std/softfloat: return early for zero operands, merge the f64 add/sub fast paths and add a faster f32_from_u32

f64_add and f64_sub now share one fast path for normal operands
(f64_add_normal) and both return early when one operand is zero. The f32
add and sub do the same, and f32_from_u32 converts values below 2^24
directly. All results are unchanged: the sweep, the old-vs-new
differential test and the ops harness report zero mismatches.
This commit is contained in:
Claude committed 2026-10-27 10:18:57 +00:00
1 parent 49356e06e4
commit ab5930ffbb
4 files changed
+195 -269

No files matched your search

+75 -124
View File
@@ -41206,14 +41206,7 @@ wuffs_softfloat__arith__f64_propagate_nan(
WUFFS_BASE__GENERATED_C_CODE
static uint64_t
wuffs_softfloat__arith__f64_add_normal_mags(
const wuffs_softfloat__arith* self,
uint64_t a_a,
uint64_t a_b);
WUFFS_BASE__GENERATED_C_CODE
static uint64_t
wuffs_softfloat__arith__f64_sub_normal_mags(
wuffs_softfloat__arith__f64_add_normal(
const wuffs_softfloat__arith* self,
uint64_t a_a,
uint64_t a_b);
@@ -41897,13 +41890,14 @@ wuffs_softfloat__arith__f64_add(
uint64_t v_z = 0;
if ((((uint64_t)(((a_a >> 52u) & 2047u) - 1u)) < 2046u) && (((uint64_t)(((a_b >> 52u) & 2047u) - 1u)) < 2046u)) {
if (((a_a ^ a_b) >> 63u) == 0u) {
return wuffs_softfloat__arith__f64_add_normal_mags(self, a_a, a_b);
}
v_z = wuffs_softfloat__arith__f64_sub_normal_mags(self, a_a, a_b);
v_z = wuffs_softfloat__arith__f64_add_normal(self, a_a, a_b);
if (v_z != 18446744073709551615u) {
return v_z;
}
} else if (((a_b & 9223372036854775807u) == 0u) && (((uint64_t)((a_a & 9223372036854775807u) - 1u)) < 9218868437227405311u)) {
return a_a;
} else if (((a_a & 9223372036854775807u) == 0u) && (((uint64_t)((a_b & 9223372036854775807u) - 1u)) < 9218868437227405311u)) {
return a_b;
}
if (((a_a ^ a_b) >> 63u) == 0u) {
return wuffs_softfloat__arith__f64_add_mags(self, a_a, a_b, (a_a & 9223372036854775808u));
@@ -41930,13 +41924,14 @@ wuffs_softfloat__arith__f64_sub(
uint64_t v_z = 0;
if ((((uint64_t)(((a_a >> 52u) & 2047u) - 1u)) < 2046u) && (((uint64_t)(((a_b >> 52u) & 2047u) - 1u)) < 2046u)) {
if (((a_a ^ a_b) >> 63u) != 0u) {
return wuffs_softfloat__arith__f64_add_normal_mags(self, a_a, a_b);
}
v_z = wuffs_softfloat__arith__f64_sub_normal_mags(self, a_a, a_b);
v_z = wuffs_softfloat__arith__f64_add_normal(self, a_a, (a_b ^ 9223372036854775808u));
if (v_z != 18446744073709551615u) {
return v_z;
}
} else if (((a_b & 9223372036854775807u) == 0u) && (((uint64_t)((a_a & 9223372036854775807u) - 1u)) < 9218868437227405311u)) {
return a_a;
} else if (((a_a & 9223372036854775807u) == 0u) && (((uint64_t)((a_b & 9223372036854775807u) - 1u)) < 9218868437227405311u)) {
return (a_b ^ 9223372036854775808u);
}
if (((a_a ^ a_b) >> 63u) == 0u) {
return wuffs_softfloat__arith__f64_sub_mags(self, a_a, a_b, (a_a & 9223372036854775808u));
@@ -41944,133 +41939,71 @@ wuffs_softfloat__arith__f64_sub(
