snailed
/
void-packages
2
0
Fork 0
Code Issues 1 Pull Requests 1 Actions Activity
void-packages/srcpkgs/nodejs/patches/ppc-fixes-for-older-models....

973 lines
34 KiB
Diff
Raw Blame History

Fix PowerPC CPU detection and codegen to work with more processors.
This patch defines the correct optional Power ISA features that the
PPC code generator needs in order to run without crashing on v2.01
and older CPUs such as PPC 970 (G5) or NXP e6500, and to run more
efficiently on CPUs with features that weren't being used before.
PowerPC ISA v2.01 and older CPUs don't have FP round to int instructions,
and PowerPC ISA v2.06 and older are missing support for unsigned 64-bit
to/from double, as well as integer to/from single-precision float.
Use the current PPC_5_PLUS CPU feature to determine whether to generate
FP round to int, and use the PPC_7_PLUS feature to determine whether
to use the v2.06 ISA instructions or whether to generate an alternate
generic PPC sequence to handle the cases of 64-bit unsigned integer
to/from floating point, integers to single-precision floating point,
and loading and storing 64-bit integers with byte reversal.
Add a new PPC_7_PLUS_NXP feature for the popcnt and ldbrx/stdbrx
opcodes added in Power ISA v2.06, which are also present in the NXP
e5500 and e6500 cores, which are otherwise missing many of the
features added since v2.01. This enables NXP cores to use a few
more features. Additionally, bring back the ISELECT feature flag,
which is also supported by NXP cores, including older ones, and
has its own AT_HWCAP2 feature flag in Linux.
By defining a new ICACHE_SNOOP feature bit to replace the use of
PPC_6_PLUS, the meaning of the instruction cache flushing fast path,
and the CPUs that can use it, is more clearly defined. In addition,
for the other PowerPC chips, the loop to flush the data and instruction
cache blocks has been split into two loops, with a single "sync" and
"isync" after each loop, which should be more efficient, and also handles
the few CPUs with differing data and instruction cache line sizes.
In the macro assembler methods, in addition to providing an alternate
path for FP conversion opcodes added in POWER7 (ISA v2.06), unnecessary
instructions to move sp down and then immediately back up were replaced
with negative offsets from the current sp. This should be faster, and also
sp is supposed to point to a back chain at all times (V8 may not do this).
--- a/deps/v8/src/base/cpu.cc 2022-02-08 04:37:48.000000000 -0800
+++ b/deps/v8/src/base/cpu.cc 2022-02-19 14:38:37.997161835 -0800
@@ -14,15 +14,13 @@
#if V8_OS_LINUX
#include <linux/auxvec.h> // AT_HWCAP
#endif
-#if V8_GLIBC_PREREQ(2, 16)
+#if V8_GLIBC_PREREQ(2, 16) || \
+ (V8_OS_LINUX && (V8_HOST_ARCH_PPC || V8_HOST_ARCH_PPC64))
#include <sys/auxv.h> // getauxval()
#endif
#if V8_OS_QNX
#include <sys/syspage.h> // cpuinfo
#endif
-#if V8_OS_LINUX && (V8_HOST_ARCH_PPC || V8_HOST_ARCH_PPC64)
-#include <elf.h>
-#endif
#if V8_OS_AIX
#include <sys/systemcfg.h> // _system_configuration
#ifndef POWER_8
@@ -772,56 +770,55 @@
#ifndef USE_SIMULATOR
#if V8_OS_LINUX
- // Read processor info from /proc/self/auxv.
- char* auxv_cpu_type = nullptr;
- FILE* fp = base::Fopen("/proc/self/auxv", "r");
- if (fp != nullptr) {
-#if V8_TARGET_ARCH_PPC64
- Elf64_auxv_t entry;
-#else
- Elf32_auxv_t entry;
-#endif
- for (;;) {
- size_t n = fread(&entry, sizeof(entry), 1, fp);
- if (n == 0 || entry.a_type == AT_NULL) {
- break;
- }
- switch (entry.a_type) {
- case AT_PLATFORM:
- auxv_cpu_type = reinterpret_cast<char*>(entry.a_un.a_val);
- break;
- case AT_ICACHEBSIZE:
- icache_line_size_ = entry.a_un.a_val;
- break;
- case AT_DCACHEBSIZE:
- dcache_line_size_ = entry.a_un.a_val;
- break;
- }
- }
- base::Fclose(fp);
- }
-
- part_ = -1;
- if (auxv_cpu_type) {
- if (strcmp(auxv_cpu_type, "power10") == 0) {
- part_ = kPPCPower10;
- } else if (strcmp(auxv_cpu_type, "power9") == 0) {
- part_ = kPPCPower9;
- } else if (strcmp(auxv_cpu_type, "power8") == 0) {
- part_ = kPPCPower8;
- } else if (strcmp(auxv_cpu_type, "power7") == 0) {
- part_ = kPPCPower7;
- } else if (strcmp(auxv_cpu_type, "power6") == 0) {
- part_ = kPPCPower6;
- } else if (strcmp(auxv_cpu_type, "power5") == 0) {
- part_ = kPPCPower5;
- } else if (strcmp(auxv_cpu_type, "ppc970") == 0) {
- part_ = kPPCG5;
- } else if (strcmp(auxv_cpu_type, "ppc7450") == 0) {
- part_ = kPPCG4;
- } else if (strcmp(auxv_cpu_type, "pa6t") == 0) {
- part_ = kPPCPA6T;
- }
+ // Read processor info from getauxval() (needs at least glibc 2.18 or musl).
