On 05/01/15 21:32, Gisle Vanem wrote:
Pádraig Brady wrote:
Does the following address the issue for you?
It should drop back to the more generic slow path on 32 bit.
But I do have _BitReverse. That could be an intrinsic.
I was assuming we could do the simpler patch and be
slightly slower on the shrinking 32 bit target.
Though I suppose it's only slightly more awkward to use
builtins where available, so I've done that in the attached.
BTW do we have the same issue with __popcnt64 in count-one-bits.h?
Seems so:
__MACHINEX64(unsigned __int64 __popcnt64(unsigned __int64))
X86 compilers only.
I've included that too, and also count-trailing-zeros is updated.
thanks,
Pádraig.
p.s. I noticed that count_one_bits() branches on the popcount_support
variable on each call, which might negate much of the gain from using this
instruction?
From 061484434eb09a80f872cab0e173847668b46e4f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?P=C3=A1draig=20Brady?= p...@draigbrady.com
Date: Tue, 6 Jan 2015 02:27:16 +
Subject: [PATCH] count-leading-zeros: avoid 64-bit intrinsics on 32-bit
Windows
* lib/count-leading-zeros.h (count_leading_zeros_ll): Use 32 bit
intrinsics in this case.
* lib/count-trailing-zeros.h: Likewise.
* lib/count-one-bits.h: Likewise.
---
ChangeLog | 8
lib/count-leading-zeros.h | 5 +
lib/count-one-bits.h | 4
lib/count-trailing-zeros.h | 5 +
4 files changed, 22 insertions(+)
diff --git a/ChangeLog b/ChangeLog
index 938e031..3b08279 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,11 @@
+2015-01-06 Pádraig Brady p...@draigbrady.com
+
+ count-leading-zeros: avoid 64-bit intrinsics on 32-bit Windows
+ * lib/count-leading-zeros.h (count_leading_zeros_ll): Use 32 bit
+ intrinsics in this case.
+ * lib/count-trailing-zeros.h: Likewise.
+ * lib/count-one-bits.h: Likewise.
+
2015-01-04 Benno Schulenberg bensb...@justemail.net
maint: fix grammar nits in propername (trivial change)
diff --git a/lib/count-leading-zeros.h b/lib/count-leading-zeros.h
index 722615f..9e72fff 100644
--- a/lib/count-leading-zeros.h
+++ b/lib/count-leading-zeros.h
@@ -104,8 +104,13 @@ count_leading_zeros_l (unsigned long int x)
COUNT_LEADING_ZEROS_INLINE int
count_leading_zeros_ll (unsigned long long int x)
{
+# if _MSC_VER ! defined _M_X64
+ int count = count_leading_zeros (x 31 1);
+ return count 32 ? count : 32 + count_leading_zeros (x);
+# else
COUNT_LEADING_ZEROS (__builtin_clzll, _BitScanReverse64,
unsigned long long int);
+# endif
}
#endif
diff --git a/lib/count-one-bits.h b/lib/count-one-bits.h
index d54397f..bd79fc5 100644
--- a/lib/count-one-bits.h
+++ b/lib/count-one-bits.h
@@ -127,7 +127,11 @@ count_one_bits_l (unsigned long int x)
COUNT_ONE_BITS_INLINE int
count_one_bits_ll (unsigned long long int x)
{
+# if _MSC_VER ! defined _M_X64
+ return count_one_bits (x 31 1) + count_one_bits (x);
+# else
COUNT_ONE_BITS (__builtin_popcountll, __popcnt64, unsigned long long int);
+# endif
}
#endif
diff --git a/lib/count-trailing-zeros.h b/lib/count-trailing-zeros.h
index 83ce2fb..1e71977 100644
--- a/lib/count-trailing-zeros.h
+++ b/lib/count-trailing-zeros.h
@@ -96,8 +96,13 @@ count_trailing_zeros_l (unsigned long int x)
COUNT_TRAILING_ZEROS_INLINE int
count_trailing_zeros_ll (unsigned long long int x)
{
+# if _MSC_VER ! defined _M_X64
+ int count = count_trailing_zeros (x);
+ return count 32 ? count : 32 + count_trailing_zeros (x 31 1);
+# else
COUNT_TRAILING_ZEROS (__builtin_ctzll, _BitScanForward64,
unsigned long long int);
+# endif
}
#endif
--
2.1.0