return wuffs_softfloat__arith__f64_add_mags(self, a_a, a_b, (a_a & 9223372036854775808u));
}
// -------- func softfloat.arith.f64_add_normal_mags
// -------- func softfloat.arith.f64_add_normal
WUFFS_BASE__GENERATED_C_CODE
static uint64_t
wuffs_softfloat__arith__f64_add_normal_mags(
wuffs_softfloat__arith__f64_add_normal(
const wuffs_softfloat__arith* self,
uint64_t a_a,
uint64_t a_b) {
uint32_t v_ea = 0;
uint32_t v_eb = 0;
uint64_t v_ma = 0;
uint64_t v_mb = 0;
uint64_t v_x = 0;
uint64_t v_y = 0;
uint32_t v_ex = 0;
uint32_t v_d = 0;
uint64_t v_mx = 0;
uint64_t v_my = 0;
uint64_t v_z = 0;
uint64_t v_rb = 0;
v_ea = ((uint32_t)(((a_a >> 52u) & 2047u)));
v_eb = ((uint32_t)(((a_b >> 52u) & 2047u)));
v_ma = (((a_a & 4503599627370495u) | 4503599627370496u) << 9u);
v_mb = (((a_b & 4503599627370495u) | 4503599627370496u) << 9u);
if (v_ea < v_eb) {
v_d = (((uint32_t)(v_eb - v_ea)) & 2047u);
v_ea = v_eb;
v_z = v_ma;
v_ma = v_mb;
v_mb = v_z;
} else {
v_d = (((uint32_t)(v_ea - v_eb)) & 2047u);
}
if (v_d >= 63u) {
v_mb = 1u;
} else if (v_d > 0u) {
if (((uint64_t)(v_mb << (64u - v_d))) != 0u) {
v_mb = ((v_mb >> v_d) | 1u);
} else {
v_mb = (v_mb >> v_d);
}
}
v_z = (v_ma + v_mb);
if (v_z >= 4611686018427387904u) {
v_ea += 1u;
v_z = ((v_z >> 1u) | (v_z & 1u));
}
v_rb = (v_z & 511u);
v_z = (((uint64_t)(v_z + 256u)) >> 9u);
if (v_rb == 256u) {
v_z &= 18446744073709551614u;
}
v_z = ((uint64_t)(((uint64_t)(((uint64_t)(((uint32_t)(v_ea - 1u)))) << 52u)) + v_z));
if (v_z >= 9218868437227405312u) {
v_z = 9218868437227405312u;
}
return ((a_a & 9223372036854775808u) | v_z);
}
// -------- func softfloat.arith.f64_sub_normal_mags
WUFFS_BASE__GENERATED_C_CODE
static uint64_t
wuffs_softfloat__arith__f64_sub_normal_mags(
const wuffs_softfloat__arith* self,
uint64_t a_a,
uint64_t a_b) {
uint64_t v_sign = 0;
uint32_t v_ea = 0;
uint32_t v_eb = 0;
uint64_t v_fa = 0;
uint64_t v_fb = 0;
uint64_t v_ma = 0;
uint64_t v_mb = 0;
uint32_t v_d = 0;
uint32_t v_s = 0;
uint64_t v_z = 0;
uint64_t v_rb = 0;
v_sign = (a_a & 9223372036854775808u);
v_ea = ((uint32_t)(((a_a >> 52u) & 2047u)));
v_eb = ((uint32_t)(((a_b >> 52u) & 2047u)));
v_fa = (a_a & 4503599627370495u);
v_fb = (a_b & 4503599627370495u);
if (v_ea == v_eb) {
if (v_fa == v_fb) {
v_x = a_a;
v_y = a_b;
if ((v_y & 9223372036854775807u) > (v_x & 9223372036854775807u)) {
v_x = a_b;
v_y = a_a;
}
v_ex = ((uint32_t)(((v_x >> 52u) & 2047u)));
v_d = ((uint32_t)(v_ex - ((uint32_t)(((v_y >> 52u) & 2047u)))));
if (v_d > 63u) {
v_d = 63u;
}
v_mx = (((v_x & 4503599627370495u) | 4503599627370496u) << 10u);
v_my = (((v_y & 4503599627370495u) | 4503599627370496u) << 10u);
if ((v_my & ((uint64_t)((((uint64_t)(1u)) << (v_d & 63u)) - 1u))) != 0u) {
v_my = ((v_my >> (v_d & 63u)) | 1u);
} else {
v_my = (v_my >> (v_d & 63u));
}
if (((v_x ^ v_y) >> 63u) == 0u) {
v_z = (v_mx + v_my);
if (v_z >= 9223372036854775808u) {
v_ex += 1u;
v_z = ((v_z >> 1u) | (v_z & 1u));
}
} else {
v_z = ((uint64_t)(v_mx - v_my));
if (v_z == 0u) {
return 0u;
} else if (v_fa < v_fb) {