+ icache_line_size_ = static_cast<int>(getauxval(AT_ICACHEBSIZE));
+ dcache_line_size_ = static_cast<int>(getauxval(AT_DCACHEBSIZE));
+ const unsigned long hwcap = getauxval(AT_HWCAP);
+ const unsigned long hwcap2 = getauxval(AT_HWCAP2);
+ const char* platform = reinterpret_cast<const char*>(getauxval(AT_PLATFORM));
+
+ // NOTE: AT_HWCAP ISA version bits aren't cumulative, so it's necessary
+ // to compare against a mask of all supported versions and CPUs, up to
+ // ISA v2.06, which *is* set for later CPUs. In contrast, the AT_HWCAP2
+ // ISA version bits from v2.07 onward are set cumulatively, so POWER10
+ // will set the ISA version bits from v2.06 (in AT_HWCAP) through v3.1.
+
+ // i-cache coherency requires Power ISA v2.02 or later; has its own flag.
+ has_icache_snoop_ = (hwcap & PPC_FEATURE_ICACHE_SNOOP);
+
+ // requires Power ISA v2.03 or later, or the HAS_ISEL bit (e.g. e6500).
+ has_isel_ = (hwcap & (PPC_FEATURE_POWER5_PLUS | PPC_FEATURE_ARCH_2_05 |
+ PPC_FEATURE_PA6T | PPC_FEATURE_POWER6_EXT | PPC_FEATURE_ARCH_2_06)) ||
+ (hwcap2 & PPC_FEATURE2_HAS_ISEL);
+
+ // hwcap mask for older 64-bit PPC CPUs with Altivec, e.g. G5, Cell.
+ static const unsigned long kHwcapMaskPPCG5 =
+ (PPC_FEATURE_64 | PPC_FEATURE_HAS_ALTIVEC);
+
+ if (hwcap2 & PPC_FEATURE2_ARCH_3_1) {
+ part_ = kPPCPower10;
+ } else if (hwcap2 & PPC_FEATURE2_ARCH_3_00) {
+ part_ = kPPCPower9;
+ } else if (hwcap2 & PPC_FEATURE2_ARCH_2_07) {
+ part_ = kPPCPower8;
+ } else if (hwcap & PPC_FEATURE_ARCH_2_06) {
+ part_ = kPPCPower7;
+ } else if (hwcap & PPC_FEATURE_ARCH_2_05) {
+ part_ = kPPCPower6;
+ } else if (hwcap & (PPC_FEATURE_POWER5 | PPC_FEATURE_POWER5_PLUS)) {
+ part_ = kPPCPower5;
+ } else if (hwcap & PPC_FEATURE_PA6T) {
+ part_ = kPPCPA6T;
+ } else if (strcmp(platform, "ppce6500") == 0) {
+ part_ = kPPCE6500;
+ } else if (strcmp(platform, "ppce5500") == 0) {
+ part_ = kPPCE5500;
+ } else if ((hwcap & kHwcapMaskPPCG5) == kHwcapMaskPPCG5) {
+ part_ = kPPCG5;
+ } else if (hwcap & PPC_FEATURE_HAS_ALTIVEC) {
+ part_ = kPPCG4;
+ } else {
+ part_ = kPPCG3;
}
#elif V8_OS_AIX
@@ -842,9 +839,13 @@
part_ = kPPCPower6;
break;
case POWER_5:
+ default:
part_ = kPPCPower5;
break;
}
+
+ has_icache_snoop_ = true;
+ has_isel_ = (part_ != kPPCPower5); // isel was added in POWER5+ (v2.03)
#endif // V8_OS_AIX
#endif // !USE_SIMULATOR
#endif // V8_HOST_ARCH_PPC || V8_HOST_ARCH_PPC64
--- a/deps/v8/src/base/cpu.h 2022-02-08 04:37:48.000000000 -0800
+++ b/deps/v8/src/base/cpu.h 2022-02-19 14:32:09.831579133 -0800
@@ -71,9 +71,12 @@
kPPCPower8,
kPPCPower9,
kPPCPower10,
+ kPPCG3,
kPPCG4,
kPPCG5,
- kPPCPA6T
+ kPPCPA6T,
+ kPPCE5500,
+ kPPCE6500
};
// General features
@@ -119,6 +122,10 @@
bool is_fp64_mode() const { return is_fp64_mode_; }
bool has_msa() const { return has_msa_; }
+ // PowerPC features
+ bool has_icache_snoop() const { return has_icache_snoop_; }
+ bool has_isel() const { return has_isel_; }
+
private:
#if defined(V8_OS_STARBOARD)
bool StarboardDetectCPU();
@@ -166,6 +173,8 @@
bool has_non_stop_time_stamp_counter_;
bool is_running_in_vm_;
bool has_msa_;
+ bool has_icache_snoop_;
+ bool has_isel_;
};
} // namespace base
--- a/deps/v8/src/builtins/ppc/builtins-ppc.cc 2022-02-08 04:37:48.000000000 -0800
+++ b/deps/v8/src/builtins/ppc/builtins-ppc.cc 2022-02-19 15:18:36.373031457 -0800
@@ -2823,7 +2823,7 @@
__ lbz(scratch, MemOperand(scratch, 0));
__ cmpi(scratch, Operand::Zero());
- if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ if (CpuFeatures::IsSupported(ISELECT)) {
__ Move(scratch, thunk_ref);
__ isel(eq, scratch, function_address, scratch);
} else {
--- a/deps/v8/src/codegen/ppc/macro-assembler-ppc.cc 2022-02-19 18:45:44.687593194 -0800
+++ b/deps/v8/src/codegen/ppc/macro-assembler-ppc.cc 2022-02-19 18:33:15.060674389 -0800
@@ -925,13 +925,25 @@
