[X86] LLVM >= 18 folding bitcast -> and -> bitcast to @llvm.fabs.f16 generates call to __extendhfsf2

[overmighty]

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:

• [X86] LLVM >= 17 generates call to __extendhfsf2 for bitcast -> and -> bitcast #104914
• [libc][math][c23] Optimize fabsf16 on x86 with Clang #104869 (comment)

cc @arsenm @lntue

[arsenm]

Forming fabs is correct, the codegen is not. Promotion should bitcast to integer

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: OverMighty (overmighty)

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&lt;_Float16, 0&gt;(_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&lt;_Float16, 0&gt;(_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:

• [X86] LLVM >= 17 generates call to __extendhfsf2 for bitcast -> and -> bitcast #104914
• [libc][math][c23] Optimize fabsf16 on x86 with Clang #104869 (comment)

cc @arsenm @lntue

[arsenm]

The same bug exists for fneg and copysign, and this is repeated in the type legalizer and promotion action

[arsenm]

Attaching partial patch. It's incomplete because it asserts on targets with legal f16, but no legal i16. I don't have the patience to fight with SelectionDAG's insistence that there can be no illegal types after type legalization

0001-DAG-Do-not-use-FP-conversions-to-promote-fabs.patch

[beetrees]

This isn't x86 specific; it affects FABS and FNEG for any target that uses TypeSoftPromoteHalf. The issue is that the lowering uses the generic SoftPromoteHalfRes_UnaryOp method. The resulting lowering is actually a miscompilation as __extendhfsf2 and __truncsfhf2 quieten signalling NaNs, whereas FABS and FNEG are defined as only affecting the sign bit (FCOPYSIGN doesn't have this issue as its lowering in SoftPromoteHalfRes_FCOPYSIGN doesn't convert to and from a f32).