[X86] LLVM >= 18 folding bitcast -> and -> bitcast to @llvm.fabs.f16 generates call to __extendhfsf2 · Issue #104915 · llvm/llvm-project (original) (raw)
https://godbolt.org/z/on6Pj6Tsc
Starting from 5c0da58, InstCombine transforms this IR:
define linkonce_odr hidden noundef half @_Float16 __llvm_libc_20_0_0_git::fputil::abs<_Float16, 0>(_Float16)(half noundef %x) local_unnamed_addr #1 comdat { entry: %0 = bitcast half %x to i16 %1 = and i16 %0, 32767 %2 = bitcast i16 %1 to half ret half %2 }
into:
define linkonce_odr hidden noundef half @_Float16 __llvm_libc_20_0_0_git::fputil::abs<_Float16, 0>(_Float16)(half noundef %x) local_unnamed_addr #1 comdat { entry: %0 = call half @llvm.fabs.f16(half %x) ret half %0 }
On x86, when AVX-512 FP16 is not available, this generates something like:
.LCPI0_0: .long 0x7fffffff __llvm_libc_20_0_0_git::fabsf16(_Float16): push rbp mov rbp, rsp call __extendhfsf2@PLT vbroadcastss xmm1, dword ptr [rip + .LCPI0_0] vandps xmm0, xmm0, xmm1 call __truncsfhf2@PLT pop rbp ret
whereas with LLVM 16, bitcast -> and -> bitcast generates something like:
__llvm_libc_20_0_0_git::fabsf16(_Float16): push rbp mov rbp, rsp vpextrw eax, xmm0, 0 and eax, 32767 vpinsrw xmm0, xmm0, eax, 0 pop rbp ret
Related: