Darwin: re-enable new assembler fill*word variants

Work around with an extra jump to an extra function.

rtl/i386/i386.inc

@@ -168,12 +168,14 @@ end;
 
   Since FPC does not attempt it even for local labels, this kind of code is inherently unsafe.
 }
-{$if not defined(darwin) and
-  (not defined(FPC_SYSTEM_HAS_FILLCHAR)
-   or not defined(FPC_SYSTEM_HAS_FILLWORD)
-   or not defined(FPC_SYSTEM_HAS_FILLDWORD)
-   or not defined(FPC_SYSTEM_HAS_FILLQWORD)
-)}
+{$ifndef darwin}
+  {$define can_jump_into_the_middle_of_a_procedure}
+{$endif darwin}
+
+{$if not defined(FPC_SYSTEM_HAS_FILLCHAR)
+  or not defined(FPC_SYSTEM_HAS_FILLWORD)
+  or not defined(FPC_SYSTEM_HAS_FILLDWORD)
+  or not defined(FPC_SYSTEM_HAS_FILLQWORD)}
 
 {$if not defined(FPC_SYSTEM_HAS_FILLCHAR)
    or not defined(FPC_SYSTEM_HAS_FILLWORD)
@@ -209,13 +211,15 @@ asm
 end;
 {$endif FillChar/Word/DWord required.}
 
+{$ifdef can_jump_into_the_middle_of_a_procedure}
 label
   FillXxxx_MoreThanTwoXMMs;
+{$else can_jump_into_the_middle_of_a_procedure}
+procedure FillXxxx_MoreThanTwoXMMs; forward;
+{$endif can_jump_into_the_middle_of_a_procedure}
 
 procedure FillXxxx_U32Pattern_SSE2_16OrMore; assembler; nostackframe;
 { eax — x, ecx — uint32 pattern, edx — byte count >= 16 (preferably > 16). }
-const
-  NtThreshold = 4 * 1024 * 1024;
 asm
         movd   %ecx, %xmm0
         pshufd $0, %xmm0, %xmm0 { xmm0 = pattern for unaligned writes }
@@ -240,10 +244,17 @@ asm
         movd   %esi, %xmm0
         pshufd $0, %xmm0, %xmm0
         pop    %esi
-
+{$ifdef can_jump_into_the_middle_of_a_procedure}
 { FillChar (to skip the misaligning above) and FillQWord jump here.
   eax — x, edx — byte count > 32, xmm0 = pattern for ALIGNED writes, first and last 16 bytes written. }
 FillXxxx_MoreThanTwoXMMs:
+{$else can_jump_into_the_middle_of_a_procedure}
+        jmp    FillXxxx_MoreThanTwoXMMs
+end;
+
+procedure FillXxxx_MoreThanTwoXMMs; assembler; nostackframe;
+asm
+{$endif can_jump_into_the_middle_of_a_procedure}
         lea    -65(%eax,%edx), %ecx
         and    $-16, %ecx { ecx = “T4” (possibly fictive) = loop bound. }
         mov    %ecx, %edx { Remember T4 to edx. }
@@ -259,7 +270,7 @@ FillXxxx_MoreThanTwoXMMs:
         jle    .LFourAlignedTailWrites { ecx was ≤ 96−48 }
 
         add    $48, %eax { eax = H3. }
-        cmp    $NtThreshold, %ecx
+        cmp    $4 * 1024 * 1024, %ecx { Non-temporal fill threshold. }
         jae    .L64xNT_Body
 
 .balign 16 { no-op }
@@ -339,8 +350,7 @@ end;
 {$endif FillChar/Word/DWord/QWord required.}
 
 
-{$if not defined(darwin) and
-     not defined(FPC_SYSTEM_HAS_FILLCHAR)}
+{$if not defined(FPC_SYSTEM_HAS_FILLCHAR)}
 {$define FPC_SYSTEM_HAS_FILLCHAR}
 procedure FillChar_3OrLess; assembler; nostackframe;
 { cl — x, edx — byte count, Low(int32) <= edx <= 3. }
@@ -438,8 +448,7 @@ end;
 {$endif FPC_SYSTEM_HAS_FILLCHAR}
 
