From d95865a2a2daeb7dd8b80c18e7409b28e7e4738a Mon Sep 17 00:00:00 2001 From: Richard Sandiford Date: Wed, 21 Aug 2013 09:04:20 +0000 Subject: [PATCH] [SystemZ] Use FI[EDX]BRA for codegen git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188895 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/SystemZ/SystemZISelLowering.cpp | 9 ++ lib/Target/SystemZ/SystemZInstrFP.td | 30 +++++ test/CodeGen/SystemZ/fp-round-01.ll | 126 ++++++++++++++++++- test/CodeGen/SystemZ/fp-round-02.ll | 195 +++++++++++++++++++++++++++++ 4 files changed, 354 insertions(+), 6 deletions(-) create mode 100644 test/CodeGen/SystemZ/fp-round-02.ll diff --git a/lib/Target/SystemZ/SystemZISelLowering.cpp b/lib/Target/SystemZ/SystemZISelLowering.cpp index 6710f89a1a4..7772b9ed443 100644 --- a/lib/Target/SystemZ/SystemZISelLowering.cpp +++ b/lib/Target/SystemZ/SystemZISelLowering.cpp @@ -209,6 +209,15 @@ SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) // We can use FI for FRINT. setOperationAction(ISD::FRINT, VT, Legal); + // We can use the extended form of FI for other rounding operations. + if (Subtarget.hasFPExtension()) { + setOperationAction(ISD::FNEARBYINT, VT, Legal); + setOperationAction(ISD::FFLOOR, VT, Legal); + setOperationAction(ISD::FCEIL, VT, Legal); + setOperationAction(ISD::FTRUNC, VT, Legal); + setOperationAction(ISD::FROUND, VT, Legal); + } + // No special instructions for these. setOperationAction(ISD::FSIN, VT, Expand); setOperationAction(ISD::FCOS, VT, Expand); diff --git a/lib/Target/SystemZ/SystemZInstrFP.td b/lib/Target/SystemZ/SystemZInstrFP.td index b407b86c2bd..dbe0fb5cac3 100644 --- a/lib/Target/SystemZ/SystemZInstrFP.td +++ b/lib/Target/SystemZ/SystemZInstrFP.td @@ -232,6 +232,36 @@ def : Pat<(frint FP32:$src), (FIEBR 0, FP32:$src)>; def : Pat<(frint FP64:$src), (FIDBR 0, FP64:$src)>; def : Pat<(frint FP128:$src), (FIXBR 0, FP128:$src)>; +let Predicates = [FeatureFPExtension] in { + // fnearbyint is like frint but does not detect inexact conditions. + def : Pat<(fnearbyint FP32:$src), (FIEBRA 0, FP32:$src, 4)>; + def : Pat<(fnearbyint FP64:$src), (FIDBRA 0, FP64:$src, 4)>; + def : Pat<(fnearbyint FP128:$src), (FIXBRA 0, FP128:$src, 4)>; + + // floor is no longer allowed to raise an inexact condition, + // so restrict it to the cases where the condition can be suppressed. + // Mode 7 is round towards -inf. + def : Pat<(ffloor FP32:$src), (FIEBRA 7, FP32:$src, 4)>; + def : Pat<(ffloor FP64:$src), (FIDBRA 7, FP64:$src, 4)>; + def : Pat<(ffloor FP128:$src), (FIXBRA 7, FP128:$src, 4)>; + + // Same idea for ceil, where mode 6 is round towards +inf. + def : Pat<(fceil FP32:$src), (FIEBRA 6, FP32:$src, 4)>; + def : Pat<(fceil FP64:$src), (FIDBRA 6, FP64:$src, 4)>; + def : Pat<(fceil FP128:$src), (FIXBRA 6, FP128:$src, 4)>; + + // Same idea for trunc, where mode 5 is round towards zero. + def : Pat<(ftrunc FP32:$src), (FIEBRA 5, FP32:$src, 4)>; + def : Pat<(ftrunc FP64:$src), (FIDBRA 5, FP64:$src, 4)>; + def : Pat<(ftrunc FP128:$src), (FIXBRA 5, FP128:$src, 4)>; + + // Same idea for round, where mode 1 is round towards nearest with + // ties away from zero. + def : Pat<(frnd FP32:$src), (FIEBRA 1, FP32:$src, 4)>; + def : Pat<(frnd FP64:$src), (FIDBRA 1, FP64:$src, 4)>; + def : Pat<(frnd FP128:$src), (FIXBRA 1, FP128:$src, 4)>; +} + //===----------------------------------------------------------------------===// // Binary arithmetic //===----------------------------------------------------------------------===// diff --git a/test/CodeGen/SystemZ/fp-round-01.ll b/test/CodeGen/SystemZ/fp-round-01.ll index f2530dc60d5..565db5ad4f5 100644 --- a/test/CodeGen/SystemZ/fp-round-01.ll +++ b/test/CodeGen/SystemZ/fp-round-01.ll @@ -1,9 +1,8 @@ -; Test rint()-like rounding, with non-integer values triggering an -; inexact condition. +; Test rounding functions for z10. ; -; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s +; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck %s -; Test f32. +; Test rint for f32. declare float @llvm.rint.f32(float %f) define float @f1(float %f) { ; CHECK-LABEL: f1: @@ -13,7 +12,7 @@ define float @f1(float %f) { ret float %res } -; Test f64. +; Test rint for f64. declare double @llvm.rint.f64(double %f) define double @f2(double %f) { ; CHECK-LABEL: f2: @@ -23,7 +22,7 @@ define double @f2(double %f) { ret double %res } -; Test f128. +; Test rint for f128. declare fp128 @llvm.rint.f128(fp128 %f) define void @f3(fp128 *%ptr) { ; CHECK-LABEL: f3: @@ -34,3 +33,118 @@ define void @f3(fp128 *%ptr) { store fp128 %res, fp128 *%ptr ret void } + +; Test nearbyint for f32. +declare float @llvm.nearbyint.f32(float %f) +define float @f4(float %f) { +; CHECK-LABEL: f4: +; CHECK: brasl %r14, nearbyintf@PLT +; CHECK: br %r14 + %res = call float @llvm.nearbyint.f32(float %f) + ret float %res +} + +; Test nearbyint for f64. +declare double @llvm.nearbyint.f64(double %f) +define double @f5(double %f) { +; CHECK-LABEL: f5: +; CHECK: brasl %r14, nearbyint@PLT +; CHECK: br %r14 + %res = call double @llvm.nearbyint.f64(double %f) + ret double %res +} + +; Test nearbyint for f128: omitted for now because we cannot handle +; indirect arguments. + +; Test floor for f32. +declare float @llvm.floor.f32(float %f) +define float @f7(float %f) { +; CHECK-LABEL: f7: +; CHECK: brasl %r14, floorf@PLT +; CHECK: br %r14 + %res = call float @llvm.floor.f32(float %f) + ret float %res +} + +; Test floor for f64. +declare double @llvm.floor.f64(double %f) +define double @f8(double %f) { +; CHECK-LABEL: f8: +; CHECK: brasl %r14, floor@PLT +; CHECK: br %r14 + %res = call double @llvm.floor.f64(double %f) + ret double %res +} + +; Test floor for f128: omitted for now because we cannot handle +; indirect arguments. + +; Test ceil for f32. +declare float @llvm.ceil.f32(float %f) +define float @f10(float %f) { +; CHECK-LABEL: f10: +; CHECK: brasl %r14, ceilf@PLT +; CHECK: br %r14 + %res = call float @llvm.ceil.f32(float %f) + ret float %res +} + +; Test ceil for f64. +declare double @llvm.ceil.f64(double %f) +define double @f11(double %f) { +; CHECK-LABEL: f11: +; CHECK: brasl %r14, ceil@PLT +; CHECK: br %r14 + %res = call double @llvm.ceil.f64(double %f) + ret double %res +} + +; Test ceil for f128: omitted for now because we cannot handle +; indirect arguments. + +; Test trunc for f32. +declare float @llvm.trunc.f32(float %f) +define float @f13(float %f) { +; CHECK-LABEL: f13: +; CHECK: brasl %r14, truncf@PLT +; CHECK: br %r14 + %res = call float @llvm.trunc.f32(float %f) + ret float %res +} + +; Test trunc for f64. +declare double @llvm.trunc.f64(double %f) +define double @f14(double %f) { +; CHECK-LABEL: f14: +; CHECK: brasl %r14, trunc@PLT +; CHECK: br %r14 + %res = call double @llvm.trunc.f64(double %f) + ret double %res +} + +; Test trunc for f128: omitted for now because we cannot handle +; indirect arguments. + +; Test round for f32. +declare float @llvm.round.f32(float %f) +define float @f16(float %f) { +; CHECK-LABEL: f16: +; CHECK: brasl %r14, roundf@PLT +; CHECK: br %r14 + %res = call float @llvm.round.f32(float %f) + ret float %res +} + +; Test round for f64. +declare double @llvm.round.f64(double %f) +define double @f17(double %f) { +; CHECK-LABEL: f17: +; CHECK: brasl %r14, round@PLT +; CHECK: br %r14 + %res = call double @llvm.round.f64(double %f) + ret double %res +} + +; Test round for f128: omitted for now because we cannot handle +; indirect arguments. diff --git a/test/CodeGen/SystemZ/fp-round-02.ll b/test/CodeGen/SystemZ/fp-round-02.ll new file mode 100644 index 00000000000..d79c9c47050 --- /dev/null +++ b/test/CodeGen/SystemZ/fp-round-02.ll @@ -0,0 +1,195 @@ +; Test rounding functions for z196 and above. +; +; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z196 | FileCheck %s + +; Test rint for f32. +declare float @llvm.rint.f32(float %f) +define float @f1(float %f) { +; CHECK-LABEL: f1: +; CHECK: fiebr %f0, 0, %f0 +; CHECK: br %r14 + %res = call float @llvm.rint.f32(float %f) + ret float %res +} + +; Test rint for f64. +declare double @llvm.rint.f64(double %f) +define double @f2(double %f) { +; CHECK-LABEL: f2: +; CHECK: fidbr %f0, 0, %f0 +; CHECK: br %r14 + %res = call double @llvm.rint.f64(double %f) + ret double %res +} + +; Test rint for f128. +declare fp128 @llvm.rint.f128(fp128 %f) +define void @f3(fp128 *%ptr) { +; CHECK-LABEL: f3: +; CHECK: fixbr %f0, 0, %f0 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.rint.