v_sign ^= 9223372036854775808u;
v_z = ((uint64_t)(v_fb - v_fa));
} else {
v_z = ((uint64_t)(v_fa - v_fb));
} else if (v_z < 4611686018427387904u) {
v_s = (((uint32_t)(wuffs_softfloat__arith__clz64(self, v_z) - 1u)) & 63u);
if (v_s >= v_ex) {
return 18446744073709551615u;
}
v_ex = (((uint32_t)(v_ex - v_s)) & 2047u);
v_z = ((uint64_t)(v_z << v_s));
}
v_s = (((uint32_t)(wuffs_softfloat__arith__clz64(self, v_z) - 11u)) & 63u);
if (v_s >= v_ea) {
return 18446744073709551615u;
}
return (v_sign | ((uint64_t)(((uint64_t)(((uint64_t)(((uint32_t)(((uint32_t)(v_ea - v_s)) - 1u)))) << 52u)) + ((uint64_t)(v_z << v_s)))));
}
v_ma = ((v_fa | 4503599627370496u) << 10u);
v_mb = ((v_fb | 4503599627370496u) << 10u);
if (v_ea < v_eb) {
v_sign ^= 9223372036854775808u;
v_d = (((uint32_t)(v_eb - v_ea)) & 2047u);
v_ea = v_eb;
v_z = v_ma;
v_ma = v_mb;
v_mb = v_z;
} else {
v_d = (((uint32_t)(v_ea - v_eb)) & 2047u);
}
if (v_d >= 63u) {
v_mb = 1u;
} else if (((uint64_t)(v_mb << ((64u - v_d) & 63u))) != 0u) {
v_mb = ((v_mb >> v_d) | 1u);
} else {
v_mb = (v_mb >> v_d);
}
v_z = ((uint64_t)(v_ma - v_mb));
v_s = (((uint32_t)(wuffs_softfloat__arith__clz64(self, v_z) - 1u)) & 63u);
if (v_s >= v_ea) {
return 18446744073709551615u;
}
v_ea = (((uint32_t)(v_ea - v_s)) & 2047u);
v_z = ((uint64_t)(v_z << v_s));
v_rb = (v_z & 1023u);
v_z = (((uint64_t)(v_z + 512u)) >> 10u);
if (v_rb == 512u) {
v_z &= 18446744073709551614u;
}
return (v_sign | ((uint64_t)(((uint64_t)(((uint64_t)(((uint32_t)(v_ea - 1u)))) << 52u)) + v_z)));
v_z = ((uint64_t)(((uint64_t)(((uint64_t)(((uint32_t)(v_ex - 1u)))) << 52u)) + v_z));
if (v_z >= 9218868437227405312u) {
v_z = 9218868437227405312u;
}
return ((v_x & 9223372036854775808u) | v_z);
}
// -------- func softfloat.arith.f64_add_mags
@@ -42832,6 +42765,10 @@ wuffs_softfloat__arith__f32_add(
if (v_z != 4294967295u) {
return v_z;
}
} else if (((a_b & 2147483647u) == 0u) && (((uint32_t)((a_a & 2147483647u) - 1u)) < 2139095039u)) {
return a_a;
} else if (((a_a & 2147483647u) == 0u) && (((uint32_t)((a_b & 2147483647u) - 1u)) < 2139095039u)) {
return a_b;
}
return wuffs_softfloat__arith__slow_f32_add(self, a_a, a_b);
}
@@ -42862,6 +42799,10 @@ wuffs_softfloat__arith__f32_sub(
if (v_z != 4294967295u) {
return v_z;
}
} else if (((a_b & 2147483647u) == 0u) && (((uint32_t)((a_a & 2147483647u) - 1u)) < 2139095039u)) {
return a_a;
} else if (((a_a & 2147483647u) == 0u) && (((uint32_t)((a_b & 2147483647u) - 1u)) < 2139095039u)) {
return (a_b ^ 2147483648u);
}
return wuffs_softfloat__arith__slow_f32_sub(self, a_a, a_b);
}
@@ -43334,6 +43275,16 @@ wuffs_softfloat__arith__f32_from_u32(
return 0;
}
uint32_t v_c = 0;
if (a_a == 0u) {
return 0u;
} else if (a_a < 16777216u) {
v_c = wuffs_softfloat__arith__clz32(self, a_a);
if (v_c >= 8u) {
return ((uint32_t)(((uint32_t)(((uint32_t)(157u - v_c)) << 23u)) + ((uint32_t)(a_a << (v_c - 8u)))));
}
}
return wuffs_softfloat__arith__f64_to_f32(self, wuffs_softfloat__arith__f64_from_u32(self, a_a));
}