From a6a17efa42d9de2a1206311b536b5de0487e6a67 Mon Sep 17 00:00:00 2001
From: Jonas Maebe <jonas@freepascal.org>
Date: Mon, 30 Dec 2019 15:05:13 +0000
Subject: [PATCH] * use LLVM constrained fpext/fptrunc intrinsics when
fastmath is not enabled for accurate exception behaviour
git-svn-id: trunk@43820 -
---
compiler/llvm/hlcgllvm.pas | 54 +++++++++++++++++++++++++++++++++++---
compiler/llvm/llvmdef.pas | 19 ++++++++++++++
compiler/llvm/llvminfo.pas | 5 ++--
rtl/inc/llvmintr.inc | 19 ++++++++++++++
4 files changed, 92 insertions(+), 5 deletions(-)
diff --git a/compiler/llvm/hlcgllvm.pas b/compiler/llvm/hlcgllvm.pas
index 318cbe3f64..75d1b4a8e5 100644
--- a/compiler/llvm/hlcgllvm.pas
+++ b/compiler/llvm/hlcgllvm.pas
@@ -96,6 +96,9 @@ uses
procedure a_loadfpu_ref_reg(list: TAsmList; fromsize, tosize: tdef; const ref: treference; reg: tregister); override;
procedure a_loadfpu_reg_ref(list: TAsmList; fromsize, tosize: tdef; reg: tregister; const ref: treference); override;
procedure a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister); override;
+ protected
+ procedure gen_fpconstrained_intrinsic(list: TAsmList; const intrinsic: TIDString; fromsize, tosize: tdef; fromreg, toreg: tregister);
+ public
procedure gen_proc_symbol(list: TAsmList); override;
procedure handle_external_proc(list: TAsmList; pd: tprocdef; const importname: TSymStr); override;
@@ -165,7 +168,7 @@ implementation
verbose,cutils,globals,fmodule,constexp,systems,
defutil,llvmdef,llvmsym,
aasmtai,aasmcpu,
- aasmllvm,llvmbase,llvminfo,tgllvm,
+ aasmllvm,aasmllvmmetadata,llvmbase,llvminfo,tgllvm,
symtable,symllvm,
paramgr,
pass_2,procinfo,llvmpi,cpuinfo,cgobj,cgllvm,cghlcpu,
@@ -1321,10 +1324,55 @@ implementation
procedure thlcgllvm.a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister);
var
op: tllvmop;
+ intrinsic: TIDString;
begin
op:=llvmconvop(fromsize,tosize,true);
- { reg2 = bitcast fromllsize reg1 to tollsize }
- list.concat(taillvm.op_reg_size_reg_size(op,reg2,fromsize,reg1,tosize));
+ if (cs_opt_fastmath in current_settings.optimizerswitches) or
+ not(llvmflag_constrained_fptrunc_fpext in llvmversion_properties[current_settings.llvmversion]) or
+ not(op in [la_fptrunc,la_fpext]) then
+ list.concat(taillvm.op_reg_size_reg_size(op,reg2,fromsize,reg1,tosize))
+ else
+ begin
+ if op=la_fptrunc then
+ intrinsic:='llvm_experimental_constrained_fptrunc'
+ else
+ intrinsic:='llvm_experimental_constrained_fpext';
+ gen_fpconstrained_intrinsic(list,
+ intrinsic+llvmfloatintrinsicsuffix(tfloatdef(tosize))+llvmfloatintrinsicsuffix(tfloatdef(fromsize)),
+ fromsize,tosize,reg1,reg2);
+ end;
+ end;
+
+
+ procedure thlcgllvm.gen_fpconstrained_intrinsic(list: TAsmList; const intrinsic: TIDString; fromsize, tosize: tdef; fromreg, toreg: tregister);
+ var
+ frompara, roundpara, exceptpara, respara: tcgpara;
+ tmploc: tlocation;
+ pd: tprocdef;
+ begin
+ frompara.init;
+ roundpara.init;
+ exceptpara.init;
+ pd:=search_system_proc(intrinsic);
+
+ paramanager.getcgtempparaloc(list,pd,1,frompara);
+ paramanager.getcgtempparaloc(list,pd,2,roundpara);
+ paramanager.getcgtempparaloc(list,pd,3,exceptpara);
+
+ location_reset(tmploc,frompara.location^.loc,def_cgsize(fromsize));
+ tmploc.register:=fromreg;
+ gen_load_loc_cgpara(list,fromsize,tmploc,frompara);
+ a_load_reg_cgpara(list,llvm_metadatatype,tllvmmetadata.getstringreg('round.dynamic'),roundpara);
+ a_load_reg_cgpara(list,llvm_metadatatype,tllvmmetadata.getstringreg('fpexcept.strict'),exceptpara);
+ respara:=g_call_system_proc(list,pd,[@frompara,@roundpara,@exceptpara],nil);
+
+ location_reset(tmploc,respara.location^.loc,def_cgsize(tosize));
+ tmploc.register:=toreg;
+ gen_load_cgpara_loc(list,tosize,respara,tmploc,false);
+ frompara.done;
+ roundpara.done;
+ exceptpara.done;
+ respara.resetiftemp;
end;
diff --git a/compiler/llvm/llvmdef.pas b/compiler/llvm/llvmdef.pas
index ea8757a7dc..34d3d99af2 100644
--- a/compiler/llvm/llvmdef.pas
+++ b/compiler/llvm/llvmdef.pas
@@ -109,6 +109,8 @@ interface
function llvmasmsymname(const sym: TAsmSymbol): TSymStr;
+ function llvmfloatintrinsicsuffix(def: tfloatdef): TIDString;
+
implementation
@@ -290,6 +292,23 @@ implementation
result:='label %'+sym.name;
end;
+ function llvmfloatintrinsicsuffix(def: tfloatdef): TIDString;
+ begin
+ case def.floattype of
+ s32real:
+ result:='_f32';
+ s64real:
+ result:='_f64';
+ s80real,sc80real:
+ result:='_f80';
+ s128real:
