[llvm] [AMDGPU][True16][CodeGen] fp conversion in true/fake16 format (PR #101678)

[Matt Arsenault via llvm-commits]

================
@@ -1131,25 +1131,42 @@ multiclass f16_fp_Pats<Instruction cvt_f16_f32_inst_e64, Instruction cvt_f32_f16
     (cvt_f32_f16_inst_e64 SRCMODS.NEG, $src0)
   >;
 
+  // fp_to_fp16 patterns
   def : GCNPat <
-    (f64 (any_fpextend f16:$src)),
-    (V_CVT_F64_F32_e32 (cvt_f32_f16_inst_e64 SRCMODS.NONE, $src))
+    (i32 (AMDGPUfp_to_f16 (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
+    (cvt_f16_f32_inst_e64 $src0_modifiers, f32:$src0)
   >;
 
-  // fp_to_fp16 patterns
+  // This is only used on targets without half support
+  // TODO: Introduce strict variant of AMDGPUfp_to_f16 and share custom lowering
   def : GCNPat <
-    (i32 (AMDGPUfp_to_f16 (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
+    (i32 (strict_fp_to_f16 (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
     (cvt_f16_f32_inst_e64 $src0_modifiers, f32:$src0)
   >;
+}
+
+let SubtargetPredicate = NotHasTrue16BitInsts in
+defm : f16_to_fp_Pats<V_CVT_F16_F32_e64, V_CVT_F32_F16_e64>;
+
+let SubtargetPredicate = UseFakeTrue16Insts in
+defm : f16_to_fp_Pats<V_CVT_F16_F32_fake16_e64, V_CVT_F32_F16_fake16_e64>;
+
+multiclass f16_fp_Pats<Instruction cvt_f16_f32_inst_e64,
+                       Instruction cvt_f32_f16_inst_e64,
+                       RegOrImmOperand VSrc> {
+  def : GCNPat <
+    (f64 (any_fpextend f16:$src)),
+    (V_CVT_F64_F32_e32 (cvt_f32_f16_inst_e64 SRCMODS.NONE, $src))
+  >;
----------------
arsenm wrote:

This is wrong, you can't do this conversion in 2 steps. f64->f16 needs to go through custom lowering 

https://github.com/llvm/llvm-project/pull/101678