fastmath fminnum.f16 could be better on non avx512_fp16 x86

[gbaraldi]

This was found in JuliaLang/julia#48848. Where LLVM started folding a function into fast fminnum.f16 and that led to a regression.

original IR

; ModuleID = 'min_fast'
source_filename = "min_fast"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

;  @ fastmath.jl:244 within `min_fast`
define half @julia_min_fast_681(half %0, half %1) #0 {
top:
; ┌ @ essentials.jl:575 within `ifelse`
   %2 = fpext half %0 to float
   %3 = fpext half %1 to float
   %4 = fcmp fast olt  float %2, %3
   %5 = select i1 %4, half %0, half %1
; └
  ret half %5
}

attributes #0 = { "frame-pointer"="all" "probe-stack"="inline-asm" }

!llvm.module.flags = !{!0, !1}

!0 = !{i32 2, !"Dwarf Version", i32 4}
!1 = !{i32 2, !"Debug Info Version", i32 3}

new IR

; ModuleID = 'min_fast'
source_filename = "min_fast"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

;  @ fastmath.jl:252 within `min_fast`
define half @julia_min_fast_678(half %0, half %1) #0 {
top:
; ┌ @ essentials.jl:586 within `ifelse`
   %2 = call fast half @llvm.minnum.f16(half %0, half %1)
; └
  ret half %2
}

; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
declare half @llvm.minnum.f16(half, half) #2

attributes #0 = { "frame-pointer"="all" "probe-stack"="inline-asm" }
attributes #2 = { nofree nosync nounwind readnone speculatable willreturn }
!llvm.module.flags = !{!0, !1}

!0 = !{i32 2, !"Dwarf Version", i32 4}
!1 = !{i32 2, !"Debug Info Version", i32 3}

Old assembly

	.text
	.file	"min_fast"
	.globl	julia_min_fast_681              # -- Begin function julia_min_fast_681
	.p2align	4, 0x90
	.type	julia_min_fast_681,@function
julia_min_fast_681:                     # @julia_min_fast_681
	.cfi_startproc
# %bb.0:                                # %top
	pushq	%rbp
	.cfi_def_cfa_offset 16
	.cfi_offset %rbp, -16
	movq	%rsp, %rbp
	.cfi_def_cfa_register %rbp
	vpextrw	$0, %xmm1, %eax
	vpextrw	$0, %xmm0, %ecx
	movzwl	%cx, %ecx
	vmovd	%ecx, %xmm0
	vcvtph2ps	%xmm0, %xmm0
	movzwl	%ax, %eax
	vmovd	%eax, %xmm1
	vcvtph2ps	%xmm1, %xmm1
	vucomiss	%xmm1, %xmm0
	cmovbl	%ecx, %eax
	vpinsrw	$0, %eax, %xmm0, %xmm0
	popq	%rbp
	.cfi_def_cfa %rsp, 8
	retq
.Lfunc_end0:
	.size	julia_min_fast_681, .Lfunc_end0-julia_min_fast_681
	.cfi_endproc
                                        # -- End function
	.section	".note.GNU-stack","",@progbits

New assembly

	.text
	.file	"min_fast"
	.globl	julia_min_fast_678              # -- Begin function julia_min_fast_678
	.p2align	4, 0x90
	.type	julia_min_fast_678,@function
julia_min_fast_678:                     # @julia_min_fast_678
	.cfi_startproc
# %bb.0:                                # %top
	pushq	%rbp
	.cfi_def_cfa_offset 16
	.cfi_offset %rbp, -16
	movq	%rsp, %rbp
	.cfi_def_cfa_register %rbp
	vpextrw	$0, %xmm0, %eax
	vpextrw	$0, %xmm1, %ecx
	movzwl	%cx, %ecx
	vmovd	%ecx, %xmm0
	vcvtph2ps	%xmm0, %xmm0
	movzwl	%ax, %eax
	vmovd	%eax, %xmm1
	vcvtph2ps	%xmm1, %xmm1
	vcmpltss	%xmm0, %xmm1, %xmm2
	vblendvps	%xmm2, %xmm1, %xmm0, %xmm0
	vcvtps2ph	$4, %xmm0, %xmm0
	vmovd	%xmm0, %eax
	vpinsrw	$0, %eax, %xmm0, %xmm0
	popq	%rbp
	.cfi_def_cfa %rsp, 8
	retq
.Lfunc_end0:
	.size	julia_min_fast_678, .Lfunc_end0-julia_min_fast_678
	.cfi_endproc
                                        # -- End function
	.section	".note.GNU-stack","",@progbits

This is with llc-15 -mcpu alderlake

[llvmbot]

@llvm/issue-subscribers-backend-x86