void TurboAssembler::ConvertIntToFloat(Register src, DoubleRegister dst) {
MovIntToDouble(dst, src, r0);
- fcfids(dst, dst);
+
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ fcfids(dst, dst);
+ } else {
+ fcfid(dst, dst);
+ frsp(dst, dst);
+ }
}
void TurboAssembler::ConvertUnsignedIntToFloat(Register src,
DoubleRegister dst) {
MovUnsignedIntToDouble(dst, src, r0);
- fcfids(dst, dst);
+
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ fcfids(dst, dst);
+ } else {
+ fcfid(dst, dst);
+ frsp(dst, dst);
+ }
}
#if V8_TARGET_ARCH_PPC64
@@ -943,20 +955,52 @@
void TurboAssembler::ConvertUnsignedInt64ToFloat(Register src,
DoubleRegister double_dst) {
- MovInt64ToDouble(double_dst, src);
- fcfidus(double_dst, double_dst);
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ MovInt64ToDouble(double_dst, src);
+ fcfidus(double_dst, double_dst);
+ } else {
+ ConvertUnsignedInt64ToDouble(src, double_dst);
+ frsp(double_dst, double_dst);
+ }
}
void TurboAssembler::ConvertUnsignedInt64ToDouble(Register src,
DoubleRegister double_dst) {
- MovInt64ToDouble(double_dst, src);
- fcfidu(double_dst, double_dst);
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ MovInt64ToDouble(double_dst, src);
+ fcfidu(double_dst, double_dst);
+ } else {
+ Label negative;
+ Label done;
+ cmpi(src, Operand::Zero());
+ blt(&negative);
+ std(src, MemOperand(sp, -kDoubleSize));
+ nop(GROUP_ENDING_NOP); // LHS/RAW optimization
+ lfd(double_dst, MemOperand(sp, -kDoubleSize));
+ fcfid(double_dst, double_dst);
+ b(&done);
+ bind(&negative);
+ // Note: GCC saves the lowest bit, then ORs it after shifting right 1 bit,
+ // presumably for better rounding. This version only shifts right 1 bit.
+ srdi(r0, src, Operand(1));
+ std(r0, MemOperand(sp, -kDoubleSize));
+ nop(GROUP_ENDING_NOP); // LHS/RAW optimization
+ lfd(double_dst, MemOperand(sp, -kDoubleSize));
+ fcfid(double_dst, double_dst);
+ fadd(double_dst, double_dst, double_dst);
+ bind(&done);
+ }
}
void TurboAssembler::ConvertInt64ToFloat(Register src,
DoubleRegister double_dst) {
MovInt64ToDouble(double_dst, src);
- fcfids(double_dst, double_dst);
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ fcfids(double_dst, double_dst);
+ } else {
+ fcfid(double_dst, double_dst);
+ frsp(double_dst, double_dst);
+ }
}
#endif
@@ -986,15 +1030,56 @@
void TurboAssembler::ConvertDoubleToUnsignedInt64(
const DoubleRegister double_input, const Register dst,
const DoubleRegister double_dst, FPRoundingMode rounding_mode) {
- if (rounding_mode == kRoundToZero) {
- fctiduz(double_dst, double_input);
+ if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ if (rounding_mode == kRoundToZero) {
+ fctiduz(double_dst, double_input);
+ } else {
+ SetRoundingMode(rounding_mode);
+ fctidu(double_dst, double_input);
+ ResetRoundingMode();
+ }
+
+ MovDoubleToInt64(dst, double_dst);
} else {
- SetRoundingMode(rounding_mode);
- fctidu(double_dst, double_input);
- ResetRoundingMode();
+ Label safe_size;
+ Label done;
+ mov(dst, Operand(1593835520)); // bit pattern for 2^63 as a float
+ stw(dst, MemOperand(sp, -kFloatSize));
+ nop(GROUP_ENDING_NOP); // LHS/RAW optimization
+ lfs(double_dst, MemOperand(sp, -kFloatSize));
+ fcmpu(double_input, double_dst);
+ blt(&safe_size);
+ // Subtract 2^63, then OR the top bit of the uint64 to add back
+ fsub(double_dst, double_input, double_dst);
+ if (rounding_mode == kRoundToZero) {
+ fctidz(double_dst, double_dst);
+ } else {
+ SetRoundingMode(rounding_mode);
+ fctid(double_dst, double_dst);
+ ResetRoundingMode();
+ }
+ // set r0 to -1, then clear all but the MSB.
+ mov(r0, Operand(-1));
+ rldicr(r0, r0, 0, 0);
+ stfd(double_dst, MemOperand(sp, -kDoubleSize));
+ nop(GROUP_ENDING_NOP); // LHS/RAW optimization
+ ld(dst, MemOperand(sp, -kDoubleSize));
+ orx(dst, dst, r0);
+ b(&done);
+ // Handling for values smaller than 2^63.
