Fix FreeBSD/PowerPC* issues by kgotlinux · Pull Request #2147 · abseil/abseil-cpp

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.

kgotlinux · 2026-08-31T18:24:16Z

This is a cumulative PR for two FreeBSD/PowerPC* changes:

Fix for UnscaledCycleClock::Frequency().
Fix for saved-LR slot on 32-bits.

google-cla · 2026-08-31T18:24:39Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

StacktracePowerPCGetLRPtr() falls back to OS checks when the compiler does not provide the _CALL_* ABI macros. Clang defines neither _CALL_SYSV nor _CALL_AIX on any FreeBSD PowerPC target (64-bit gets only _CALL_ELF=2), so 32-bit powerpc-*-freebsd hits the __FreeBSD__ fallback and reads the LR from sp[2]. The 32-bit SVR4 ABI saves the LR in the second word of the frame, sp[1]; sp[2] is correct only for the 64-bit ELF ABIs and Darwin/AIX. Verified against clang codegen: powerpc-unknown-freebsd emits "stwu 1,-16(1); stw 0,20(1)" (old SP + 4) while powerpc64-unknown-freebsd emits "stdu 1,-48(1); std 0,64(1)" (old SP + 16). Key the fallback on the word size instead of the OS: any __PPC64__ target uses sp[2], and 32-bit ELF targets (Linux, FreeBSD) use sp[1]. This is behavior-preserving for Linux and for 64-bit FreeBSD, and fixes the 32-bit FreeBSD case (currently latent, since stacktrace_config.h does not select this unwinder on FreeBSD).

The FreeBSD branch reads the kern.timecounter.tc.timebase.frequency sysctl straight into a double. The kernel registers that OID as CTLTYPE_U64 (sys/kern/kern_tc.c, tc_init()), so sysctlbyname() fills the 8-byte buffer with the integer 512000000 and the bits are then interpreted as an IEEE double: Frequency() returns ~2.5e-315 instead of 5.12e+08. The size check cannot catch this because uint64_t and double are both 8 bytes. Observed on FreeBSD 15.1 powerpc64le and powerpc64 (POWER9): - LOG_EVERY_N_SEC / VLOG_EVERY_N_SEC never rate-limit, because LogEveryNSecState::ShouldLog() computes the next log time as now + seconds * Frequency() ~= now (1000 of 1000 iterations logged with a 3600 s period). - Mutex waiters refresh their pthread_getschedparam() cache on every block instead of once per second. Read the sysctl into a uint64_t and convert to double; keep 0.0 if the sysctl fails.