| Message ID | 20250127112715.2936555-1-peter.maydell@linaro.org |
|---|---|
| Headers | show |
| Series | target/arm: Deprecate iwMMXt emulation and associated CPUs | expand |
On 1/27/25 03:27, Peter Maydell wrote: > This patchset marks all the CPUs that support iwMMXt as deprecated, > because I don't believe anybody is using them, and we have no way to > test the thousands of lines of code we have that's only there for > iwMMXt emulation. (See for instance the recent thread where a patch > was submitted to fix an issue detected by a static analyzer: we > didn't take it, largely because we had no way to know if the > behaviour change the patch would produce was correct or not.) > > The pxa2xx CPUs are now only useful with user-mode emulation, because > we dropped all the machine types that used them in 9.2. (Technically > you could alse use "-cpu pxa270" with a board model like versatilepb > which doesn't sanity-check the CPU type, but that has never been a > supported config.) > > To use them (or iwMMXt emulation) with QEMU user-mode you would need > to explicitly select them with the -cpu option or the QEMU_CPU > environment variable, because the '-cpu max' default CPU does not > include iwMMXt emulation. A google search finds no examples of > anybody doing this in the last decade. > > I asked some of the Linaro GCC folks if they were using QEMU to test > their iwMMXt codegen, or knew anybody doing that upstream, and the > answer was "no". In fact, GCC is in the process of dropping support > for iwMMXt entirely. > > We have one test case in check-tcg which purports to be checking > iwMMXt. In fact it is doing no such thing: it runs the test without > selecting an iwMMXt CPU, which means the iwMMXt insns are interpreted > as FPA11 insns by the linux-user emulate_arm_fpa11() code. So the test > prints garbage and then succeeds anyway. Modern distro toolchains > can't generate a binary that will run with -cpu pxa270 (because their > crt startup code uses Thumb insns); rather than putting in a lot of > effort trying to salvage the test case to really test a feature we've > deprecated, I opted to just remove the test. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> r~
On 27/1/25 12:27, Peter Maydell wrote: > This patchset marks all the CPUs that support iwMMXt as deprecated, > because I don't believe anybody is using them, and we have no way to > test the thousands of lines of code we have that's only there for > iwMMXt emulation. (See for instance the recent thread where a patch > was submitted to fix an issue detected by a static analyzer: we > didn't take it, largely because we had no way to know if the > behaviour change the patch would produce was correct or not.) > > The pxa2xx CPUs are now only useful with user-mode emulation, because > we dropped all the machine types that used them in 9.2. (Technically > you could alse use "-cpu pxa270" with a board model like versatilepb > which doesn't sanity-check the CPU type, but that has never been a > supported config.) > > To use them (or iwMMXt emulation) with QEMU user-mode you would need > to explicitly select them with the -cpu option or the QEMU_CPU > environment variable, because the '-cpu max' default CPU does not > include iwMMXt emulation. A google search finds no examples of > anybody doing this in the last decade. > > I asked some of the Linaro GCC folks if they were using QEMU to test > their iwMMXt codegen, or knew anybody doing that upstream, and the > answer was "no". In fact, GCC is in the process of dropping support > for iwMMXt entirely. > > We have one test case in check-tcg which purports to be checking > iwMMXt. In fact it is doing no such thing: it runs the test without > selecting an iwMMXt CPU, which means the iwMMXt insns are interpreted > as FPA11 insns by the linux-user emulate_arm_fpa11() code. So the test > prints garbage and then succeeds anyway. Modern distro toolchains Uh... > can't generate a binary that will run with -cpu pxa270 (because their > crt startup code uses Thumb insns); rather than putting in a lot of > effort trying to salvage the test case to really test a feature we've > deprecated, I opted to just remove the test. > > -- PMM > > Peter Maydell (2): > target/arm: deprecate the pxa2xx CPUs and iwMMXt emulation > tests/tcg/arm: Remove test-arm-iwmmxt test Series: Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
