AMDGPU: Make v32bf16 a legal type (#76679)

Depends #76678

Author: arsenm

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -173,6 +173,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     addRegisterClass(MVT::v16bf16, &AMDGPU::SGPR_256RegClass);
     addRegisterClass(MVT::v32i16, &AMDGPU::SGPR_512RegClass);
     addRegisterClass(MVT::v32f16, &AMDGPU::SGPR_512RegClass);
+    addRegisterClass(MVT::v32bf16, &AMDGPU::SGPR_512RegClass);
   }
 
   addRegisterClass(MVT::v32i32, &AMDGPU::VReg_1024RegClass);
@@ -719,11 +720,15 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     AddPromotedToType(ISD::LOAD, MVT::v32i16, MVT::v16i32);
     setOperationAction(ISD::LOAD, MVT::v32f16, Promote);
     AddPromotedToType(ISD::LOAD, MVT::v32f16, MVT::v16i32);
+    setOperationAction(ISD::LOAD, MVT::v32bf16, Promote);
+    AddPromotedToType(ISD::LOAD, MVT::v32bf16, MVT::v16i32);
 
     setOperationAction(ISD::STORE, MVT::v32i16, Promote);
     AddPromotedToType(ISD::STORE, MVT::v32i16, MVT::v16i32);
     setOperationAction(ISD::STORE, MVT::v32f16, Promote);
     AddPromotedToType(ISD::STORE, MVT::v32f16, MVT::v16i32);
+    setOperationAction(ISD::STORE, MVT::v32bf16, Promote);
+    AddPromotedToType(ISD::STORE, MVT::v32bf16, MVT::v16i32);
 
     setOperationAction({ISD::ANY_EXTEND, ISD::ZERO_EXTEND, ISD::SIGN_EXTEND},
                        MVT::v2i32, Expand);

llvm/lib/Target/AMDGPU/SIInstructions.td

@@ -1801,6 +1801,38 @@ def : BitConvert <v8f64,  v16f32, VReg_512>;
 def : BitConvert <v16f32, v8i64,  VReg_512>;
 def : BitConvert <v16f32, v8f64,  VReg_512>;
 
+
+
+def : BitConvert <v32bf16, v32i16, VReg_512>;
+def : BitConvert <v32i16, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v32i16, SReg_512>;
+def : BitConvert <v32i16, v32bf16, SReg_512>;
+
+def : BitConvert <v32bf16, v32f16, VReg_512>;
+def : BitConvert <v32f16, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v32f16, SReg_512>;
+def : BitConvert <v32f16, v32bf16, SReg_512>;
+
+def : BitConvert <v32bf16, v16i32, VReg_512>;
+def : BitConvert <v16i32, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v16i32, SReg_512>;
+def : BitConvert <v16i32, v32bf16, SReg_512>;
+
+def : BitConvert <v32bf16, v16f32, VReg_512>;
+def : BitConvert <v16f32, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v16f32, SReg_512>;
+def : BitConvert <v16f32, v32bf16, SReg_512>;
+
+def : BitConvert <v32bf16, v8f64, VReg_512>;
+def : BitConvert <v8f64, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v8f64, SReg_512>;
+def : BitConvert <v8f64, v32bf16, SReg_512>;
+
+def : BitConvert <v32bf16, v8i64, VReg_512>;
+def : BitConvert <v8i64, v32bf16, VReg_512>;
+def : BitConvert <v32bf16, v8i64, SReg_512>;
+def : BitConvert <v8i64, v32bf16, SReg_512>;
+
 // 1024-bit bitcast
 def : BitConvert <v32i32, v32f32, VReg_1024>;
 def : BitConvert <v32f32, v32i32, VReg_1024>;

llvm/lib/Target/AMDGPU/SIRegisterInfo.td

@@ -916,7 +916,7 @@ defm "" : SRegClass<11, [v11i32, v11f32], SGPR_352Regs, TTMP_352Regs>;
 defm "" : SRegClass<12, [v12i32, v12f32], SGPR_384Regs, TTMP_384Regs>;
 
 let GlobalPriority = true in {
-defm "" : SRegClass<16, [v16i32, v16f32, v8i64, v8f64, v32i16, v32f16], SGPR_512Regs, TTMP_512Regs>;
+defm "" : SRegClass<16, [v16i32, v16f32, v8i64, v8f64, v32i16, v32f16, v32bf16], SGPR_512Regs, TTMP_512Regs>;
 defm "" : SRegClass<32, [v32i32, v32f32, v16i64, v16f64], SGPR_1024Regs>;
 }
 
@@ -970,7 +970,7 @@ defm VReg_352 : VRegClass<11, [v11i32, v11f32], (add VGPR_352)>;
 defm VReg_384 : VRegClass<12, [v12i32, v12f32], (add VGPR_384)>;
 
 let GlobalPriority = true in {
-defm VReg_512 : VRegClass<16, [v16i32, v16f32, v8i64, v8f64, v32i16, v32f16], (add VGPR_512)>;
+defm VReg_512 : VRegClass<16, [v16i32, v16f32, v8i64, v8f64, v32i16, v32f16, v32bf16], (add VGPR_512)>;
 defm VReg_1024 : VRegClass<32, [v32i32, v32f32, v16i64, v16f64], (add VGPR_1024)>;
 }