Changeset 102938 in vbox for trunk/include/iprt/asm.h

Timestamp:

Jan 17, 2024 7:39:48 PM (13 months ago)

Author:

vboxsync

Message:

iprt/asm.h: Added a CLREX instruction to the cmpxchg sequences on arm64, so we don't keep the CPU in the locked state unnecessarily long. This speeds up the negative (test fail) benchmarks of the ASMAtomicCmpXchg* functions. bugref:9898

File:

• trunk/include/iprt/asm.h (modified) (15 diffs)

trunk/include/iprt/asm.h

@@ -304 +304 @@
 #   define RTASM_ARM_DMB_LD             RTASM_ARM_DMB_SY
 #   define RTASM_ARM_DMB_LD_IN_REG      RTASM_ARM_DMB_SY_IN_REG
+
 #  elif RT_ARCH_ARM32 >= 4
 #   warning armv5 or older
@@ -1094 +1095 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrexb    %[uOld], %[pMem]\n\t"
@@ -1102 +1105 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgU8_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu8)
                          , [uOld]   "=&r" (u32Spill)
@@ -1231 +1235 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrex     %[uOld], %[pMem]\n\t"
@@ -1239 +1245 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgU32_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu32)
                          , [uOld]   "=&r" (u32Spill)
@@ -1393 +1400 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrexd    %[uOld], %H[uOld], %[pMem]\n\t"
@@ -1402 +1411 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu64)
                          , [uOld]   "=&r" (u64Spill)
@@ -1738 +1748 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrexb     %[uOld], %[pMem]\n\t"
@@ -1746 +1758 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgExU8_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu8)
                          , [uOld]   "=&r" (u8ActualOld)
@@ -1863 +1876 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrexh     %[uOld], %[pMem]\n\t"
@@ -1871 +1886 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgExU16_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu16)
                          , [uOld]   "=&r" (u16ActualOld)
@@ -1988 +2004 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrex     %[uOld], %[pMem]\n\t"
@@ -1996 +2014 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgExU32_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu32)
                          , [uOld]   "=&r" (u32ActualOld)
@@ -2158 +2177 @@
                          "cbnz      %w[rc], Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
                          "mov       %w[fXchg], #1\n\t"
+                         "1:\n\t"
+                         "clrex\n\t"
 #  else
                          "ldrexd    %[uOld], %H[uOld], %[pMem]\n\t"
@@ -2167 +2188 @@
                          "bne       Ltry_again_ASMAtomicCmpXchgU64_%=\n\t"
                          "mov       %[fXchg], #1\n\t"
-#  endif
                          "1:\n\t"
+                         /** @todo clrexne on armv7? */
+#  endif
                          : [pMem]   "+Q"  (*pu64)
                          , [uOld]   "=&r" (u64ActualOld)