[llvm] [Xtensa] Implement Xtensa Floating Point Option. (PR #136086)

[Andrei Safronov via llvm-commits]

================
@@ -815,6 +936,21 @@ SDValue XtensaTargetLowering::LowerImmediate(SDValue Op,
   return Op;
 }
 
+SDValue XtensaTargetLowering::LowerImmediateFP(SDValue Op,
----------------
andreisfr wrote:

  Thank you for comment. I've made an experiment with 2 simple tests using i32 and f32 constants. For f32 immediate I used lowering by expansion. And I've got such results:
  
  Integer constant:
  "
    IR code
	
	%1 = sub i32 %c, 132132
	
	it is translated to
    
	t12: i32 = XtensaISD::PCREL_WRAPPER TargetConstantPool:i32<i32 -132132> 0
	t9: i32 = add t2, t12
    
	and then such code generated
	
    t12: i32 = L32R TargetConstantPool:i32<i32 -132132> 0
    t9: i32 = ADD t2, t12
	
	This code is correctly loads i32 constant from ConstantPool using L32R instruction
 "
 
 Floating point  constant:
 "
    IR code
	
	%c_ = fadd float 0.0, %c
	
	it is translated to

    t22: i32 = XtensaISD::PCREL_WRAPPER TargetConstantPool:i32<float 0.000000e+00> 0
    t20: f32,ch = load<(load (s32) from constant-pool)> t0, t22, undef:i32
 
	and then such code generated
	
    t22: i32 = L32R TargetConstantPool:i32<float 0.000000e+00> 0
    t20: f32,ch = LSI<Mem:(load (s32) from constant-pool)> t22, TargetConstant:i32<0>, t0 
	
	This code loads i32 constant from ConstantPool using L32R instruction and then consider this constant as new address to 
	load another constant using LSI instruction. This is incorrect.
 "
 
 So, it seems that if we need to just use expasnsion of the f32 immediate, then we need to make some kind of refactoring of the constant pool implementation.

https://github.com/llvm/llvm-project/pull/136086