[clang] [llvm] [LoongArch] [CodeGen] Add options for Clang to generate LoongArch-specific frecipe & frsqrte instructions (PR #109917)

Lu Weining via llvm-commits llvm-commits at lists.llvm.org
Thu Oct 10 20:22:10 PDT 2024


================
@@ -0,0 +1,396 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch32 --mattr=+f,-d,-frecipe < %s | FileCheck %s --check-prefix=LA32F
+; RUN: llc --mtriple=loongarch32 --mattr=+f,-d,+frecipe < %s | FileCheck %s --check-prefix=LA32F-FRECIPE
+; RUN: llc --mtriple=loongarch64 --mattr=+d,-frecipe < %s | FileCheck %s --check-prefix=LA64D
+; RUN: llc --mtriple=loongarch64 --mattr=+d,+frecipe < %s | FileCheck %s --check-prefix=LA64D-FRECIPE
+
+
+declare float @llvm.sqrt.f32(float)
+declare double @llvm.sqrt.f64(double)
+
+define float @frsqrt_f32(float %a) nounwind {
+; LA32F-LABEL: frsqrt_f32:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    frsqrt.s $fa0, $fa0
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: frsqrt_f32:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    frsqrte.s	$fa1, $fa0
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI0_0)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa2, $a0, %pc_lo12(.LCPI0_0)
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI0_1)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa3, $a0, %pc_lo12(.LCPI0_1)
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa0, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    fmadd.s	$fa0, $fa0, $fa1, $fa2
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa1, $fa3
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa0, $fa1, $fa0
+; LA32F-FRECIPE-NEXT:    ret
+;
+; LA64D-LABEL: frsqrt_f32:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    frsqrt.s $fa0, $fa0
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: frsqrt_f32:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frsqrte.s	$fa1, $fa0
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI0_0)
+; LA64D-FRECIPE-NEXT:    fld.s	$fa2, $a0, %pc_lo12(.LCPI0_0)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI0_1)
+; LA64D-FRECIPE-NEXT:    fld.s	$fa3, $a0, %pc_lo12(.LCPI0_1)
+; LA64D-FRECIPE-NEXT:    fmul.s	$fa1, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.s	$fa0, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.s	$fa0, $fa0, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.s	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.s	$fa0, $fa1, $fa0
+; LA64D-FRECIPE-NEXT:    ret
+
+  %1 = call fast float @llvm.sqrt.f32(float %a)
+  %2 = fdiv fast float 1.0, %1
+  ret float %2
+}
+
+define double @frsqrt_f64(double %a) nounwind {
+; LA32F-LABEL: frsqrt_f64:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    addi.w	$sp, $sp, -16
+; LA32F-NEXT:    st.w	$ra, $sp, 12 
+; LA32F-NEXT:    bl	%plt(sqrt)
+; LA32F-NEXT:    move	$a2, $a0
+; LA32F-NEXT:    move	$a3, $a1
+; LA32F-NEXT:    lu12i.w	$a1, 261888
+; LA32F-NEXT:    move	$a0, $zero
+; LA32F-NEXT:    bl	%plt(__divdf3)
+; LA32F-NEXT:    ld.w	$ra, $sp, 12  
+; LA32F-NEXT:    addi.w	$sp, $sp, 16
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: frsqrt_f64:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    addi.w	$sp, $sp, -16
+; LA32F-FRECIPE-NEXT:    st.w	$ra, $sp, 12                    # 4-byte Folded Spill
+; LA32F-FRECIPE-NEXT:    bl	%plt(sqrt)
+; LA32F-FRECIPE-NEXT:    move	$a2, $a0
+; LA32F-FRECIPE-NEXT:    move	$a3, $a1
+; LA32F-FRECIPE-NEXT:    lu12i.w	$a1, 261888
+; LA32F-FRECIPE-NEXT:    move	$a0, $zero
+; LA32F-FRECIPE-NEXT:    bl	%plt(__divdf3)
+; LA32F-FRECIPE-NEXT:    ld.w	$ra, $sp, 12                    # 4-byte Folded Reload
+; LA32F-FRECIPE-NEXT:    addi.w	$sp, $sp, 16
+; LA32F-FRECIPE-NEXT:    ret
+;
