[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH 7/8] softfloat: Use x86_64 assembly for {add,sub}{192,256}
From: |
Richard Henderson |
Subject: |
[PATCH 7/8] softfloat: Use x86_64 assembly for {add,sub}{192,256} |
Date: |
Wed, 23 Sep 2020 18:24:52 -0700 |
The compiler cannot chain more than two additions together.
Use inline assembly for 3 or 4 additions.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/fpu/softfloat-macros.h | 18 ++++++++++++++++--
fpu/softfloat.c | 28 ++++++++++++++++++++++++++++
2 files changed, 44 insertions(+), 2 deletions(-)
diff --git a/include/fpu/softfloat-macros.h b/include/fpu/softfloat-macros.h
index 95d88d05b8..99fa124e56 100644
--- a/include/fpu/softfloat-macros.h
+++ b/include/fpu/softfloat-macros.h
@@ -436,6 +436,13 @@ static inline void
uint64_t *z2Ptr
)
{
+#ifdef __x86_64__
+ asm("add %5, %2\n\t"
+ "adc %4, %1\n\t"
+ "adc %3, %0"
+ : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
+ : "rm"(b0), "rm"(b1), "rm"(b2), "0"(a0), "1"(a1), "2"(a2));
+#else
uint64_t z0, z1, z2;
int8_t carry0, carry1;
@@ -450,7 +457,7 @@ static inline void
*z2Ptr = z2;
*z1Ptr = z1;
*z0Ptr = z0;
-
+#endif
}
/*----------------------------------------------------------------------------
@@ -494,6 +501,13 @@ static inline void
uint64_t *z2Ptr
)
{
+#ifdef __x86_64__
+ asm("sub %5, %2\n\t"
+ "sbb %4, %1\n\t"
+ "sbb %3, %0"
+ : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
+ : "rm"(b0), "rm"(b1), "rm"(b2), "0"(a0), "1"(a1), "2"(a2));
+#else
uint64_t z0, z1, z2;
int8_t borrow0, borrow1;
@@ -508,7 +522,7 @@ static inline void
*z2Ptr = z2;
*z1Ptr = z1;
*z0Ptr = z0;
-
+#endif
}
/*----------------------------------------------------------------------------
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index 5b714fbd82..d8e5d90fd7 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -7297,6 +7297,15 @@ static void shift256RightJamming(uint64_t p[4], int
count)
/* R = A - B */
static void sub256(uint64_t r[4], uint64_t a[4], uint64_t b[4])
{
+#if defined(__x86_64__)
+ asm("sub %7, %3\n\t"
+ "sbb %6, %2\n\t"
+ "sbb %5, %1\n\t"
+ "sbb %4, %0"
+ : "=&r"(r[0]), "=&r"(r[1]), "=&r"(r[2]), "=&r"(r[3])
+ : "rme"(b[0]), "rme"(b[1]), "rme"(b[2]), "rme"(b[3]),
+ "0"(a[0]), "1"(a[1]), "2"(a[2]), "3"(a[3]));
+#else
bool borrow = false;
for (int i = 3; i >= 0; --i) {
@@ -7308,11 +7317,20 @@ static void sub256(uint64_t r[4], uint64_t a[4],
uint64_t b[4])
r[i] = a[i] - b[i];
}
}
+#endif
}
/* A = -A */
static void neg256(uint64_t a[4])
{
+#if defined(__x86_64__)
+ asm("negq %3\n\t"
+ "sbb %6, %2\n\t"
+ "sbb %5, %1\n\t"
+ "sbb %4, %0"
+ : "=&r"(a[0]), "=&r"(a[1]), "=&r"(a[2]), "+rm"(a[3])
+ : "rme"(a[0]), "rme"(a[1]), "rme"(a[2]), "0"(0), "1"(0), "2"(0));
+#else
a[3] = -a[3];
if (likely(a[3])) {
goto not2;
@@ -7333,11 +7351,20 @@ static void neg256(uint64_t a[4])
a[1] = ~a[1];
not0:
a[0] = ~a[0];
+#endif
}
/* A += B */
static void add256(uint64_t a[4], uint64_t b[4])
{
+#if defined(__x86_64__)
+ asm("add %7, %3\n\t"
+ "adc %6, %2\n\t"
+ "adc %5, %1\n\t"
+ "adc %4, %0"
+ : "+r"(a[0]), "+r"(a[1]), "+r"(a[2]), "+r"(a[3])
+ : "rme"(b[0]), "rme"(b[1]), "rme"(b[2]), "rme"(b[3]));
+#else
bool carry = false;
for (int i = 3; i >= 0; --i) {
@@ -7350,6 +7377,7 @@ static void add256(uint64_t a[4], uint64_t b[4])
}
a[i] = t;
}
+#endif
}
float128 float128_muladd(float128 a_f, float128 b_f, float128 c_f,
--
2.25.1
- [PATCH 3/8] softfloat: Tidy a * b + inf return, (continued)
- [PATCH 7/8] softfloat: Use x86_64 assembly for {add,sub}{192,256},
Richard Henderson <=
- [PATCH 8/8] softfloat: Use aarch64 assembly for {add,sub}{192,256}, Richard Henderson, 2020/09/23
- Re: [PATCH 0/8] softfloat: Implement float128_muladd, David Hildenbrand, 2020/09/24