]], [[VGPR1]]
+    %6:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.maxnum.f32), %1, %2
+
+    G_STORE %4, %3 :: (store 4 into %ir.global0)
+    G_STORE %5, %3 :: (store 4 into %ir.global0)
+    G_STORE %6, %3 :: (store 4 into %ir.global0)
+
+    ; 64-bit
+
+    ; maxnum vs
+    ; GCN: V_MAX_F64 0, [[SGPR64_0]], 0, [[VGPR64_0]], 0, 0
+    %14:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.maxnum.f64), %10, %11
+
+    ; maxnum sv
+    ; GCN: V_MAX_F64 0, [[VGPR64_0]], 0, [[SGPR64_0]], 0, 0
+    %15:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.maxnum.f64), %11, %10
+
+    ; maxnum vv
+    ; GCN: V_MAX_F64 0, [[VGPR64_0]], 0, [[VGPR64_1]], 0, 0
+    %16:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.maxnum.f64), %11, %12
+
+    G_STORE %14, %3 :: (store 8 into %ir.global0)
+    G_STORE %15, %3 :: (store 8 into %ir.global0)
+    G_STORE %16, %3 :: (store 8 into %ir.global0)
+
+...
+---

test/CodeGen/AMDGPU/GlobalISel/inst-select-minnum.mir

+# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefixes=GCN
+
+--- |
+  define void @minnum(i32 addrspace(1)* %global0) { ret void }
+...
+---
+
+name:            minnum
+legalized:       true
+regBankSelected: true
+
+# GCN-LABEL: name: minnum
+body: |
+  bb.0:
+    liveins: $sgpr0, $vgpr0, $vgpr1, $vgpr3_vgpr4, $sgpr10_sgpr11, $vgpr10_vgpr11, $vgpr12_vgpr13
+    ; GCN: [[SGPR0:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: [[VGPR0:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GCN: [[VGPR1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    %0:sgpr(s32) = COPY $sgpr0
+    %1:vgpr(s32) = COPY $vgpr0
+    %2:vgpr(s32) = COPY $vgpr1
+    %3:vgpr(s64) = COPY $vgpr3_vgpr4
+
+    ; GCN: [[SGPR64_0:%[0-9]+]]:sreg_64_xexec = COPY $sgpr10_sgpr11
+    ; GCN: [[VGPR64_0:%[0-9]+]]:vreg_64 = COPY $vgpr10_vgpr11
+    ; GCN: [[VGPR64_1:%[0-9]+]]:vreg_64 = COPY $vgpr12_vgpr13
+    %10:sgpr(s64) = COPY $sgpr10_sgpr11
+    %11:vgpr(s64) = COPY $vgpr10_vgpr11
+    %12:vgpr(s64) = COPY $vgpr12_vgpr13
+
+    ; minnum vs
+    ; GCN: V_MIN_F32_e32 [[SGPR0]], [[VGPR0]]
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.minnum.f32), %1, %0
+
+    ; minnum sv
+    ; GCN: V_MIN_F32_e32 [[SGPR0]], [[VGPR0]]
+    %5:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.minnum.f32), %0, %1
+
+    ; minnum vv
+    ; GCN: V_MIN_F32_e32 [[VGPR0]], [[VGPR1]]
+    %6:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.minnum.f32), %1, %2
+
+    G_STORE %4, %3 :: (store 4 into %ir.global0)
+    G_STORE %5, %3 :: (store 4 into %ir.global0)
+    G_STORE %6, %3 :: (store 4 into %ir.global0)
+
+    ; 64-bit
+
+    ; minnum vs
+    ; GCN: V_MIN_F64 0, [[SGPR64_0]], 0, [[VGPR64_0]], 0, 0
+    %14:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.minnum.f64), %10, %11
+
+    ; minnum sv
+    ; GCN: V_MIN_F64 0, [[VGPR64_0]], 0, [[SGPR64_0]], 0, 0
+    %15:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.minnum.f64), %11, %10
+
+    ; minnum vv
+    ; GCN: V_MIN_F64 0, [[VGPR64_0]], 0, [[VGPR64_1]], 0, 0
+    %16:vgpr(s64) = G_INTRINSIC intrinsic(@llvm.minnum.f64), %11, %12
+
+    G_STORE %14, %3 :: (store 8 into %ir.global0)
+    G_STORE %15, %3 :: (store 8 into %ir.global0)
+    G_STORE %16, %3 :: (store 8 into %ir.global0)
+...
+---