+ bind(&safe_size);
+ if (rounding_mode == kRoundToZero) {
+ fctidz(double_dst, double_input);
+ } else {
+ SetRoundingMode(rounding_mode);
+ fctid(double_dst, double_input);
+ ResetRoundingMode();
+ }
+ stfd(double_dst, MemOperand(sp, -kDoubleSize));
+ nop(GROUP_ENDING_NOP); // LHS/RAW optimization
+ ld(dst, MemOperand(sp, -kDoubleSize));
+ bind(&done);
}
-
- MovDoubleToInt64(dst, double_dst);
}
#endif
@@ -2459,19 +2544,17 @@
}
#endif
- addi(sp, sp, Operand(-kDoubleSize));
#if V8_TARGET_ARCH_PPC64
mov(scratch, Operand(litVal.ival));
- std(scratch, MemOperand(sp));
+ std(scratch, MemOperand(sp, -kDoubleSize));
#else
LoadIntLiteral(scratch, litVal.ival[0]);
- stw(scratch, MemOperand(sp, 0));
+ stw(scratch, MemOperand(sp, -kDoubleSize));
LoadIntLiteral(scratch, litVal.ival[1]);
- stw(scratch, MemOperand(sp, 4));
+ stw(scratch, MemOperand(sp, -kDoubleSize + 4));
#endif
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(result, MemOperand(sp, 0));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(result, MemOperand(sp, -kDoubleSize));
}
void TurboAssembler::MovIntToDouble(DoubleRegister dst, Register src,
@@ -2485,18 +2568,16 @@
#endif
DCHECK(src != scratch);
- subi(sp, sp, Operand(kDoubleSize));
#if V8_TARGET_ARCH_PPC64
extsw(scratch, src);
- std(scratch, MemOperand(sp, 0));
+ std(scratch, MemOperand(sp, -kDoubleSize));
#else
srawi(scratch, src, 31);
- stw(scratch, MemOperand(sp, Register::kExponentOffset));
- stw(src, MemOperand(sp, Register::kMantissaOffset));
+ stw(scratch, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
+ stw(src, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
#endif
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp, 0));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
void TurboAssembler::MovUnsignedIntToDouble(DoubleRegister dst, Register src,
@@ -2510,18 +2591,16 @@
#endif
DCHECK(src != scratch);
- subi(sp, sp, Operand(kDoubleSize));
#if V8_TARGET_ARCH_PPC64
clrldi(scratch, src, Operand(32));
- std(scratch, MemOperand(sp, 0));
+ std(scratch, MemOperand(sp, -kDoubleSize));
#else
li(scratch, Operand::Zero());
- stw(scratch, MemOperand(sp, Register::kExponentOffset));
- stw(src, MemOperand(sp, Register::kMantissaOffset));
+ stw(scratch, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
+ stw(src, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
#endif
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp, 0));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
void TurboAssembler::MovInt64ToDouble(DoubleRegister dst,
@@ -2536,16 +2615,14 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
#if V8_TARGET_ARCH_PPC64
- std(src, MemOperand(sp, 0));
+ std(src, MemOperand(sp, -kDoubleSize));
#else
- stw(src_hi, MemOperand(sp, Register::kExponentOffset));
- stw(src, MemOperand(sp, Register::kMantissaOffset));
+ stw(src_hi, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
+ stw(src, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
#endif
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp, 0));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
#if V8_TARGET_ARCH_PPC64
@@ -2560,12 +2637,10 @@
return;
}
- subi(sp, sp, Operand(kDoubleSize));
- stw(src_hi, MemOperand(sp, Register::kExponentOffset));
- stw(src_lo, MemOperand(sp, Register::kMantissaOffset));
+ stw(src_hi, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
+ stw(src_lo, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
#endif
@@ -2580,12 +2655,10 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
- stfd(dst, MemOperand(sp));
- stw(src, MemOperand(sp, Register::kMantissaOffset));
+ stfd(dst, MemOperand(sp, -kDoubleSize));
+ stw(src, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
void TurboAssembler::InsertDoubleHigh(DoubleRegister dst, Register src,
@@ -2599,12 +2672,10 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
- stfd(dst, MemOperand(sp));
- stw(src, MemOperand(sp, Register::kExponentOffset));
+ stfd(dst, MemOperand(sp, -kDoubleSize));
+ stw(src, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfd(dst, MemOperand(sp));
- addi(sp, sp, Operand(kDoubleSize));
+ lfd(dst, MemOperand(sp, -kDoubleSize));
}
void TurboAssembler::MovDoubleLowToInt(Register dst, DoubleRegister src) {
@@ -2615,11 +2686,9 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
- stfd(src, MemOperand(sp));
+ stfd(src, MemOperand(sp, -kDoubleSize));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lwz(dst, MemOperand(sp, Register::kMantissaOffset));
- addi(sp, sp, Operand(kDoubleSize));