 
-{$if not defined(darwin) and
-     not defined(FPC_SYSTEM_HAS_FILLWORD)}
+{$if not defined(FPC_SYSTEM_HAS_FILLWORD)}
 {$define FPC_SYSTEM_HAS_FILLWORD}
 procedure FillWord_3OrLess; assembler; nostackframe;
 asm
@@ -527,8 +536,7 @@ end;
 {$endif FPC_SYSTEM_HAS_FILLWORD}
 
 
-{$if not defined(darwin) and
-     not defined(FPC_SYSTEM_HAS_FILLDWORD)}
+{$if not defined(FPC_SYSTEM_HAS_FILLDWORD)}
 {$define FPC_SYSTEM_HAS_FILLDWORD}
 procedure FillDWord_4OrLess; assembler; nostackframe;
 asm
@@ -602,8 +610,7 @@ end;
 {$endif FPC_SYSTEM_HAS_FILLDWORD}
 
 
-{$if not defined(darwin) and
-     not defined(FPC_SYSTEM_HAS_FILLQWORD)}
+{$if not defined(FPC_SYSTEM_HAS_FILLQWORD)}
 {$define FPC_SYSTEM_HAS_FILLQWORD}
 {$ifndef CPUX86_HAS_SSE2}
 procedure FillQWord_Plain(var x;count:SizeInt;value:QWord);assembler;nostackframe;

rtl/i386/set.inc

@@ -15,14 +15,10 @@
 
 {$ifndef FPC_SYSTEM_HAS_FPC_VARSET_ADD_SETS}
 {$define FPC_SYSTEM_HAS_FPC_VARSET_ADD_SETS}
-label
-  fpc_varset_add_sets_plain_fallback;
-
 procedure fpc_varset_add_sets_plain(const set1,set2; var dest;size : ptrint); assembler; nostackframe;
 { eax = set1, edx = set2, ecx = dest, [esp + 4] = size }
 asm
     push   %ebx
-fpc_varset_add_sets_plain_fallback:
     push   %esi
     mov    12(%esp), %esi { esi = size }
     sub    $4, %esi
@@ -60,7 +56,7 @@ asm
     push   %ebx
     mov    8(%esp), %ebx
     sub    $16, %ebx { ebx = position }
-    jl     fpc_varset_add_sets_plain_fallback { probably dead branch... }
+    jl     .LFallback { Hopefully dead branch... }
 
 .L16x_Loop:
     movups (%eax,%ebx), %xmm0
@@ -75,6 +71,11 @@ asm
     orps   %xmm1, %xmm0
     movups %xmm0, (%ecx)
     pop    %ebx
+    ret    $4
+
+.LFallback:
+    pop    %ebx
+    jmp    fpc_varset_add_sets_plain
 end;
 
 {$ifndef CPUX86_HAS_SSEUNIT}
@@ -101,14 +102,10 @@ end;
 
 {$ifndef FPC_SYSTEM_HAS_FPC_VARSET_MUL_SETS}
 {$define FPC_SYSTEM_HAS_FPC_VARSET_MUL_SETS}
-label
-  fpc_varset_mul_sets_plain_fallback;
-
 procedure fpc_varset_mul_sets_plain(const set1,set2; var dest;size : ptrint); assembler; nostackframe;
 { Same as fpc_varset_add_sets_plain but with 'and' instead of 'or'. }
 asm
     push   %ebx
-fpc_varset_mul_sets_plain_fallback:
     push   %esi
     mov    12(%esp), %esi { esi = size }
     sub    $4, %esi
@@ -146,7 +143,7 @@ asm
     push   %ebx
     mov    8(%esp), %ebx
     sub    $16, %ebx { ebx = position }
-    jl     fpc_varset_mul_sets_plain_fallback { probably dead branch... }
+    jl     .LFallback { Hopefully dead branch... }
 
