124 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -O3 -mtriple=riscv64 -mattr=+f,+d < %s | FileCheck %s3 4; Make sure WriteFRM is not hoisted out of loop due to dead implicit-def.5 6define double @foo(double %0, double %1, i64 %n) strictfp {7; CHECK-LABEL: foo:8; CHECK: # %bb.0: # %entry9; CHECK-NEXT: fmv.d.x fa5, zero10; CHECK-NEXT: .LBB0_1: # %loop11; CHECK-NEXT: # =>This Inner Loop Header: Depth=112; CHECK-NEXT: fsrmi 313; CHECK-NEXT: fadd.d fa5, fa5, fa014; CHECK-NEXT: addi a0, a0, -115; CHECK-NEXT: fsrmi 016; CHECK-NEXT: fadd.d fa5, fa5, fa117; CHECK-NEXT: beqz a0, .LBB0_118; CHECK-NEXT: # %bb.2: # %exit19; CHECK-NEXT: fmv.d fa0, fa520; CHECK-NEXT: ret21entry:22 br label %loop23loop:24 %cnt = phi i64 [0, %entry], [%cnt_inc, %loop]25 %acc = phi double [0.0, %entry], [%f2, %loop]26 call void @llvm.set.rounding(i32 2) strictfp27 %f1 = call double @llvm.experimental.constrained.fadd.f64(double %acc, double %0, metadata !"round.dynamic", metadata !"fpexcept.ignore") strictfp28 call void @llvm.set.rounding(i32 1) strictfp29 %f2 = call double @llvm.experimental.constrained.fadd.f64(double %f1, double %1, metadata !"round.dynamic", metadata !"fpexcept.ignore") strictfp30 %cnt_inc = add i64 %cnt, 131 %cond = icmp eq i64 %cnt_inc, %n32 br i1 %cond, label %loop, label %exit33exit:34 ret double %f235}36 37declare double @baz(double, double)38 39define double @bar(double %0, double %1, i64 %n) strictfp {40; CHECK-LABEL: bar:41; CHECK: # %bb.0: # %entry42; CHECK-NEXT: addi sp, sp, -3243; CHECK-NEXT: .cfi_def_cfa_offset 3244; CHECK-NEXT: sd ra, 24(sp) # 8-byte Folded Spill45; CHECK-NEXT: sd s0, 16(sp) # 8-byte Folded Spill46; CHECK-NEXT: fsd fs0, 8(sp) # 8-byte Folded Spill47; CHECK-NEXT: fsd fs1, 0(sp) # 8-byte Folded Spill48; CHECK-NEXT: .cfi_offset ra, -849; CHECK-NEXT: .cfi_offset s0, -1650; CHECK-NEXT: .cfi_offset fs0, -2451; CHECK-NEXT: .cfi_offset fs1, -3252; CHECK-NEXT: mv s0, a053; CHECK-NEXT: fmv.d fs0, fa154; CHECK-NEXT: fmv.d fs1, fa055; CHECK-NEXT: fmv.d.x fa0, zero56; CHECK-NEXT: .LBB1_1: # %loop57; CHECK-NEXT: # =>This Inner Loop Header: Depth=158; CHECK-NEXT: fsrmi 359; CHECK-NEXT: fmv.d fa1, fs160; CHECK-NEXT: call baz61; CHECK-NEXT: fsrmi 062; CHECK-NEXT: fmv.d fa1, fs063; CHECK-NEXT: call baz64; CHECK-NEXT: addi s0, s0, -165; CHECK-NEXT: beqz s0, .LBB1_166; CHECK-NEXT: # %bb.2: # %exit67; CHECK-NEXT: ld ra, 24(sp) # 8-byte Folded Reload68; CHECK-NEXT: ld s0, 16(sp) # 8-byte Folded Reload69; CHECK-NEXT: fld fs0, 8(sp) # 8-byte Folded Reload70; CHECK-NEXT: fld fs1, 0(sp) # 8-byte Folded Reload71; CHECK-NEXT: .cfi_restore ra72; CHECK-NEXT: .cfi_restore s073; CHECK-NEXT: .cfi_restore fs074; CHECK-NEXT: .cfi_restore fs175; CHECK-NEXT: addi sp, sp, 3276; CHECK-NEXT: .cfi_def_cfa_offset 077; CHECK-NEXT: ret78entry:79 br label %loop80loop:81 %cnt = phi i64 [0, %entry], [%cnt_inc, %loop]82 %acc = phi double [0.0, %entry], [%f2, %loop]83 call void @llvm.set.rounding(i32 2) strictfp84 %f1 = call double @baz(double %acc, double %0) strictfp85 call void @llvm.set.rounding(i32 1) strictfp86 %f2 = call double @baz(double %f1, double %1) strictfp87 %cnt_inc = add i64 %cnt, 188 %cond = icmp eq i64 %cnt_inc, %n89 br i1 %cond, label %loop, label %exit90exit:91 ret double %f292}93 94define double @foo2(double %0, double %1, i64 %n, i64 %fcsr) strictfp {95; CHECK-LABEL: foo2:96; CHECK: # %bb.0: # %entry97; CHECK-NEXT: fmv.d.x fa5, zero98; CHECK-NEXT: .LBB2_1: # %loop99; CHECK-NEXT: # =>This Inner Loop Header: Depth=1100; CHECK-NEXT: csrwi fcsr, 0101; CHECK-NEXT: fadd.d fa5, fa5, fa0102; CHECK-NEXT: addi a0, a0, -1103; CHECK-NEXT: fscsr a1104; CHECK-NEXT: fadd.d fa5, fa5, fa1105; CHECK-NEXT: beqz a0, .LBB2_1106; CHECK-NEXT: # %bb.2: # %exit107; CHECK-NEXT: fmv.d fa0, fa5108; CHECK-NEXT: ret109entry:110 br label %loop111loop:112 %cnt = phi i64 [0, %entry], [%cnt_inc, %loop]113 %acc = phi double [0.0, %entry], [%f2, %loop]114 call void @llvm.set.fpenv(i64 0) strictfp115 %f1 = call double @llvm.experimental.constrained.fadd.f64(double %acc, double %0, metadata !"round.dynamic", metadata !"fpexcept.ignore") strictfp116 call void @llvm.set.fpenv(i64 %fcsr) strictfp117 %f2 = call double @llvm.experimental.constrained.fadd.f64(double %f1, double %1, metadata !"round.dynamic", metadata !"fpexcept.ignore") strictfp118 %cnt_inc = add i64 %cnt, 1119 %cond = icmp eq i64 %cnt_inc, %n120 br i1 %cond, label %loop, label %exit121exit:122 ret double %f2123}124