+ lwz(dst, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
}
void TurboAssembler::MovDoubleHighToInt(Register dst, DoubleRegister src) {
@@ -2631,11 +2700,9 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
- stfd(src, MemOperand(sp));
+ stfd(src, MemOperand(sp, -kDoubleSize));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lwz(dst, MemOperand(sp, Register::kExponentOffset));
- addi(sp, sp, Operand(kDoubleSize));
+ lwz(dst, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
}
void TurboAssembler::MovDoubleToInt64(
@@ -2650,32 +2717,26 @@
}
#endif
- subi(sp, sp, Operand(kDoubleSize));
- stfd(src, MemOperand(sp));
+ stfd(src, MemOperand(sp, -kDoubleSize));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
#if V8_TARGET_ARCH_PPC64
- ld(dst, MemOperand(sp, 0));
+ ld(dst, MemOperand(sp, -kDoubleSize));
#else
- lwz(dst_hi, MemOperand(sp, Register::kExponentOffset));
- lwz(dst, MemOperand(sp, Register::kMantissaOffset));
+ lwz(dst_hi, MemOperand(sp, -kDoubleSize + Register::kExponentOffset));
+ lwz(dst, MemOperand(sp, -kDoubleSize + Register::kMantissaOffset));
#endif
- addi(sp, sp, Operand(kDoubleSize));
}
void TurboAssembler::MovIntToFloat(DoubleRegister dst, Register src) {
- subi(sp, sp, Operand(kFloatSize));
- stw(src, MemOperand(sp, 0));
+ stw(src, MemOperand(sp, -kFloatSize));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lfs(dst, MemOperand(sp, 0));
- addi(sp, sp, Operand(kFloatSize));
+ lfs(dst, MemOperand(sp, -kFloatSize));
}
void TurboAssembler::MovFloatToInt(Register dst, DoubleRegister src) {
- subi(sp, sp, Operand(kFloatSize));
- stfs(src, MemOperand(sp, 0));
+ stfs(src, MemOperand(sp, -kFloatSize));
nop(GROUP_ENDING_NOP); // LHS/RAW optimization
- lwz(dst, MemOperand(sp, 0));
- addi(sp, sp, Operand(kFloatSize));
+ lwz(dst, MemOperand(sp, -kFloatSize));
}
void TurboAssembler::AddS64(Register dst, Register src, Register value, OEBit s,
@@ -3189,10 +3250,8 @@
}
#define MEM_LE_OP_LIST(V) \
- V(LoadU64, ldbrx) \
V(LoadU32, lwbrx) \
V(LoadU16, lhbrx) \
- V(StoreU64, stdbrx) \
V(StoreU32, stwbrx) \
V(StoreU16, sthbrx)
@@ -3214,6 +3273,37 @@
#undef MEM_LE_OP_FUNCTION
#undef MEM_LE_OP_LIST
+void TurboAssembler::LoadU64LE(Register dst, const MemOperand& mem,
+ Register scratch) {
+#ifdef V8_TARGET_BIG_ENDIAN
+ if (CpuFeatures::IsSupported(PPC_7_PLUS_NXP)) {
+ GenerateMemoryLEOperation(dst, mem, ldbrx);
+ } else {
+ lwbrx(dst, mem);
+ lwbrx(scratch, MemOperand(mem.ra(), mem.rb(), mem.offset() + 4));
+ rldicr(scratch, scratch, 32, 31);
+ orx(dst, dst, scratch);
+ }
+#else
+ LoadU64(dst, mem, scratch);
+#endif
+}
+
+void TurboAssembler::StoreU64LE(Register src, const MemOperand& mem,
+ Register scratch) {
+#ifdef V8_TARGET_BIG_ENDIAN
+ if (CpuFeatures::IsSupported(PPC_7_PLUS_NXP)) {
+ GenerateMemoryLEOperation(src, mem, stdbrx);
+ } else {
+ stwbrx(src, mem);
+ rldicl(scratch, src, 32, 32);
+ stwbrx(scratch, MemOperand(mem.ra(), mem.rb(), mem.offset() + 4));
+ }
+#else
+ StoreU64(src, mem, scratch);
+#endif
+}
+
void TurboAssembler::LoadS32LE(Register dst, const MemOperand& mem,
Register scratch) {
#ifdef V8_TARGET_BIG_ENDIAN
--- a/deps/v8/src/codegen/cpu-features.h 2022-02-19 21:19:15.982288690 -0800
+++ b/deps/v8/src/codegen/cpu-features.h 2022-02-19 21:22:43.071487369 -0800
@@ -52,11 +52,15 @@
MIPS_SIMD, // MSA instructions
#elif V8_TARGET_ARCH_PPC || V8_TARGET_ARCH_PPC64
+ PPC_5_PLUS,
PPC_6_PLUS,
PPC_7_PLUS,
PPC_8_PLUS,
PPC_9_PLUS,
PPC_10_PLUS,
+ ICACHE_SNOOP, // ISA v2.02 (POWER5)
+ ISELECT, // ISA v2.03 (POWER5+ and some NXP cores)
+ PPC_7_PLUS_NXP, // ISA v2.06 (POWER7 and NXP e5500/e6500)
#elif V8_TARGET_ARCH_S390X
FPU,
--- a/deps/v8/src/compiler/backend/ppc/instruction-selector-ppc.cc 2022-02-20 23:35:21.212337639 -0800
+++ b/deps/v8/src/compiler/backend/ppc/instruction-selector-ppc.cc 2022-02-20 23:36:20.925858840 -0800
@@ -2702,16 +2702,26 @@
// static
MachineOperatorBuilder::Flags
InstructionSelector::SupportedMachineOperatorFlags() {
- return MachineOperatorBuilder::kFloat32RoundDown |
- MachineOperatorBuilder::kFloat64RoundDown |
- MachineOperatorBuilder::kFloat32RoundUp |
- MachineOperatorBuilder::kFloat64RoundUp |
- MachineOperatorBuilder::kFloat32RoundTruncate |
- MachineOperatorBuilder::kFloat64RoundTruncate |
- MachineOperatorBuilder::kFloat64RoundTiesAway |
- MachineOperatorBuilder::kWord32Popcnt |
- MachineOperatorBuilder::kWord64Popcnt;
+ MachineOperatorBuilder::Flags flags = MachineOperatorBuilder::Flag::kNoFlags;
+ // FP rounding to integer instructions require Power ISA v2.02 or later.
+ if (CpuFeatures::IsSupported(PPC_5_PLUS)) {
+ flags |= MachineOperatorBuilder::kFloat32RoundDown |
+ MachineOperatorBuilder::kFloat64RoundDown |
+ MachineOperatorBuilder::kFloat32RoundUp |
+ MachineOperatorBuilder::kFloat64RoundUp |
+ MachineOperatorBuilder::kFloat32RoundTruncate |
+ MachineOperatorBuilder::kFloat64RoundTruncate |
+ MachineOperatorBuilder::kFloat64RoundTiesAway;
+ }
+ // Population count requires Power ISA v2.06, or NXP e5500/e6500.
+ if (CpuFeatures::IsSupported(PPC_7_PLUS_NXP)) {
+ flags |= MachineOperatorBuilder::kWord32Popcnt;
+#if V8_TARGET_ARCH_PPC64
+ flags |= MachineOperatorBuilder::kWord64Popcnt;
+#endif
+ }
// We omit kWord32ShiftIsSafe as s[rl]w use 0x3F as a mask rather than 0x1F.
+ return flags;
}
// static
--- a/deps/v8/src/compiler/backend/ppc/code-generator-ppc.cc 2022-02-20 23:35:21.216337741 -0800
+++ b/deps/v8/src/compiler/backend/ppc/code-generator-ppc.cc 2022-02-20 23:39:56.479351482 -0800
@@ -785,6 +785,7 @@
// Calculate a mask which has all bits set in the normal case, but has all
// bits cleared if we are speculatively executing the wrong PC.
__ CmpS64(kJavaScriptCallCodeStartRegister, scratch);
+ // TODO: is alternate sequence needed for CPUs without isel?
__ li(scratch, Operand::Zero());
__ notx(kSpeculationPoisonRegister, scratch);
__ isel(eq, kSpeculationPoisonRegister, kSpeculationPoisonRegister, scratch);
@@ -1823,6 +1824,7 @@
int crbit = v8::internal::Assembler::encode_crbit(
cr, static_cast<CRBit>(VXCVI % CRWIDTH));
__ mcrfs(cr, VXCVI); // extract FPSCR field containing VXCVI into cr7
+ // TODO: is alternate sequence needed for CPUs without isel?
__ li(kScratchReg, Operand(1));
__ ShiftLeftU64(kScratchReg, kScratchReg,
Operand(31)); // generate INT32_MIN.
@@ -1844,6 +1846,7 @@
int crbit = v8::internal::Assembler::encode_crbit(
cr, static_cast<CRBit>(VXCVI % CRWIDTH));
__ mcrfs(cr, VXCVI); // extract FPSCR field containing VXCVI into cr7
+ // TODO: is alternate sequence needed for CPUs without isel?
__ li(kScratchReg, Operand::Zero());
__ isel(i.OutputRegister(0), kScratchReg, i.OutputRegister(0), crbit);
}
@@ -1868,7 +1871,7 @@
cr, static_cast<CRBit>(VXCVI % CRWIDTH));
__ mcrfs(cr, VXCVI); // extract FPSCR field containing VXCVI into cr7
// Handle conversion failures (such as overflow).
