[AMDGPU] Add instruction selection for i1 to f16 conversion

[critson]

Summary

This is required for GPUs with 16 bit instructions where f16 is a
legal register type and hence int_to_fp i1 to f16 is not lowered
by legalizing.

Change-Id: Ie4c0fd6ced7cf10ad612023c6879724d9ded5851

[nhaehnle]

What about sitofp?

Why the detour via V_CVT_F16_F32 instead of selecting an fp16 1.0 constant directly?

[critson]

Add support and test for sint_to_fp.

[critson]

In D52018#1232860, @nhaehnle wrote:

Why the detour via V_CVT_F16_F32 instead of selecting an fp16 1.0 constant directly?

It is not possible select an fp16 constant with V_CNDMASK_B32.
In principle the VOP2 version could be used to select an inline literal, but only when the i1 is vcc.

[nhaehnle]

In D52018#1235002, @critson wrote:

In D52018#1232860, @nhaehnle wrote:

Why the detour via V_CVT_F16_F32 instead of selecting an fp16 1.0 constant directly?

It is not possible select an fp16 constant with V_CNDMASK_B32.
In principle the VOP2 version could be used to select an inline literal, but only when the i1 is vcc.

Oh yeah, that makes sense.

Thanks for the sitofp change, LGTM.

llvm/trunk/lib/Target/AMDGPU/SIInstructions.td

 >;
 
 def : GCNPat <
   (i1 (xor i1:$src0, i1:$src1)),
   (S_XOR_B64 $src0, $src1)
 >;
 
 def : GCNPat <
+  (f16 (sint_to_fp i1:$src)),
+  (V_CVT_F16_F32_e32 (V_CNDMASK_B32_e64 (i32 0), (i32 CONST.FP32_NEG_ONE), $src))
+>;
+
+def : GCNPat <
+  (f16 (uint_to_fp i1:$src)),
+  (V_CVT_F16_F32_e32 (V_CNDMASK_B32_e64 (i32 0), (i32 CONST.FP32_ONE), $src))
+>;
+
+def : GCNPat <
   (f32 (sint_to_fp i1:$src)),
   (V_CNDMASK_B32_e64 (i32 0), (i32 CONST.FP32_NEG_ONE), $src)
 >;
 
 def : GCNPat <
   (f32 (uint_to_fp i1:$src)),
   (V_CNDMASK_B32_e64 (i32 0), (i32 CONST.FP32_ONE), $src)
 >;

llvm/trunk/test/CodeGen/AMDGPU/sitofp.f16.ll

define amdgpu_kernel void @sitofp_v2i32_to_v2f16(
     <2 x i32> addrspace(1)* %a) {
 entry:
   %a.val = load <2 x i32>, <2 x i32> addrspace(1)* %a
   %r.val = sitofp <2 x i32> %a.val to <2 x half>
   store <2 x half> %r.val, <2 x half> addrspace(1)* %r
   ret void
 }
 
+; FUNC-LABEL: {{^}}s_sint_to_fp_i1_to_f16:
+; GCN-DAG: v_cmp_le_f32_e32 [[CMP0:vcc]], 1.0, {{v[0-9]+}}
+; GCN-DAG: v_cmp_le_f32_e64 [[CMP1:s\[[0-9]+:[0-9]+\]]], 0, {{v[0-9]+}}
+; GCN: s_xor_b64 [[R_CMP:s\[[0-9]+:[0-9]+\]]], [[CMP1]], [[CMP0]]
+; GCN: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1.0, [[R_CMP]]
+; GCN-NEXT: v_cvt_f16_f32_e32 [[R_F16:v[0-9]+]], [[RESULT]]
+; GCN: buffer_store_short
+; GCN: s_endpgm
+define amdgpu_kernel void @s_sint_to_fp_i1_to_f16(half addrspace(1)* %out, float addrspace(1)* %in0, float addrspace(1)* %in1) {
+  %a = load float, float addrspace(1) * %in0
+  %b = load float, float addrspace(1) * %in1
+  %acmp = fcmp oge float %a, 0.000000e+00
+  %bcmp = fcmp oge float %b, 1.000000e+00
+  %result = xor i1 %acmp, %bcmp
+  %fp = sitofp i1 %result to half
+  store half %fp, half addrspace(1)* %out
+  ret void
+}
+
 ; v2f16 = sitofp v2i64 is in sint_to_fp.i64.ll

llvm/trunk/test/CodeGen/AMDGPU/uitofp.f16.ll

define amdgpu_kernel void @uitofp_v2i32_to_v2f16(
     <2 x i32> addrspace(1)* %a) {
 entry:
   %a.val = load <2 x i32>, <2 x i32> addrspace(1)* %a
   %r.val = uitofp <2 x i32> %a.val to <2 x half>
   store <2 x half> %r.val, <2 x half> addrspace(1)* %r
   ret void
 }
 
+; FUNC-LABEL: {{^}}s_uint_to_fp_i1_to_f16:
+; GCN-DAG: v_cmp_le_f32_e32 [[CMP0:vcc]], 1.0, {{v[0-9]+}}
+; GCN-DAG: v_cmp_le_f32_e64 [[CMP1:s\[[0-9]+:[0-9]+\]]], 0, {{v[0-9]+}}
+; GCN: s_xor_b64 [[R_CMP:s\[[0-9]+:[0-9]+\]]], [[CMP1]], [[CMP0]]
+; GCN: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1.0, [[R_CMP]]
+; GCN-NEXT: v_cvt_f16_f32_e32 [[R_F16:v[0-9]+]], [[RESULT]]
+; GCN: buffer_store_short
+; GCN: s_endpgm
+define amdgpu_kernel void @s_uint_to_fp_i1_to_f16(half addrspace(1)* %out, float addrspace(1)* %in0, float addrspace(1)* %in1) {
+  %a = load float, float addrspace(1) * %in0
+  %b = load float, float addrspace(1) * %in1
+  %acmp = fcmp oge float %a, 0.000000e+00
+  %bcmp = fcmp oge float %b, 1.000000e+00
+  %result = xor i1 %acmp, %bcmp
+  %fp = uitofp i1 %result to half
+  store half %fp, half addrspace(1)* %out
+  ret void
+}
+
 ; f16 = uitofp i64 is in uint_to_fp.i64.ll