llvm-project

af31bd52 - [X86] Lower bf16->f32/f64 fpext without a GPR round-trip (#218359)

[X86] Lower bf16->f32/f64 fpext without a GPR round-trip (#218359) `fpext bfloat to float/double` currently lowers `ISD::BF16_TO_FP` via the generic target-independent expansion (extend to i32, shl 16, bitcast to f32), which always forces the value out of its XMM register into a GPR and back. Fixes #155270 Add a custom `BF16_TO_FP` lowering (mirroring the existing `FP_TO_BF16` function) for f32/f64 results that does the shift in-register instead. Only on targets with SSE2 or higher otherwise still falls back to the existing generic expansion. The updated lowering then for the example ```llvm define float @src32(bfloat %a0) { %res = fpext bfloat %a0 to float ret float %res } define double @src64(bfloat %a0) { %res = fpext bfloat %a0 to double ret double %res } ``` generates ```asm src32: pslld $16, %xmm0 retq src64: pslld $16, %xmm0 cvtss2sd %xmm0, %xmm0 retq ``` instead of the previous roundtrip ```asm src32: pextrw $0, %xmm0, %eax shll $16, %eax movd %eax, %xmm0 retq src64: pextrw $0, %xmm0, %eax shll $16, %eax movd %eax, %xmm0 cvtss2sd %xmm0, %xmm0 retq ``` AI used for generating and cleaning up comments and reviewing the code.