From 21348cfda233e9acb175f7025c123b4d30c238f9 Mon Sep 17 00:00:00 2001 From: Oliver Stannard Date: Wed, 7 Oct 2015 16:58:49 +0000 Subject: [PATCH] [ARM] Use correct half-precision functions in EABI mode The ARM RTABI defines the half- to single-precision float conversion functions with an __aeabi prefix, but libgcc only has them with a __gnu prefix. Therefore we need to emit the __aeabi version when compiling with an eabi or eabihf triple, and the __gnu version with a gnueabi or gnueabihf triple. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@249565 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/ARM/ARMISelLowering.cpp | 8 +++ test/CodeGen/ARM/fp16.ll | 62 +++++++++++-------- .../CodeGen/Thumb2/float-intrinsics-double.ll | 2 +- test/CodeGen/Thumb2/float-intrinsics-float.ll | 4 +- 4 files changed, 47 insertions(+), 29 deletions(-) diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp index cafb4500d1a..64e494be514 100644 --- a/lib/Target/ARM/ARMISelLowering.cpp +++ b/lib/Target/ARM/ARMISelLowering.cpp @@ -396,6 +396,14 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM, setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS); } + // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have + // a __gnu_ prefix (which is the default). + if (Subtarget->isTargetAEABI()) { + setLibcallName(RTLIB::FPROUND_F32_F16, "__aeabi_f2h"); + setLibcallName(RTLIB::FPROUND_F64_F16, "__aeabi_d2h"); + setLibcallName(RTLIB::FPEXT_F16_F32, "__aeabi_h2f"); + } + if (Subtarget->isThumb1Only()) addRegisterClass(MVT::i32, &ARM::tGPRRegClass); else diff --git a/test/CodeGen/ARM/fp16.ll b/test/CodeGen/ARM/fp16.ll index 25fbf9070cb..73d5c36a9c2 100644 --- a/test/CodeGen/ARM/fp16.ll +++ b/test/CodeGen/ARM/fp16.ll @@ -1,7 +1,9 @@ -; RUN: llc < %s | FileCheck %s -; RUN: llc -mattr=+vfp3,+fp16 < %s | FileCheck --check-prefix=CHECK-FP16 %s -; RUN: llc -mtriple=armv8-eabihf < %s | FileCheck --check-prefix=CHECK-ARMV8 %s -; RUN: llc -mtriple=thumbv7m-eabi < %s | FileCheck --check-prefix=CHECK-SOFTFLOAT %s +; RUN: llc -mtriple=armv7a--none-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-EABI %s +; RUN: llc -mtriple=armv7a--none-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-GNU %s +; RUN: llc -mattr=+vfp3,+fp16 < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 %s +; RUN: llc -mtriple=armv8-eabihf < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-ARMV8 %s +; RUN: llc -mtriple=thumbv7m-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-EABI %s +; RUN: llc -mtriple=thumbv7m-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-GNU %s target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-n32" target triple = "armv7---eabihf" @@ -12,41 +14,45 @@ target triple = "armv7---eabihf" define void @foo() nounwind { ; CHECK-LABEL: foo: -; CHECK-FP16-LABEL: foo: -; CHECK-ARMV8-LABEL: foo: -; CHECK-SOFTFLOAT-LABEL: foo: entry: %0 = load i16, i16* @x, align 2 %1 = load i16, i16* @y, align 2 %2 = tail call float @llvm.convert.from.fp16.f32(i16 %0) -; CHECK: __gnu_h2f_ieee +; CHECK-HARDFLOAT-EABI: __aeabi_h2f +; CHECK-HARDFLOAT-GNU: __gnu_h2f_ieee ; CHECK-FP16: vcvtb.f32.f16 ; CHECK-ARMv8: vcvtb.f32.f16 -; CHECK-SOFTFLOAT: __gnu_h2f_ieee +; CHECK-SOFTFLOAT-EABI: __aeabi_h2f +; CHECK-SOFTFLOAT-GNU: __gnu_h2f_ieee %3 = tail call float @llvm.convert.from.fp16.f32(i16 %1) -; CHECK: __gnu_h2f_ieee +; CHECK-HARDFLOAT-EABI: __aeabi_h2f +; CHECK-HARDFLOAT-GNU: __gnu_h2f_ieee ; CHECK-FP16: vcvtb.f32.f16 ; CHECK-ARMV8: vcvtb.f32.f16 -; CHECK-SOFTFLOAT: __gnu_h2f_ieee +; CHECK-SOFTFLOAT-EABI: __aeabi_h2f +; CHECK-SOFTFLOAT-GNU: __gnu_h2f_ieee %4 = fadd float %2, %3 %5 = tail call i16 @llvm.convert.to.fp16.f32(float %4) -; CHECK: __gnu_f2h_ieee +; CHECK-HARDFLOAT-EABI: __aeabi_f2h +; CHECK-HARDFLOAT-GNU: __gnu_f2h_ieee ; CHECK-FP16: vcvtb.f16.f32 ; CHECK-ARMV8: vcvtb.f16.f32 -; CHECK-SOFTFLOAT: __gnu_f2h_ieee +; CHECK-SOFTFLOAT-EABI: __aeabi_f2h +; CHECK-SOFTFLOAT-GNU: __gnu_f2h_ieee store i16 %5, i16* @x, align 2 ret void } define double @test_from_fp16(i16 %in) { ; CHECK-LABEL: test_from_fp16: -; CHECK-FP16-LABEL: test_from_fp16: -; CHECK-ARMV8-LABEL: test_from_fp16: -; CHECK-SOFTFLOAT-LABEL: test_from_fp16: %val = call double @llvm.convert.from.fp16.f64(i16 %in) -; CHECK: bl __gnu_h2f_ieee -; CHECK: vmov [[TMP:s[0-9]+]], r0 -; CHECK: vcvt.f64.f32 d0, [[TMP]] +; CHECK-HARDFLOAT-EABI: bl __aeabi_h2f +; CHECK-HARDFLOAT-EABI: vmov [[TMP:s[0-9]+]], r0 +; CHECK-HARDFLOAT-EABI: vcvt.f64.f32 {{d[0-9]+}}, [[TMP]] + +; CHECK-HARDFLOAT-GNU: bl __gnu_h2f_ieee +; CHECK-HARDFLOAT-GNU: vmov [[TMP:s[0-9]+]], r0 +; CHECK-HARDFLOAT-GNU: vcvt.f64.f32 {{d[0-9]+}}, [[TMP]] ; CHECK-FP16: vmov [[TMP16:s[0-9]+]], r0 ; CHECK-FP16: vcvtb.f32.f16 [[TMP32:s[0-9]+]], [[TMP16]] @@ -55,25 +61,29 @@ define double @test_from_fp16(i16 %in) { ; CHECK-ARMV8: vmov [[TMP:s[0-9]+]], r0 ; CHECK-ARMV8: vcvtb.f64.f16 d0, [[TMP]] -; CHECK-SOFTFLOAT: bl __gnu_h2f_ieee -; CHECK-SOFTFLOAT: bl __aeabi_f2d +; CHECK-SOFTFLOAT-EABI: bl __aeabi_h2f +; CHECK-SOFTFLOAT-EABI: bl __aeabi_f2d + +; CHECK-SOFTFLOAT-GNU: bl __gnu_h2f_ieee +; CHECK-SOFTFLOAT-GNU: bl __aeabi_f2d ret double %val } define i16 @test_to_fp16(double %in) { ; CHECK-LABEL: test_to_fp16: -; CHECK-FP16-LABEL: test_to_fp16: -; CHECK-ARMV8-LABEL: test_to_fp16: -; CHECK-SOFTFLOAT-LABEL: test_to_fp16: %val = call i16 @llvm.convert.to.fp16.f64(double %in) -; CHECK: bl __aeabi_d2h +; CHECK-HARDFLOAT-EABI: bl __aeabi_d2h + +; CHECK-HARDFLOAT-GNU: bl __aeabi_d2h ; CHECK-FP16: bl __aeabi_d2h ; CHECK-ARMV8: vcvtb.f16.f64 [[TMP:s[0-9]+]], d0 ; CHECK-ARMV8: vmov r0, [[TMP]] -; CHECK-SOFTFLOAT: bl __aeabi_d2h +; CHECK-SOFTFLOAT-EABI: bl __aeabi_d2h + +; CHECK-SOFTFLOAT-GNU: bl __aeabi_d2h ret i16 %val } diff --git a/test/CodeGen/Thumb2/float-intrinsics-double.ll b/test/CodeGen/Thumb2/float-intrinsics-double.ll index 38d04d30b87..fb9df2f9c42 100644 --- a/test/CodeGen/Thumb2/float-intrinsics-double.ll +++ b/test/CodeGen/Thumb2/float-intrinsics-double.ll @@ -215,7 +215,7 @@ define i16 @d_to_h(double %a) { declare double @llvm.convert.from.fp16.f64(i16 %a) define double @h_to_d(i16 %a) { ; CHECK-LABEL: h_to_d: -; NONE: bl __gnu_h2f_ieee +; NONE: bl __aeabi_h2f ; NONE: bl __aeabi_f2d ; SP: vcvt{{[bt]}}.f32.f16 ; SP: bl __aeabi_f2d diff --git a/test/CodeGen/Thumb2/float-intrinsics-float.ll b/test/CodeGen/Thumb2/float-intrinsics-float.ll index ec1bcd3708a..847aeacd2f9 100644 --- a/test/CodeGen/Thumb2/float-intrinsics-float.ll +++ b/test/CodeGen/Thumb2/float-intrinsics-float.ll @@ -205,7 +205,7 @@ define float @fmuladd_f(float %a, float %b, float %c) { declare i16 @llvm.convert.to.fp16.f32(float %a) define i16 @f_to_h(float %a) { ; CHECK-LABEL: f_to_h: -; SOFT: bl __gnu_f2h_ieee +; SOFT: bl __aeabi_f2h ; HARD: vcvt{{[bt]}}.f16.f32 %1 = call i16 @llvm.convert.to.fp16.f32(float %a) ret i16 %1 @@ -214,7 +214,7 @@ define i16 @f_to_h(float %a) { declare float @llvm.convert.from.fp16.f32(i16 %a) define float @h_to_f(i16 %a) { ; CHECK-LABEL: h_to_f: -; SOFT: bl __gnu_h2f_ieee +; SOFT: bl __aeabi_h2f ; HARD: vcvt{{[bt]}}.f32.f16 %1 = call float @llvm.convert.from.fp16.f32(i16 %a) ret float %1 -- 2.34.1