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} + +; Test nearbyint for f32. +declare float @llvm.nearbyint.f32(float %f) +define float @f4(float %f) { +; CHECK-LABEL: f4: +; CHECK: fiebra %f0, 0, %f0, 4 +; CHECK: br %r14 + %res = call float @llvm.nearbyint.f32(float %f) + ret float %res +} + +; Test nearbyint for f64. +declare double @llvm.nearbyint.f64(double %f) +define double @f5(double %f) { +; CHECK-LABEL: f5: +; CHECK: fidbra %f0, 0, %f0, 4 +; CHECK: br %r14 + %res = call double @llvm.nearbyint.f64(double %f) + ret double %res +} + +; Test nearbyint for f128. +declare fp128 @llvm.nearbyint.f128(fp128 %f) +define void @f6(fp128 *%ptr) { +; CHECK-LABEL: f6: +; CHECK: fixbra %f0, 0, %f0, 4 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.nearbyint.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} + +; Test floor for f32. +declare float @llvm.floor.f32(float %f) +define float @f7(float %f) { +; CHECK-LABEL: f7: +; CHECK: fiebra %f0, 7, %f0, 4 +; CHECK: br %r14 + %res = call float @llvm.floor.f32(float %f) + ret float %res +} + +; Test floor for f64. +declare double @llvm.floor.f64(double %f) +define double @f8(double %f) { +; CHECK-LABEL: f8: +; CHECK: fidbra %f0, 7, %f0, 4 +; CHECK: br %r14 + %res = call double @llvm.floor.f64(double %f) + ret double %res +} + +; Test floor for f128. +declare fp128 @llvm.floor.f128(fp128 %f) +define void @f9(fp128 *%ptr) { +; CHECK-LABEL: f9: +; CHECK: fixbra %f0, 7, %f0, 4 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.floor.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} + +; Test ceil for f32. +declare float @llvm.ceil.f32(float %f) +define float @f10(float %f) { +; CHECK-LABEL: f10: +; CHECK: fiebra %f0, 6, %f0, 4 +; CHECK: br %r14 + %res = call float @llvm.ceil.f32(float %f) + ret float %res +} + +; Test ceil for f64. +declare double @llvm.ceil.f64(double %f) +define double @f11(double %f) { +; CHECK-LABEL: f11: +; CHECK: fidbra %f0, 6, %f0, 4 +; CHECK: br %r14 + %res = call double @llvm.ceil.f64(double %f) + ret double %res +} + +; Test ceil for f128. +declare fp128 @llvm.ceil.f128(fp128 %f) +define void @f12(fp128 *%ptr) { +; CHECK-LABEL: f12: +; CHECK: fixbra %f0, 6, %f0, 4 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.ceil.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} + +; Test trunc for f32. +declare float @llvm.trunc.f32(float %f) +define float @f13(float %f) { +; CHECK-LABEL: f13: +; CHECK: fiebra %f0, 5, %f0, 4 +; CHECK: br %r14 + %res = call float @llvm.trunc.f32(float %f) + ret float %res +} + +; Test trunc for f64. +declare double @llvm.trunc.f64(double %f) +define double @f14(double %f) { +; CHECK-LABEL: f14: +; CHECK: fidbra %f0, 5, %f0, 4 +; CHECK: br %r14 + %res = call double @llvm.trunc.f64(double %f) + ret double %res +} + +; Test trunc for f128. +declare fp128 @llvm.trunc.f128(fp128 %f) +define void @f15(fp128 *%ptr) { +; CHECK-LABEL: f15: +; CHECK: fixbra %f0, 5, %f0, 4 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.trunc.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} + +; Test round for f32. +declare float @llvm.round.f32(float %f) +define float @f16(float %f) { +; CHECK-LABEL: f16: +; CHECK: fiebra %f0, 1, %f0, 4 +; CHECK: br %r14 + %res = call float @llvm.round.f32(float %f) + ret float %res +} + +; Test round for f64. +declare double @llvm.round.f64(double %f) +define double @f17(double %f) { +; CHECK-LABEL: f17: +; CHECK: fidbra %f0, 1, %f0, 4 +; CHECK: br %r14 + %res = call double @llvm.round.f64(double %f) + ret double %res +} + +; Test round for f128. +declare fp128 @llvm.round.f128(fp128 %f) +define void @f18(fp128 *%ptr) { +; CHECK-LABEL: f18: +; CHECK: fixbra %f0, 1, %f0, 4 +; CHECK: br %r14 + %src = load fp128 *%ptr + %res = call fp128 @llvm.round.f128(fp128 %src) + store fp128 %res, fp128 *%ptr + ret void +} -- 2.11.0