- if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ if (CpuFeatures::IsSupported(ISELECT)) {
if (check_conversion) {
__ li(i.OutputRegister(1), Operand(1));
__ isel(i.OutputRegister(1), r0, i.OutputRegister(1), crbit);
@@ -1905,7 +1908,7 @@
int crbit = v8::internal::Assembler::encode_crbit(
cr, static_cast<CRBit>(VXCVI % CRWIDTH));
__ mcrfs(cr, VXCVI); // extract FPSCR field containing VXCVI into cr7
- if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ if (CpuFeatures::IsSupported(ISELECT)) {
__ li(i.OutputRegister(1), Operand(1));
__ isel(i.OutputRegister(1), r0, i.OutputRegister(1), crbit);
} else {
@@ -2168,12 +2171,67 @@
break;
}
case kPPC_LoadByteRev64: {
- ASSEMBLE_LOAD_INTEGER_RR(ldbrx);
+ // inlined version of ASSEMBLE_LOAD_INTEGER_RR()
+ Register result = i.OutputRegister();
+ AddressingMode mode = kMode_None;
+ MemOperand operand = i.MemoryOperand(&mode);
+ DCHECK_EQ(mode, kMode_MRR);
+ bool is_atomic = i.InputInt32(2);
+ if (CpuFeatures::IsSupported(PPC_7_PLUS_NXP)) {
+ __ ldbrx(result, operand);
+ } else {
+#ifdef V8_TARGET_BIG_ENDIAN
+ // low and high words from reversed perspective
+ MemOperand op_low = operand;
+ MemOperand op_high = MemOperand(operand.ra(), operand.rb(),
+ operand.offset() + 4);
+#else
+ // low and high words from reversed perspective
+ MemOperand op_high = operand;
+ MemOperand op_low = MemOperand(operand.ra(), operand.rb(),
+ operand.offset() + 4);
+#endif
+ Register temp1 = r0;
+ __ lwbrx(result, op_low);
+ __ lwbrx(temp1, op_high);
+ __ rldicr(temp1, temp1, 32, 31);
+ __ orx(result, result, temp1);
+ }
+ if (is_atomic) __ lwsync();
+ DCHECK_EQ(LeaveRC, i.OutputRCBit());
EmitWordLoadPoisoningIfNeeded(this, instr, i);
break;
}
case kPPC_StoreByteRev64: {
- ASSEMBLE_STORE_INTEGER_RR(stdbrx);
+ // inlined version of ASSEMBLE_STORE_INTEGER_RR()
+ size_t index = 0;
+ AddressingMode mode = kMode_None;
+ MemOperand operand = i.MemoryOperand(&mode, &index);
+ DCHECK_EQ(mode, kMode_MRR);
+ Register value = i.InputRegister(index);
+ bool is_atomic = i.InputInt32(3);
+ if (is_atomic) __ lwsync();
+ if (CpuFeatures::IsSupported(PPC_7_PLUS_NXP)) {
+ __ stdbrx(value, operand);
+ } else {
+#ifdef V8_TARGET_BIG_ENDIAN
+ // low and high words from reversed perspective
+ MemOperand op_low = operand;
+ MemOperand op_high = MemOperand(operand.ra(), operand.rb(),
+ operand.offset() + 4);
+#else
+ // low and high words from reversed perspective
+ MemOperand op_high = operand;
+ MemOperand op_low = MemOperand(operand.ra(), operand.rb(),
+ operand.offset() + 4);
+#endif
+ Register temp1 = r0;
+ __ stwbrx(value, op_low);
+ __ rldicl(temp1, value, 32, 32);
+ __ stwbrx(temp1, op_high);
+ }
+ if (is_atomic) __ sync();
+ DCHECK_EQ(LeaveRC, i.OutputRCBit());
break;
}
case kPPC_F64x2Splat: {
@@ -2911,11 +2969,13 @@
__ li(ip, Operand(1));
// Check if both lanes are 0, if so then return false.
__ vxor(kScratchSimd128Reg, kScratchSimd128Reg, kScratchSimd128Reg);
+ // TODO: is alternate sequence needed for CPUs without isel?
__ mtcrf(r0, fxm); // Clear cr6.
__ vcmpequd(kScratchSimd128Reg, src, kScratchSimd128Reg, SetRC);
__ isel(dst, r0, ip, bit_number);
break;
}
+// TODO: is alternate sequence needed for CPUs without isel?
#define SIMD_ALL_TRUE(opcode) \
Simd128Register src = i.InputSimd128Register(0); \
Register dst = i.OutputRegister(); \
@@ -3809,6 +3869,7 @@
ArchOpcode op = instr->arch_opcode();
condition = NegateFlagsCondition(condition);
+ // TODO: is alternate sequence needed for CPUs without isel?
__ li(kScratchReg, Operand::Zero());
__ isel(FlagsConditionToCondition(condition, op), kSpeculationPoisonRegister,
kScratchReg, kSpeculationPoisonRegister, cr0);
@@ -3922,7 +3983,7 @@
// Unnecessary for eq/lt & ne/ge since only FU bit will be set.
}
- if (CpuFeatures::IsSupported(PPC_7_PLUS)) {
+ if (CpuFeatures::IsSupported(ISELECT)) {
switch (cond) {
case eq:
case lt:
--- a/deps/v8/src/codegen/ppc/cpu-ppc.cc 2022-02-08 04:37:48.000000000 -0800
+++ b/deps/v8/src/codegen/ppc/cpu-ppc.cc 2022-02-20 17:02:17.900000785 -0800
@@ -8,14 +8,12 @@
#include "src/codegen/cpu-features.h"
-#define INSTR_AND_DATA_CACHE_COHERENCY PPC_6_PLUS
-
namespace v8 {
namespace internal {
void CpuFeatures::FlushICache(void* buffer, size_t size) {
#if !defined(USE_SIMULATOR)
- if (CpuFeatures::IsSupported(INSTR_AND_DATA_CACHE_COHERENCY)) {
+ if (CpuFeatures::IsSupported(ICACHE_SNOOP)) {
__asm__ __volatile__(
"sync \n"
"icbi 0, %0 \n"
@@ -26,25 +24,33 @@
return;
}
- const int kCacheLineSize = CpuFeatures::icache_line_size();
- intptr_t mask = kCacheLineSize - 1;
+ const int kInstrCacheLineSize = CpuFeatures::icache_line_size();
+ const int kDataCacheLineSize = CpuFeatures::dcache_line_size();
+ intptr_t ic_mask = kInstrCacheLineSize - 1;
+ intptr_t dc_mask = kDataCacheLineSize - 1;
byte* start =
- reinterpret_cast<byte*>(reinterpret_cast<intptr_t>(buffer) & ~mask);
+ reinterpret_cast<byte*>(reinterpret_cast<intptr_t>(buffer) & ~dc_mask);
byte* end = static_cast<byte*>(buffer) + size;
- for (byte* pointer = start; pointer < end; pointer += kCacheLineSize) {
- __asm__(
+ for (byte* pointer = start; pointer < end; pointer += kDataCacheLineSize) {
+ __asm__ __volatile__(
"dcbf 0, %0 \n"
- "sync \n"
- "icbi 0, %0 \n"
- "isync \n"
: /* no output */
: "r"(pointer));
}
+ __asm__ __volatile__("sync");
+ start =
+ reinterpret_cast<byte*>(reinterpret_cast<intptr_t>(buffer) & ~ic_mask);
+ for (byte* pointer = start; pointer < end; pointer += kInstrCacheLineSize) {
+ __asm__ __volatile__(
+ "icbi 0, %0 \n"
+ : /* no output */
+ : "r"(pointer));
+ }
+ __asm__ __volatile__("isync");
#endif // !USE_SIMULATOR
}
} // namespace internal
} // namespace v8
-#undef INSTR_AND_DATA_CACHE_COHERENCY
#endif // V8_TARGET_ARCH_PPC || V8_TARGET_ARCH_PPC64
--- ./deps/v8/src/codegen/ppc/assembler-ppc.cc.orig 2022-02-08 04:37:48.000000000 -0800
+++ ./deps/v8/src/codegen/ppc/assembler-ppc.cc 2022-02-20 17:20:25.019591225 -0800
@@ -65,6 +65,7 @@
void CpuFeatures::ProbeImpl(bool cross_compile) {
supported_ |= CpuFeaturesImpliedByCompiler();
icache_line_size_ = 128;
+ dcache_line_size_ = 128;
// Only use statically determined features for cross compile (snapshot).