+; LA64D-LABEL: frsqrt_f64:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    frsqrt.d $fa0, $fa0
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: frsqrt_f64:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frsqrte.d	$fa1, $fa0
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI1_0)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa2, $a0, %pc_lo12(.LCPI1_0)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI1_1)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa3, $a0, %pc_lo12(.LCPI1_1)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa4, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa0, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa0, $fa0, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa0, $fa1, $fa0
+; LA64D-FRECIPE-NEXT:    ret
+  %1 = call fast double @llvm.sqrt.f64(double %a)
+  %2 = fdiv fast double 1.0, %1
+  ret double %2
+}
+
+define double @sqrt_simplify_before_recip_3_uses_f64(double %x, ptr %p1, ptr %p2) nounwind {
+; LA64D-LABEL: sqrt_simplify_before_recip_3_uses_f64:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI2_0)
+; LA64D-NEXT:    fld.d	$fa2, $a2, %pc_lo12(.LCPI2_0)
+; LA64D-NEXT:    fsqrt.d	$fa1, $fa0
+; LA64D-NEXT:    frsqrt.d	$fa0, $fa0
+; LA64D-NEXT:    fdiv.d	$fa2, $fa2, $fa1
+; LA64D-NEXT:    fst.d	$fa0, $a0, 0
+; LA64D-NEXT:    fst.d	$fa2, $a1, 0
+; LA64D-NEXT:    fmov.d	$fa0, $fa1
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: sqrt_simplify_before_recip_3_uses_f64:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frsqrte.d	$fa1, $fa0
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI2_0)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa2, $a2, %pc_lo12(.LCPI2_0)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI2_1)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa3, $a2, %pc_lo12(.LCPI2_1)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa4, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI2_2)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa5, $a2, %pc_lo12(.LCPI2_2)
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa2, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa2, $fa1, $fa5
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa0, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fst.d	$fa1, $a0, 0
+; LA64D-FRECIPE-NEXT:    fst.d	$fa2, $a1, 0
+; LA64D-FRECIPE-NEXT:    ret
+  %sqrt = tail call fast double @llvm.sqrt.f64(double %x)
+  %rsqrt = fdiv fast double 1.0, %sqrt
+  %r = fdiv fast double 42.0, %sqrt
+  %sqrt_fast = fdiv fast double %x, %sqrt
+  store double %rsqrt, ptr %p1, align 8
+  store double %r, ptr %p2, align 8
+  ret double %sqrt_fast
+}
+
+
+define double @sqrt_simplify_before_recip_3_uses_order_f64(double %x, ptr %p1, ptr %p2) nounwind {
+; LA64D-LABEL: sqrt_simplify_before_recip_3_uses_order_f64:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_0)
+; LA64D-NEXT:    fld.d	$fa1, $a2, %pc_lo12(.LCPI3_0)
+; LA64D-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_1)
+; LA64D-NEXT:    fld.d	$fa2, $a2, %pc_lo12(.LCPI3_1)
+; LA64D-NEXT:    fsqrt.d	$fa0, $fa0
+; LA64D-NEXT:    fdiv.d	$fa1, $fa1, $fa0
+; LA64D-NEXT:    fdiv.d	$fa2, $fa2, $fa0
+; LA64D-NEXT:    fst.d	$fa1, $a0, 0
+; LA64D-NEXT:    fst.d	$fa2, $a1, 0
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: sqrt_simplify_before_recip_3_uses_order_f64:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frsqrte.d	$fa1, $fa0