+ result:='_f128';
+ else
+ { comp/currency need to be converted to s(c)80real first }
+ internalerror(2019122902);
+ end;
+ end;
+
function llvmbyvalparaloc(paraloc: pcgparalocation): boolean;
begin
diff --git a/compiler/llvm/llvminfo.pas b/compiler/llvm/llvminfo.pas
index 13503ba24b..05b82c1031 100644
--- a/compiler/llvm/llvminfo.pas
+++ b/compiler/llvm/llvminfo.pas
@@ -47,7 +47,8 @@ Type
type
tllvmversionflag = (
llvmflag_memcpy_indiv_align, { memcpy intrinsic supports separate alignment for source and dest }
- llvmflag_null_pointer_valid { supports "llvmflag_null_pointer_valid" attribute, which indicates access to nil should not be optimized as undefined behaviour }
+ llvmflag_null_pointer_valid, { supports "llvmflag_null_pointer_valid" attribute, which indicates access to nil should not be optimized as undefined behaviour }
+ llvmflag_constrained_fptrunc_fpext { supports constrained fptrunc and fpext intrinsics }
);
tllvmversionflags = set of tllvmversionflag;
@@ -70,7 +71,7 @@ Const
{ llvmver_7_0 } [llvmflag_memcpy_indiv_align,llvmflag_null_pointer_valid],
{ llvmver_7_1 } [llvmflag_memcpy_indiv_align,llvmflag_null_pointer_valid],
{ llvmver_8_0 } [llvmflag_memcpy_indiv_align,llvmflag_null_pointer_valid],
- { llvmver_9_0 } [llvmflag_memcpy_indiv_align,llvmflag_null_pointer_valid]
+ { llvmver_9_0 } [llvmflag_memcpy_indiv_align,llvmflag_null_pointer_valid,llvmflag_constrained_fptrunc_fpext]
);
{ Supported optimizations, only used for information }
diff --git a/rtl/inc/llvmintr.inc b/rtl/inc/llvmintr.inc
index 3aac1d1ab2..880be2ce51 100644
--- a/rtl/inc/llvmintr.inc
+++ b/rtl/inc/llvmintr.inc
@@ -86,3 +86,22 @@ function llvm_experimental_constrained_fsub(a, b: float128; rounding, exceptions
function llvm_experimental_constrained_fmul(a, b: float128; rounding, exceptions: LLVMMetadata): float128; external name 'llvm.experimental.constrained.fmul.f128';
function llvm_experimental_constrained_fdiv(a, b: float128; rounding, exceptions: LLVMMetadata): float128; external name 'llvm.experimental.constrained.fdiv.f128';
{$endif}
+
+function llvm_experimental_constrained_fptrunc_f32_f64(a: double; rounding, exceptions: LLVMMetadata): single; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f32.f64';
+function llvm_experimental_constrained_fpext_f64_f32(a: single; rounding, exceptions: LLVMMetadata): double; compilerproc; external name 'llvm.experimental.constrained.fpext.f64.f32';
+{$ifdef SUPPORT_EXTENDED}
+function llvm_experimental_constrained_fptrunc_f32_f80(a: extended; rounding, exceptions: LLVMMetadata): single; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f32.x86_fp80';
+function llvm_experimental_constrained_fptrunc_f64_f80(a: extended; rounding, exceptions: LLVMMetadata): double; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f64.x86_fp80';
+function llvm_experimental_constrained_fpext_f80_f32(a: single; rounding, exceptions: LLVMMetadata): extended; compilerproc; external name 'llvm.experimental.constrained.fpext.x86_fp80.f32';
+function llvm_experimental_constrained_fpext_f80_f64(a: double; rounding, exceptions: LLVMMetadata): extended; compilerproc; external name 'llvm.experimental.constrained.fpext.x86_fp80.f64';
+{$ifdef SUPPORT_FLOAT128}
+function llvm_experimental_constrained_fptrunc_f128_f80(a: extended; rounding, exceptions: LLVMMetadata): float128; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f128.x86_fp80';
+function llvm_experimental_constrained_fpext_f80_f32(a: float128; rounding, exceptions: LLVMMetadata): extended; compilerproc; external name 'llvm.experimental.constrained.fpext.x86_fp80.f128';
+{$endif}
+{$endif}
+{$ifdef SUPPORT_FLOAT128}
+function llvm_experimental_constrained_fptrunc_f32_f128(a: float128; rounding, exceptions: LLVMMetadata): single; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f32.f128';
+function llvm_experimental_constrained_fptrunc_f64_f128(a: float128; rounding, exceptions: LLVMMetadata): double; compilerproc; external name 'llvm.experimental.constrained.fptrunc.f64.f128';
+function llvm_experimental_constrained_fpext_f128_f32(a: single; rounding, exceptions: LLVMMetadata): float128; compilerproc; external name 'llvm.experimental.constrained.fpext.f128.f32';
+function llvm_experimental_constrained_fpext_f128_f64(a: double; rounding, exceptions: LLVMMetadata): float128; compilerproc; external name 'llvm.experimental.constrained.fpext.f128.f64';
+{$endif}