AMDGPU/SI: Promote i16 = fp_[us]int f32 for VI

[• tstellarAMD]

Summary

This fixes a regression caused by r286464.

[arsenm]

LGTM with test names fixed

test/CodeGen/AMDGPU/fp_to_sint.ll

251–254

Test name should include f32_to_i16

test/CodeGen/AMDGPU/fp_to_uint.ll

249

should include source type in name

lib/Target/AMDGPU/SIISelLowering.cpp

if (Subtarget->has16BitInsts()) {
     setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
     AddPromotedToType(ISD::UINT_TO_FP, MVT::i16, MVT::i32);
     setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
     AddPromotedToType(ISD::SINT_TO_FP, MVT::i16, MVT::i32);
     setOperationAction(ISD::FP16_TO_FP, MVT::i16, Promote);
     AddPromotedToType(ISD::FP16_TO_FP, MVT::i16, MVT::i32);
     setOperationAction(ISD::FP_TO_FP16, MVT::i16, Promote);
     AddPromotedToType(ISD::FP_TO_FP16, MVT::i16, MVT::i32);
+
+    setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
+    AddPromotedToType(ISD::FP_TO_SINT, MVT::i16, MVT::i32);
+
+    setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
+    AddPromotedToType(ISD::FP_TO_UINT, MVT::i16, MVT::i32);
   }
 
   setTargetDAGCombine(ISD::FADD);
   setTargetDAGCombine(ISD::FSUB);
   setTargetDAGCombine(ISD::FMINNUM);
   setTargetDAGCombine(ISD::FMAXNUM);
   setTargetDAGCombine(ISD::SMIN);
   setTargetDAGCombine(ISD::SMAX);

test/CodeGen/AMDGPU/fp_to_sint.ll

 ; SI: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, -1.0, |s{{[0-9]+}}|
 define void @fp_to_uint_fabs_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
   %in.fabs = call float @llvm.fabs.f32(float %in)
   %conv = fptosi float %in.fabs to i1
   store i1 %conv, i1 addrspace(1)* %out
   ret void
 }
 
+; FUNC-LABEL: {{^}}fp_to_sint_i16:
+; SI: v_cvt_i32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
+; SI: buffer_store_short [[VAL]]
+define void @fp_to_sint_i16(i16 addrspace(1)* %out, float %in) #0 {
+  %sint = fptosi float %in to i16
+  store i16 %sint, i16 addrspace(1)* %out
+  ret void
+}
+
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }

test/CodeGen/AMDGPU/fp_to_uint.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s -check-prefix=SI -check-prefix=FUNC
-; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s -check-prefix=SI -check-prefix=FUNC
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GCN,FUNC,SI
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GCN,FUNC,VI
 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck %s -check-prefix=EG -check-prefix=FUNC
 
 declare float @llvm.fabs.f32(float) #1
 
 ; FUNC-LABEL: {{^}}fp_to_uint_f32_to_i32:
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
 
-; SI: v_cvt_u32_f32_e32
-; SI: s_endpgm
+; GCN: v_cvt_u32_f32_e32
+; GCN: s_endpgm
 define void @fp_to_uint_f32_to_i32 (i32 addrspace(1)* %out, float %in) {
   %conv = fptoui float %in to i32
   store i32 %conv, i32 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}fp_to_uint_v2f32_to_v2i32:
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
-; SI: v_cvt_u32_f32_e32
-; SI: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
 define void @fp_to_uint_v2f32_to_v2i32(<2 x i32> addrspace(1)* %out, <2 x float> %in) {
   %result = fptoui <2 x float> %in to <2 x i32>
   store <2 x i32> %result, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}fp_to_uint_v4f32_to_v4i32:
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
 ; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
-; SI: v_cvt_u32_f32_e32
-; SI: v_cvt_u32_f32_e32
-; SI: v_cvt_u32_f32_e32
-; SI: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
+; GCN: v_cvt_u32_f32_e32
 
 define void @fp_to_uint_v4f32_to_v4i32(<4 x i32> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
   %value = load <4 x float>, <4 x float> addrspace(1) * %in
   %result = fptoui <4 x float> %value to <4 x i32>
   store <4 x i32> %result, <4 x i32> addrspace(1)* %out
   ret void
 }
 

