From b713c7380bbd705f224c1bc1ad1e26eb8f5694e4 Mon Sep 17 00:00:00 2001
From: florian <florian@freepascal.org>
Date: Wed, 23 Dec 2020 17:25:09 +0000
Subject: [PATCH]   * implemented UseAVX512 properly   + make use of VREDUCE*
 for frac(...) if AVX512QD is a available

git-svn-id: trunk@47840 -
---
 compiler/x86/cpubase.pas    |  3 +--
 compiler/x86/nx86inl.pas    | 22 ++++++++++++++++------
 compiler/x86_64/cpuinfo.pas |  6 ++++--
 3 files changed, 21 insertions(+), 10 deletions(-)

diff --git a/compiler/x86/cpubase.pas b/compiler/x86/cpubase.pas
index b2d5e110d6..7f56482e87 100644
--- a/compiler/x86/cpubase.pas
+++ b/compiler/x86/cpubase.pas
@@ -960,8 +960,7 @@ implementation
 
   function UseAVX512: boolean;
     begin
-      // Result:=(current_settings.fputype in fpu_avx_instructionsets) {$ifndef i8086}or (CPUX86_HAS_AVXUNIT in cpu_capabilities[current_settings.cputype]){$endif i8086};
-      Result:=false;
+      Result:={$ifdef i8086}false{$else i8086}UseAVX and (FPUX86_HAS_AVX512F in fpu_capabilities[current_settings.fputype]){$endif i8086};
     end;
 
 
diff --git a/compiler/x86/nx86inl.pas b/compiler/x86/nx86inl.pas
index b179362538..ac4fce336b 100644
--- a/compiler/x86/nx86inl.pas
+++ b/compiler/x86/nx86inl.pas
@@ -1315,15 +1315,25 @@ implementation
               case tfloatdef(left.resultdef).floattype of
                 s32real:
                   begin
-                    { using left.location.register here as 3rd parameter is crucial to break dependency chains }
-                    current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VROUNDSS,S_NO,3,left.location.register,left.location.register,location.register));
-                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg_reg(A_VSUBSS,S_NO,location.register,left.location.register,location.register));
+                    if UseAVX512 and (FPUX86_HAS_AVX512DQ in fpu_capabilities[current_settings.fputype]) then
+                      current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VREDUCESS,S_NO,3,left.location.register,left.location.register,location.register))
+                    else
+                      begin
+                        { using left.location.register here as 3rd parameter is crucial to break dependency chains }
+                        current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VROUNDSS,S_NO,3,left.location.register,left.location.register,location.register));
+                        current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg_reg(A_VSUBSS,S_NO,location.register,left.location.register,location.register));
+                      end;
                   end;
                 s64real:
                   begin
-                    { using left.location.register here as 3rd parameter is crucial to break dependency chains }
-                    current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VROUNDSD,S_NO,3,left.location.register,left.location.register,location.register));
-                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg_reg(A_VSUBSD,S_NO,location.register,left.location.register,location.register));
+                    if UseAVX512 and (FPUX86_HAS_AVX512DQ in fpu_capabilities[current_settings.fputype]) then
+                      current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VREDUCESD,S_NO,3,left.location.register,left.location.register,location.register))
+                    else
+                      begin
+                        { using left.location.register here as 3rd parameter is crucial to break dependency chains }
+                        current_asmdata.CurrAsmList.concat(taicpu.op_const_reg_reg_reg(A_VROUNDSD,S_NO,3,left.location.register,left.location.register,location.register));
+                        current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg_reg(A_VSUBSD,S_NO,location.register,left.location.register,location.register));
+                      end;
                   end;
                 else
                   internalerror(2017052102);
diff --git a/compiler/x86_64/cpuinfo.pas b/compiler/x86_64/cpuinfo.pas
index 05d0537b72..a0e99a7610 100644
--- a/compiler/x86_64/cpuinfo.pas
+++ b/compiler/x86_64/cpuinfo.pas
@@ -182,7 +182,9 @@ type
    tfpuflags =
       (FPUX86_HAS_AVXUNIT,
        FPUX86_HAS_32MMREGS,
-       FPUX86_HAS_AVX512F
+       FPUX86_HAS_AVX512F,
+       FPUX86_HAS_AVX512VL,
+       FPUX86_HAS_AVX512DQ
       );
 
  const
@@ -203,7 +205,7 @@ type
       { fpu_sse42    } [],
       { fpu_avx      } [FPUX86_HAS_AVXUNIT],
       { fpu_avx2     } [FPUX86_HAS_AVXUNIT],
-      { fpu_avx512   } [FPUX86_HAS_AVXUNIT,FPUX86_HAS_32MMREGS,FPUX86_HAS_AVX512F]
+      { fpu_avx512   } [FPUX86_HAS_AVXUNIT,FPUX86_HAS_32MMREGS,FPUX86_HAS_AVX512F,FPUX86_HAS_AVX512VL,FPUX86_HAS_AVX512DQ]
    );
 
 Implementation