if (cross_compile) return;
@@ -73,6 +74,8 @@
#ifdef USE_SIMULATOR
// Simulator
supported_ |= (1u << PPC_10_PLUS);
+ supported_ |= (1u << ICACHE_SNOOP);
+ supported_ |= (1u << ISELECT);
#else
base::CPU cpu;
if (cpu.part() == base::CPU::kPPCPower10) {
@@ -85,17 +88,37 @@
supported_ |= (1u << PPC_7_PLUS);
} else if (cpu.part() == base::CPU::kPPCPower6) {
supported_ |= (1u << PPC_6_PLUS);
+ } else if (cpu.part() == base::CPU::kPPCPower5 ||
+ cpu.part() == base::CPU::kPPCPA6T) {
+ supported_ |= (1u << PPC_5_PLUS);
+ } else if (cpu.part() == base::CPU::kPPCE6500 ||
+ cpu.part() == base::CPU::kPPCE5500) {
+ supported_ |= (1u << PPC_7_PLUS_NXP); // NXP-supported v2.06 features
+ }
+
+ if (cpu.has_icache_snoop()) {
+ supported_ |= (1u << ICACHE_SNOOP);
+ }
+ if (cpu.has_isel()) {
+ supported_ |= (1u << ISELECT);
}
#if V8_OS_LINUX
if (cpu.icache_line_size() != base::CPU::kUnknownCacheLineSize) {
icache_line_size_ = cpu.icache_line_size();
}
+ if (cpu.dcache_line_size() != base::CPU::kUnknownCacheLineSize) {
+ dcache_line_size_ = cpu.dcache_line_size();
+ }
#endif
#endif
if (supported_ & (1u << PPC_10_PLUS)) supported_ |= (1u << PPC_9_PLUS);
if (supported_ & (1u << PPC_9_PLUS)) supported_ |= (1u << PPC_8_PLUS);
if (supported_ & (1u << PPC_8_PLUS)) supported_ |= (1u << PPC_7_PLUS);
- if (supported_ & (1u << PPC_7_PLUS)) supported_ |= (1u << PPC_6_PLUS);
+ if (supported_ & (1u << PPC_7_PLUS)) {
+ supported_ |= (1u << PPC_7_PLUS_NXP); // NXP-supported v2.06 features
+ supported_ |= (1u << PPC_6_PLUS);
+ }
+ if (supported_ & (1u << PPC_6_PLUS)) supported_ |= (1u << PPC_5_PLUS);
// Set a static value on whether Simd is supported.
// This variable is only used for certain archs to query SupportWasmSimd128()
@@ -117,11 +140,15 @@
}
void CpuFeatures::PrintFeatures() {
+ printf("PPC_5_PLUS=%d\n", CpuFeatures::IsSupported(PPC_5_PLUS));
printf("PPC_6_PLUS=%d\n", CpuFeatures::IsSupported(PPC_6_PLUS));
printf("PPC_7_PLUS=%d\n", CpuFeatures::IsSupported(PPC_7_PLUS));
printf("PPC_8_PLUS=%d\n", CpuFeatures::IsSupported(PPC_8_PLUS));
printf("PPC_9_PLUS=%d\n", CpuFeatures::IsSupported(PPC_9_PLUS));
printf("PPC_10_PLUS=%d\n", CpuFeatures::IsSupported(PPC_10_PLUS));
+ printf("ICACHE_SNOOP=%d\n", CpuFeatures::IsSupported(ICACHE_SNOOP));
+ printf("ISELECT=%d\n", CpuFeatures::IsSupported(ISELECT));
+ printf("PPC_7_PLUS_NXP=%d\n", CpuFeatures::IsSupported(PPC_7_PLUS_NXP));
}
Register ToRegister(int num) {
Reference in New Issue View Git Blame Copy Permalink