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_0)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa2, $a2, %pc_lo12(.LCPI3_0)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_1)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa3, $a2, %pc_lo12(.LCPI3_1)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa4, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa2, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_2)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa3, $a2, %pc_lo12(.LCPI3_2)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI3_3)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa4, $a2, %pc_lo12(.LCPI3_3)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa0, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa2, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fst.d	$fa2, $a0, 0
+; LA64D-FRECIPE-NEXT:    fst.d	$fa1, $a1, 0
+; LA64D-FRECIPE-NEXT:    ret
+  %sqrt = tail call fast double @llvm.sqrt.f64(double %x)
+  %sqrt_fast = fdiv fast double %x, %sqrt
+  %r1 = fdiv fast double 42.0, %sqrt
+  %r2 = fdiv fast double 43.0, %sqrt
+  store double %r1, ptr %p1, align 8
+  store double %r2, ptr %p2, align 8
+  ret double %sqrt_fast
+}
+
+define double @sqrt_simplify_before_recip_4_uses_f64(double %x, ptr %p1, ptr %p2, ptr %p3) nounwind {
+; LA64D-LABEL: sqrt_simplify_before_recip_4_uses_f64:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_0)
+; LA64D-NEXT:    fld.d	$fa2, $a3, %pc_lo12(.LCPI4_0)
+; LA64D-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_1)
+; LA64D-NEXT:    fld.d	$fa3, $a3, %pc_lo12(.LCPI4_1)
+; LA64D-NEXT:    fsqrt.d	$fa1, $fa0
+; LA64D-NEXT:    frsqrt.d	$fa0, $fa0
+; LA64D-NEXT:    fdiv.d	$fa2, $fa2, $fa1
+; LA64D-NEXT:    fdiv.d	$fa3, $fa3, $fa1
+; LA64D-NEXT:    fst.d	$fa0, $a0, 0
+; LA64D-NEXT:    fst.d	$fa2, $a1, 0
+; LA64D-NEXT:    fst.d	$fa3, $a2, 0
+; LA64D-NEXT:    fmov.d	$fa0, $fa1
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: sqrt_simplify_before_recip_4_uses_f64:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frsqrte.d	$fa1, $fa0
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_0)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa2, $a3, %pc_lo12(.LCPI4_0)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_1)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa3, $a3, %pc_lo12(.LCPI4_1)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa4, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa4, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa2, $fa4, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_2)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa4, $a3, %pc_lo12(.LCPI4_2)
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI4_3)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa5, $a3, %pc_lo12(.LCPI4_3)
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa1, $fa1, $fa2
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa2, $fa1, $fa4
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa3, $fa1, $fa5
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa0, $fa0, $fa1
+; LA64D-FRECIPE-NEXT:    fst.d	$fa1, $a0, 0
+; LA64D-FRECIPE-NEXT:    fst.d	$fa2, $a1, 0
+; LA64D-FRECIPE-NEXT:    fst.d	$fa3, $a2, 0
+; LA64D-FRECIPE-NEXT:    ret
+  %sqrt = tail call fast double @llvm.sqrt.f64(double %x)
+  %rsqrt = fdiv fast double 1.0, %sqrt
+  %r1 = fdiv fast double 42.0, %sqrt
+  %r2 = fdiv fast double 43.0, %sqrt
+  %sqrt_fast = fdiv fast double %x, %sqrt
+  store double %rsqrt, ptr %p1, align 8
+  store double %r1, ptr %p2, align 8
+  store double %r2, ptr %p3, align 8
+  ret double %sqrt_fast
+}
+
+define float @sqrt_simplify_before_recip_3_uses_f32(float %x, ptr %p1, ptr %p2) nounwind {
+; LA32F-LABEL: sqrt_simplify_before_recip_3_uses_f32:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI5_0)
+; LA32F-NEXT:    fld.s	$fa2, $a2, %pc_lo12(.LCPI5_0)
+; LA32F-NEXT:    fsqrt.s	$fa1, $fa0
+; LA32F-NEXT:    frsqrt.s	$fa0, $fa0
+; LA32F-NEXT:    fdiv.s	$fa2, $fa2, $fa1
+; LA32F-NEXT:    fst.s	$fa0, $a0, 0
+; LA32F-NEXT:    fst.s	$fa2, $a1, 0