On Mon, 27 Jan 2025 at 17:27, Philippe Mathieu-Daudé <philmd@linaro.org> wrote: > On 27/1/25 12:27, Peter Maydell wrote: > > We have one test case in check-tcg which purports to be checking > > iwMMXt. In fact it is doing no such thing: it runs the test without > > selecting an iwMMXt CPU, which means the iwMMXt insns are interpreted > > as FPA11 insns by the linux-user emulate_arm_fpa11() code. So the test > > prints garbage and then succeeds anyway. > > Uh... It had me confused for a while, because I'd convinced myself by code inspection that we only enabled iwMMXt if you manually asked for a pxa2xx CPU and not in the default 'max' CPU, and then I found this test case that we were running with the default CPU type and which was apparently passing :-) -- PMM
On 1/27/25 09:46, Peter Maydell wrote: > On Mon, 27 Jan 2025 at 17:27, Philippe Mathieu-Daudé <philmd@linaro.org> wrote: >> On 27/1/25 12:27, Peter Maydell wrote: >>> We have one test case in check-tcg which purports to be checking >>> iwMMXt. In fact it is doing no such thing: it runs the test without >>> selecting an iwMMXt CPU, which means the iwMMXt insns are interpreted >>> as FPA11 insns by the linux-user emulate_arm_fpa11() code. So the test >>> prints garbage and then succeeds anyway. >> >> Uh... > > It had me confused for a while, because I'd convinced myself > by code inspection that we only enabled iwMMXt if you > manually asked for a pxa2xx CPU and not in the default 'max' > CPU, and then I found this test case that we were running > with the default CPU type and which was apparently passing :-) Any chance we can deprecate the fpa11 stuff too? I see it's still enabled by the kernel for some boards, but it depends on really old configury: config FPE_NWFPE bool "NWFPE math emulation" depends on (!AEABI || OABI_COMPAT) && !THUMB2_KERNEL help Say Y to include the NWFPE floating point emulator in the kernel. This is necessary to run most binaries. Linux does not currently support floating point hardware so you need to say Y here even if your machine has an FPA or floating point co-processor podule. No thumb2, no eabi, and oabi_compat enabled. I suspect that's a very rare configuration these days. I'll note the help message is somewhat dated. :-) r~
This patchset marks all the CPUs that support iwMMXt as deprecated, because I don't believe anybody is using them, and we have no way to test the thousands of lines of code we have that's only there for iwMMXt emulation. (See for instance the recent thread where a patch was submitted to fix an issue detected by a static analyzer: we didn't take it, largely because we had no way to know if the behaviour change the patch would produce was correct or not.) The pxa2xx CPUs are now only useful with user-mode emulation, because we dropped all the machine types that used them in 9.2. (Technically you could alse use "-cpu pxa270" with a board model like versatilepb which doesn't sanity-check the CPU type, but that has never been a supported config.) To use them (or iwMMXt emulation) with QEMU user-mode you would need to explicitly select them with the -cpu option or the QEMU_CPU environment variable, because the '-cpu max' default CPU does not include iwMMXt emulation. A google search finds no examples of anybody doing this in the last decade. I asked some of the Linaro GCC folks if they were using QEMU to test their iwMMXt codegen, or knew anybody doing that upstream, and the answer was "no". In fact, GCC is in the process of dropping support for iwMMXt entirely. We have one test case in check-tcg which purports to be checking iwMMXt. In fact it is doing no such thing: it runs the test without selecting an iwMMXt CPU, which means the iwMMXt insns are interpreted as FPA11 insns by the linux-user emulate_arm_fpa11() code. So the test prints garbage and then succeeds anyway. Modern distro toolchains can't generate a binary that will run with -cpu pxa270 (because their crt startup code uses Thumb insns); rather than putting in a lot of effort trying to salvage the test case to really test a feature we've deprecated, I opted to just remove the test. -- PMM Peter Maydell (2): target/arm: deprecate the pxa2xx CPUs and iwMMXt emulation tests/tcg/arm: Remove test-arm-iwmmxt test docs/about/deprecated.rst | 21 ++++++++++++++ target/arm/cpu.h | 1 + target/arm/cpu.c | 3 ++ target/arm/tcg/cpu32.c | 36 ++++++++++++++++-------- tests/tcg/arm/Makefile.target | 7 ----- tests/tcg/arm/README | 5 ---- tests/tcg/arm/test-arm-iwmmxt.S | 49 --------------------------------- 7 files changed, 49 insertions(+), 73 deletions(-) delete mode 100644 tests/tcg/arm/test-arm-iwmmxt.S