 .L16x_Loop:
     movups (%eax,%ebx), %xmm0
@@ -161,6 +158,11 @@ asm
     andps   %xmm1, %xmm0
     movups %xmm0, (%ecx)
     pop    %ebx
+    ret    $4
+
+.LFallback:
+    pop    %ebx
+    jmp    fpc_varset_mul_sets_plain
 end;
 
 {$ifndef CPUX86_HAS_SSEUNIT}
@@ -187,14 +189,10 @@ end;
 
 {$ifndef FPC_SYSTEM_HAS_FPC_VARSET_SUB_SETS}
 {$define FPC_SYSTEM_HAS_FPC_VARSET_SUB_SETS}
-label
-  fpc_varset_sub_sets_plain_fallback;
-
 procedure fpc_varset_sub_sets_plain(const set1,set2; var dest;size : ptrint); assembler; nostackframe;
 { eax = set1, edx = set2, ecx = dest, [esp + 4] = size }
 asm
     push   %ebx
-fpc_varset_sub_sets_plain_fallback:
     push   %esi
     mov    12(%esp), %esi { esi = size }
     sub    $4, %esi
@@ -237,7 +235,7 @@ asm
     push   %ebx
     mov    8(%esp), %ebx
     sub    $16, %ebx { ebx = position }
-    jl     fpc_varset_sub_sets_plain_fallback { probably dead branch... }
+    jl     .LFallback { Hopefully dead branch... }
 
     movups (%eax), %xmm1 { Tail, just in case (if size is always divisible by 16, 16x_Loop can be altered to handle everything instead). }
     movups (%edx), %xmm2 { Precalculated because operation is not idempotent and dest can be equal to set1/set2. }
@@ -253,6 +251,11 @@ asm
 
     movups %xmm2, (%ecx) { Write precalculated tail. }
     pop    %ebx
+    ret    $4
+
+.LFallback:
+    pop    %ebx
+    jmp    fpc_varset_sub_sets_plain
 end;
 
 {$ifndef CPUX86_HAS_SSEUNIT}
@@ -279,15 +282,11 @@ end;
 
 {$ifndef FPC_SYSTEM_HAS_FPC_VARSET_SYMDIF_SETS}
 {$define FPC_SYSTEM_HAS_FPC_VARSET_SYMDIF_SETS}
-label
-  fpc_varset_symdif_sets_plain_fallback;
-
 procedure fpc_varset_symdif_sets_plain(const set1,set2; var dest;size : ptrint); assembler; nostackframe;
 { Same as fpc_varset_sub_sets_plain but with 'xor' instead of 'and not'.
   eax = set1, edx = set2, ecx = dest, [esp + 4] = size }
 asm
     push   %ebx
-fpc_varset_symdif_sets_plain_fallback:
     push   %esi
     mov    12(%esp), %esi { esi = size }
     sub    $4, %esi
@@ -328,7 +327,7 @@ asm
     push   %ebx
     mov    8(%esp), %ebx
     sub    $16, %ebx { ebx = position }
-    jl     fpc_varset_symdif_sets_plain_fallback { probably dead branch... }
+    jl     .LFallback { Hopefully dead branch... }
 
     movups (%eax), %xmm1 { Tail, just in case (if size is always divisible by 16, 16x_Loop can be altered to handle everything instead). }
     movups (%edx), %xmm2 { Precalculated because operation is not idempotent and dest can be equal to set1/set2. }
@@ -344,6 +343,11 @@ asm
 
     movups %xmm2, (%ecx) { Write precalculated tail. }
     pop    %ebx
+    ret    $4
+
+.LFallback:
+    pop    %ebx
+    jmp    fpc_varset_symdif_sets_plain
 end;
 
 {$ifndef CPUX86_HAS_SSEUNIT}