[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH] sh4: mac.l: implement saturation arithmetic logic
From: |
Zack Buhman |
Subject: |
[PATCH] sh4: mac.l: implement saturation arithmetic logic |
Date: |
Thu, 4 Apr 2024 23:10:35 +0800 |
The saturation arithmetic logic in helper_macl is not correct.
I tested and verified this behavior on a SH7091, the general pattern
is a code sequence such as:
sets
mov.l _mach,r2
lds r2,mach
mov.l _macl,r2
lds r2,macl
mova _n,r0
mov r0,r1
mova _m,r0
mac.l @r0+,@r1+
_mach: .long 0x00007fff
_macl: .long 0x12345678
_m: .long 0x7fffffff
_n: .long 0x7fffffff
Test case 0: (no int64_t overflow)
given; prior to saturation mac.l:
mach = 0x00007fff macl = 0x12345678
@r0 = 0x7fffffff @r1 = 0x7fffffff
expected saturation mac.l result:
mach = 0x00007fff macl = 0xffffffff
qemu saturation mac.l result (prior to this commit):
mach = 0x00007ffe macl = 0x12345678
Test case 1: (no int64_t overflow)
given; prior to saturation mac.l:
mach = 0xffff8000 macl = 0x00000000
@r0 = 0xffffffff @r1 = 0x00000001
expected saturation mac.l result:
mach = 0xffff8000 macl = 0x00000000
qemu saturation mac.l result (prior to this commit):
mach = 0xffff7fff macl = 0xffffffff
Test case 2: (int64_t addition overflow)
given; prior to saturation mac.l:
mach = 0x80000000 macl = 0x00000000
@r0 = 0xffffffff @r1 = 0x00000001
expected saturation mac.l result:
mach = 0xffff8000 macl = 0x00000000
qemu saturation mac.l result (prior to this commit):
mach = 0xffff7fff macl = 0xffffffff
Test case 3: (int64_t addition overflow)
given; prior to saturation mac.l:
mach = 0x7fffffff macl = 0x00000000
@r0 = 0x7fffffff @r1 = 0x7fffffff
expected saturation mac.l result:
mach = 0x00007fff macl = 0xffffffff
qemu saturation mac.l result (prior to this commit):
mach = 0xfffffffe macl = 0x00000001
All of the above also matches the description of MAC.L as documented
in cd00147165-sh-4-32-bit-cpu-core-architecture-stmicroelectronics.pdf
---
target/sh4/op_helper.c | 45 ++++++++++++++++++++++++++++++++----------
1 file changed, 35 insertions(+), 10 deletions(-)
diff --git a/target/sh4/op_helper.c b/target/sh4/op_helper.c
index 4559d0d376..a3eb2f5281 100644
--- a/target/sh4/op_helper.c
+++ b/target/sh4/op_helper.c
@@ -160,18 +160,43 @@ void helper_ocbi(CPUSH4State *env, uint32_t address)
void helper_macl(CPUSH4State *env, uint32_t arg0, uint32_t arg1)
{
- int64_t res;
-
- res = ((uint64_t) env->mach << 32) | env->macl;
- res += (int64_t) (int32_t) arg0 *(int64_t) (int32_t) arg1;
- env->mach = (res >> 32) & 0xffffffff;
- env->macl = res & 0xffffffff;
+ int32_t value0 = (int32_t)arg0;
+ int32_t value1 = (int32_t)arg1;
+ int64_t mul = ((int64_t)value0) * ((int64_t)value1);
+ int64_t mac = (((uint64_t)env->mach) << 32) | env->macl;
+ int64_t result = mac + mul;
+ /* Perform 48-bit saturation arithmetic if the S flag is set */
if (env->sr & (1u << SR_S)) {
- if (res < 0)
- env->mach |= 0xffff0000;
- else
- env->mach &= 0x00007fff;
+ /*
+ * The following xor/and expression is necessary to detect an
+ * overflow in MSB of res; this is logic necessary because the
+ * sign bit of `mac + mul` may overflow. The MAC unit on real
+ * SH-4 hardware has carry/saturation logic that is equivalent
+ * to the following:
+ */
+ const int64_t upper_bound = ((1ull << 47) - 1);
+ const int64_t lower_bound = -((1ull << 47) - 0);
+
+ if (((((result ^ mac) & (result ^ mul)) >> 63) & 1) == 1) {
+ /* An overflow occured during 64-bit addition */
+ if (((mac >> 63) & 1) == 0) {
+ result = upper_bound;
+ } else {
+ result = lower_bound;
+ }
+ } else {
+ /* An overflow did not occur during 64-bit addition */
+ if (result > upper_bound) {
+ result = upper_bound;
+ } else if (result < lower_bound) {
+ result = lower_bound;
+ } else {
+ /* leave result unchanged */
+ }
+ }
}
+ env->macl = result;
+ env->mach = result >> 32;
}
void helper_macw(CPUSH4State *env, uint32_t arg0, uint32_t arg1)
--
2.41.0
- [PATCH] sh4: mac.l: implement saturation arithmetic logic,
Zack Buhman <=