+; LA32F-NEXT:    fmov.s	$fa0, $fa1
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: sqrt_simplify_before_recip_3_uses_f32:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    frsqrte.s	$fa1, $fa0
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa2, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI5_0)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa3, $a2, %pc_lo12(.LCPI5_0)
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI5_1)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa4, $a2, %pc_lo12(.LCPI5_1)
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a2, %pc_hi20(.LCPI5_2)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa5, $a2, %pc_lo12(.LCPI5_2)
+; LA32F-FRECIPE-NEXT:    fmadd.s	$fa2, $fa2, $fa1, $fa3
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa1, $fa4
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa1, $fa2
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa2, $fa1, $fa5
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa0, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    fst.s	$fa1, $a0, 0
+; LA32F-FRECIPE-NEXT:    fst.s	$fa2, $a1, 0
+; LA32F-FRECIPE-NEXT:    ret
+  %sqrt = tail call fast float @llvm.sqrt.f32(float %x)
+  %rsqrt = fdiv fast float 1.0, %sqrt
+  %r = fdiv fast float 42.0, %sqrt
+  %sqrt_fast = fdiv fast float %x, %sqrt
+  store float %rsqrt, ptr %p1, align 8
+  store float %r, ptr %p2, align 8
+  ret float %sqrt_fast
+}
+
+define float @sqrt_simplify_before_recip_4_uses_f32(float %x, ptr %p1, ptr %p2, ptr %p3) nounwind {
+; LA32F-LABEL: sqrt_simplify_before_recip_4_uses_f32:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_0)
+; LA32F-NEXT:    fld.s	$fa2, $a3, %pc_lo12(.LCPI6_0)
+; LA32F-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_1)
+; LA32F-NEXT:    fld.s	$fa3, $a3, %pc_lo12(.LCPI6_1)
+; LA32F-NEXT:    fsqrt.s	$fa1, $fa0
+; LA32F-NEXT:    frsqrt.s	$fa0, $fa0
+; LA32F-NEXT:    fdiv.s	$fa2, $fa2, $fa1
+; LA32F-NEXT:    fdiv.s	$fa3, $fa3, $fa1
+; LA32F-NEXT:    fst.s	$fa0, $a0, 0
+; LA32F-NEXT:    fst.s	$fa2, $a1, 0
+; LA32F-NEXT:    fst.s	$fa3, $a2, 0
+; LA32F-NEXT:    fmov.s	$fa0, $fa1
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: sqrt_simplify_before_recip_4_uses_f32:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_0)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa1, $a3, %pc_lo12(.LCPI6_0)
+; LA32F-FRECIPE-NEXT:    frsqrte.s	$fa2, $fa0
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa2, $fa0, $fa2
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa3, $fa0, $fa2
+; LA32F-FRECIPE-NEXT:    fmadd.s	$fa1, $fa3, $fa2, $fa1
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_1)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa3, $a3, %pc_lo12(.LCPI6_1)
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_2)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa4, $a3, %pc_lo12(.LCPI6_2)
+; LA32F-FRECIPE-NEXT:    pcalau12i	$a3, %pc_hi20(.LCPI6_3)
+; LA32F-FRECIPE-NEXT:    fld.s	$fa5, $a3, %pc_lo12(.LCPI6_3)
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa2, $fa2, $fa3
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa1, $fa2, $fa1
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa2, $fa1, $fa4
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa3, $fa1, $fa5
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa0, $fa0, $fa1
+; LA32F-FRECIPE-NEXT:    fst.s	$fa1, $a0, 0
+; LA32F-FRECIPE-NEXT:    fst.s	$fa2, $a1, 0
+; LA32F-FRECIPE-NEXT:    fst.s	$fa3, $a2, 0
+; LA32F-FRECIPE-NEXT:    ret
+  %sqrt = tail call fast float @llvm.sqrt.f32(float %x)
----------------
SixWeining wrote:

LA64 check is missing?

https://github.com/llvm/llvm-project/pull/109917


More information about the llvm-commits mailing list