 ; EG-DAG: SETGT_INT
 ; EG-DAG: XOR_INT
 ; EG-DAG: XOR_INT
 ; EG: SUB_INT
 ; EG-DAG: SUB_INT
 ; EG-DAG: CNDE_INT
 ; EG-DAG: CNDE_INT
 
-; SI: s_endpgm
+; GCN: s_endpgm
 define void @fp_to_uint_f32_to_i64(i64 addrspace(1)* %out, float %x) {
   %conv = fptoui float %x to i64
   store i64 %conv, i64 addrspace(1)* %out
   ret void
 }
 
 ; FUNC: {{^}}fp_to_uint_v2f32_to_v2i64:
 ; EG-DAG: AND_INT

 ; EG-DAG: SETGT_INT
 ; EG-DAG: XOR_INT
 ; EG-DAG: XOR_INT
 ; EG-DAG: SUB_INT
 ; EG-DAG: SUB_INT
 ; EG-DAG: CNDE_INT
 ; EG-DAG: CNDE_INT
 
-; SI: s_endpgm
+; GCN: s_endpgm
 define void @fp_to_uint_v2f32_to_v2i64(<2 x i64> addrspace(1)* %out, <2 x float> %x) {
   %conv = fptoui <2 x float> %x to <2 x i64>
   store <2 x i64> %conv, <2 x i64> addrspace(1)* %out
   ret void
 }
 
 ; FUNC: {{^}}fp_to_uint_v4f32_to_v4i64:
 ; EG-DAG: AND_INT

 ; EG-DAG: SETGT_INT
 ; EG-DAG: XOR_INT
 ; EG-DAG: XOR_INT
 ; EG-DAG: SUB_INT
 ; EG-DAG: SUB_INT
 ; EG-DAG: CNDE_INT
 ; EG-DAG: CNDE_INT
 
-; SI: s_endpgm
+; GCN: s_endpgm
 define void @fp_to_uint_v4f32_to_v4i64(<4 x i64> addrspace(1)* %out, <4 x float> %x) {
   %conv = fptoui <4 x float> %x to <4 x i64>
   store <4 x i64> %conv, <4 x i64> addrspace(1)* %out
   ret void
 }
 
 
 ; FUNC-LABEL: {{^}}fp_to_uint_f32_to_i1:
-; SI: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, s{{[0-9]+}}
+; GCN: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, s{{[0-9]+}}
 
 ; EG: AND_INT
 ; EG: SETE_DX10 {{[*]?}} T{{[0-9]+}}.{{[XYZW]}}, KC0[2].Z, 1.0,
 define void @fp_to_uint_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
   %conv = fptoui float %in to i1
   store i1 %conv, i1 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}fp_to_uint_fabs_f32_to_i1:
-; SI: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, |s{{[0-9]+}}|
+; GCN: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, |s{{[0-9]+}}|
 define void @fp_to_uint_fabs_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
   %in.fabs = call float @llvm.fabs.f32(float %in)
   %conv = fptoui float %in.fabs to i1
   store i1 %conv, i1 addrspace(1)* %out
   ret void
 }
 
+; FUNC-LABEL: {{^}}fp_to_uint_i16:
+; The reason different instructions are used on SI and VI is because for
+; SI fp_to_uint is legalized by the type legalizer and for VI it is
+; legalized by the dag legalizer and they legalize fp_to_uint differently.
+; SI: v_cvt_u32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
+; VI: v_cvt_i32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
+; GCN: buffer_store_short [[VAL]]
+define void @fp_to_uint_i16(i16 addrspace(1)* %out, float %in) #0 {
+  %uint = fptoui float %in to i16
+  store i16 %uint, i16 addrspace(1)* %out
+  ret void
+}
+
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }