forked from torvalds/linux
-
Notifications
You must be signed in to change notification settings - Fork 14
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
invalid output size for constraint '=q' in drivers/gpu/drm/i915/i915_query.c #781
Labels
duplicate
This issue or pull request already exists
Comments
Quick fix: --- a/arch/x86/include/asm/uaccess.h
+++ b/arch/x86/include/asm/uaccess.h
@@ -349,7 +349,7 @@ do { \
__chk_user_ptr(ptr); \
switch (size) { \
case 1: \
- __get_user_asm(x, ptr, retval, "b", "b", "=q", errret); \
+ __get_user_asm(x, ptr, retval, "b", "b", "=r", errret); \
break; \
case 2: \
__get_user_asm(x, ptr, retval, "w", "w", "=r", errret); \
@@ -399,7 +399,7 @@ do { \
__chk_user_ptr(ptr); \
switch (size) { \
case 1: \
- __get_user_asm_ex(x, ptr, "b", "b", "=q"); \
+ __get_user_asm_ex(x, ptr, "b", "b", "=r"); \
break; \
case 2: \
__get_user_asm_ex(x, ptr, "w", "w", "=r"); \ |
nathanchance
pushed a commit
that referenced
this issue
Apr 30, 2020
Running with KASAN on a VIM3L systems leads to the following splat when probing the Ethernet device: ================================================================== BUG: KASAN: global-out-of-bounds in _get_maxdiv+0x74/0xd8 Read of size 4 at addr ffffa000090615f4 by task systemd-udevd/139 CPU: 1 PID: 139 Comm: systemd-udevd Tainted: G E 5.7.0-rc1-00101-g8624b7577b9c #781 Hardware name: amlogic w400/w400, BIOS 2020.01-rc5 03/12/2020 Call trace: dump_backtrace+0x0/0x2a0 show_stack+0x20/0x30 dump_stack+0xec/0x148 print_address_description.isra.12+0x70/0x35c __kasan_report+0xfc/0x1d4 kasan_report+0x4c/0x68 __asan_load4+0x9c/0xd8 _get_maxdiv+0x74/0xd8 clk_divider_bestdiv+0x74/0x5e0 clk_divider_round_rate+0x80/0x1a8 clk_core_determine_round_nolock.part.9+0x9c/0xd0 clk_core_round_rate_nolock+0xf0/0x108 clk_hw_round_rate+0xac/0xf0 clk_factor_round_rate+0xb8/0xd0 clk_core_determine_round_nolock.part.9+0x9c/0xd0 clk_core_round_rate_nolock+0xf0/0x108 clk_core_round_rate_nolock+0xbc/0x108 clk_core_set_rate_nolock+0xc4/0x2e8 clk_set_rate+0x58/0xe0 meson8b_dwmac_probe+0x588/0x72c [dwmac_meson8b] platform_drv_probe+0x78/0xd8 really_probe+0x158/0x610 driver_probe_device+0x140/0x1b0 device_driver_attach+0xa4/0xb0 __driver_attach+0xcc/0x1c8 bus_for_each_dev+0xf4/0x168 driver_attach+0x3c/0x50 bus_add_driver+0x238/0x2e8 driver_register+0xc8/0x1e8 __platform_driver_register+0x88/0x98 meson8b_dwmac_driver_init+0x28/0x1000 [dwmac_meson8b] do_one_initcall+0xa8/0x328 do_init_module+0xe8/0x368 load_module+0x3300/0x36b0 __do_sys_finit_module+0x120/0x1a8 __arm64_sys_finit_module+0x4c/0x60 el0_svc_common.constprop.2+0xe4/0x268 do_el0_svc+0x98/0xa8 el0_svc+0x24/0x68 el0_sync_handler+0x12c/0x318 el0_sync+0x158/0x180 The buggy address belongs to the variable: div_table.63646+0x34/0xfffffffffffffa40 [dwmac_meson8b] Memory state around the buggy address: ffffa00009061480: fa fa fa fa 00 00 00 01 fa fa fa fa 00 00 00 00 ffffa00009061500: 05 fa fa fa fa fa fa fa 00 04 fa fa fa fa fa fa >ffffa00009061580: 00 03 fa fa fa fa fa fa 00 00 00 00 00 00 fa fa ^ ffffa00009061600: fa fa fa fa 00 01 fa fa fa fa fa fa 01 fa fa fa ffffa00009061680: fa fa fa fa 00 01 fa fa fa fa fa fa 04 fa fa fa ================================================================== Digging into this indeed shows that the clock divider array is lacking a final fence, and that the clock subsystems goes in the weeds. Oh well. Let's add the empty structure that indicates the end of the array. Fixes: bd6f485 ("net: stmmac: dwmac-meson8b: Fix the RGMII TX delay on Meson8b/8m2 SoCs") Signed-off-by: Marc Zyngier <maz@kernel.org> Cc: Martin Blumenstingl <martin.blumenstingl@googlemail.com> Reviewed-by: Martin Blumenstingl <martin.blumenstingl@googlemail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
fengguang
pushed a commit
to 0day-ci/linux
that referenced
this issue
May 5, 2020
GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead. This can happen with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case *before* the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings/errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. We also don't want to swap the use of "=q" with "=r". For 64b, it doesn't matter. For 32b, it's possible that a 32b register without a 8b lower alias (i.e. ESI, EDI, EBP) is selected which the assembler will then reject. With this, Clang can finally build an i386 defconfig. Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux#3 Link: ClangBuiltLinux#194 Link: ClangBuiltLinux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
krasCGQ
pushed a commit
to krasCGQ/linux
that referenced
this issue
Jun 25, 2020
GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead. This can happen with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case *before* the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings/errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. We also don't want to swap the use of "=q" with "=r". For 64b, it doesn't matter. For 32b, it's possible that a 32b register without a 8b lower alias (i.e. ESI, EDI, EBP) is selected which the assembler will then reject. With this, Clang can finally build an i386 defconfig. Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux/linux#3 Link: ClangBuiltLinux/linux#194 Link: ClangBuiltLinux/linux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Albert I <kras@raphielgang.org>
fengguang
pushed a commit
to 0day-ci/linux
that referenced
this issue
Jul 20, 2020
GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead. This can happen with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case *before* the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings/errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. We also don't want to swap the use of "=q" with "=r". For 64b, it doesn't matter. For 32b, it's possible that a 32b register without a 8b lower alias (i.e. ESI, EDI, EBP) is selected which the assembler will then reject. With this, Clang can finally build an i386 defconfig. Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux#3 Link: ClangBuiltLinux#194 Link: ClangBuiltLinux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
dup of #194 |
ruscur
pushed a commit
to ruscur/linux
that referenced
this issue
Jul 22, 2020
GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead. This can happen with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case *before* the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings/errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. We also don't want to swap the use of "=q" with "=r". For 64b, it doesn't matter. For 32b, it's possible that a 32b register without a 8b lower alias (i.e. ESI, EDI, EBP) is selected which the assembler will then reject. With this, Clang can finally build an i386 defconfig. Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux#3 Link: ClangBuiltLinux#194 Link: ClangBuiltLinux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Dennis Zhou <dennis@kernel.org>
fengguang
pushed a commit
to 0day-ci/linux
that referenced
this issue
Jul 24, 2020
Clang fails to compile __get_user_size() on 32-bit for the following code: long long val; __get_user(val, usrptr); with: error: invalid output size for constraint '=q' GCC compiles the same code without complaints. The reason is that GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead, i.e. with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case before the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings / errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. It was discussed to change the asm output constraint for the one byte case from "=q" to "=r". While it works for 64-bit, it fails on 32-bit. With '=r' the compiler could fail to chose a register accessible as high/low which is required for the byte operation. If that happens the assembly will fail. Use a local temporary variable of type 'unsigned char' as output for the byte copy inline asm and then assign it to the real output variable. This prevents Clang from failing the semantic analysis in the above case. The resulting code for the actual one byte copy is not affected as the temporary variable is optimized out. [ tglx: Amended changelog ] Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Dennis Zhou <dennis@kernel.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux#3 Link: ClangBuiltLinux#194 Link: ClangBuiltLinux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Link: https://lkml.kernel.org/r/20200720204925.3654302-12-ndesaulniers@google.com
BoredOutOfMyGit
pushed a commit
to codeaurora-unofficial/kernel-msm-3.10
that referenced
this issue
Aug 7, 2020
Clang fails to compile __get_user_size() on 32-bit for the following code: long long val; __get_user(val, usrptr); with: error: invalid output size for constraint '=q' GCC compiles the same code without complaints. The reason is that GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead, i.e. with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case before the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings / errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. It was discussed to change the asm output constraint for the one byte case from "=q" to "=r". While it works for 64-bit, it fails on 32-bit. With '=r' the compiler could fail to chose a register accessible as high/low which is required for the byte operation. If that happens the assembly will fail. Use a local temporary variable of type 'unsigned char' as output for the byte copy inline asm and then assign it to the real output variable. This prevents Clang from failing the semantic analysis in the above case. The resulting code for the actual one byte copy is not affected as the temporary variable is optimized out. [ tglx: Amended changelog ] Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Dennis Zhou <dennis@kernel.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux/linux#3 Link: ClangBuiltLinux/linux#194 Link: ClangBuiltLinux/linux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Link: https://lkml.kernel.org/r/20200720204925.3654302-12-ndesaulniers@google.com (cherry picked from commit 158807d git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git x86/asm) [adelva: Minor re-diff around __get_user_asm() API change since 5.4] Bug: 154934534 Signed-off-by: Alistair Delva <adelva@google.com> Change-Id: Icac4ed451acf788a10e35bdfe80e9aae15b77e1b
gregkh
pushed a commit
to gregkh/linux
that referenced
this issue
Aug 19, 2020
[ Upstream commit 158807d ] Clang fails to compile __get_user_size() on 32-bit for the following code: long long val; __get_user(val, usrptr); with: error: invalid output size for constraint '=q' GCC compiles the same code without complaints. The reason is that GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead, i.e. with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case before the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings / errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. It was discussed to change the asm output constraint for the one byte case from "=q" to "=r". While it works for 64-bit, it fails on 32-bit. With '=r' the compiler could fail to chose a register accessible as high/low which is required for the byte operation. If that happens the assembly will fail. Use a local temporary variable of type 'unsigned char' as output for the byte copy inline asm and then assign it to the real output variable. This prevents Clang from failing the semantic analysis in the above case. The resulting code for the actual one byte copy is not affected as the temporary variable is optimized out. [ tglx: Amended changelog ] Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Dennis Zhou <dennis@kernel.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux/linux#3 Link: ClangBuiltLinux/linux#194 Link: ClangBuiltLinux/linux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Link: https://lkml.kernel.org/r/20200720204925.3654302-12-ndesaulniers@google.com Signed-off-by: Sasha Levin <sashal@kernel.org>
gregkh
pushed a commit
to gregkh/linux
that referenced
this issue
Aug 19, 2020
[ Upstream commit 158807d ] Clang fails to compile __get_user_size() on 32-bit for the following code: long long val; __get_user(val, usrptr); with: error: invalid output size for constraint '=q' GCC compiles the same code without complaints. The reason is that GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead, i.e. with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case before the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings / errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. It was discussed to change the asm output constraint for the one byte case from "=q" to "=r". While it works for 64-bit, it fails on 32-bit. With '=r' the compiler could fail to chose a register accessible as high/low which is required for the byte operation. If that happens the assembly will fail. Use a local temporary variable of type 'unsigned char' as output for the byte copy inline asm and then assign it to the real output variable. This prevents Clang from failing the semantic analysis in the above case. The resulting code for the actual one byte copy is not affected as the temporary variable is optimized out. [ tglx: Amended changelog ] Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Dennis Zhou <dennis@kernel.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux/linux#3 Link: ClangBuiltLinux/linux#194 Link: ClangBuiltLinux/linux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Link: https://lkml.kernel.org/r/20200720204925.3654302-12-ndesaulniers@google.com Signed-off-by: Sasha Levin <sashal@kernel.org>
jackpot51
pushed a commit
to pop-os/linux
that referenced
this issue
Sep 4, 2020
BugLink: https://bugs.launchpad.net/bugs/1892215 [ Upstream commit 158807d ] Clang fails to compile __get_user_size() on 32-bit for the following code: long long val; __get_user(val, usrptr); with: error: invalid output size for constraint '=q' GCC compiles the same code without complaints. The reason is that GCC and Clang are architecturally different, which leads to subtle issues for code that's invalid but clearly dead, i.e. with code that emulates polymorphism with the preprocessor and sizeof. GCC will perform semantic analysis after early inlining and dead code elimination, so it will not warn on invalid code that's dead. Clang strictly performs optimizations after semantic analysis, so it will warn for dead code. Neither Clang nor GCC like this very much with -m32: long long ret; asm ("movb $5, %0" : "=q" (ret)); However, GCC can tolerate this variant: long long ret; switch (sizeof(ret)) { case 1: asm ("movb $5, %0" : "=q" (ret)); break; case 8:; } Clang, on the other hand, won't accept that because it validates the inline asm for the '1' case before the optimisation phase where it realises that it wouldn't have to emit it anyway. If LLVM (Clang's "back end") fails such as during instruction selection or register allocation, it cannot provide accurate diagnostics (warnings / errors) that contain line information, as the AST has been discarded from memory at that point. While there have been early discussions about having C/C++ specific language optimizations in Clang via the use of MLIR, which would enable such earlier optimizations, such work is not scoped and likely a multi-year endeavor. It was discussed to change the asm output constraint for the one byte case from "=q" to "=r". While it works for 64-bit, it fails on 32-bit. With '=r' the compiler could fail to chose a register accessible as high/low which is required for the byte operation. If that happens the assembly will fail. Use a local temporary variable of type 'unsigned char' as output for the byte copy inline asm and then assign it to the real output variable. This prevents Clang from failing the semantic analysis in the above case. The resulting code for the actual one byte copy is not affected as the temporary variable is optimized out. [ tglx: Amended changelog ] Reported-by: Arnd Bergmann <arnd@arndb.de> Reported-by: David Woodhouse <dwmw2@infradead.org> Reported-by: Dmitry Golovin <dima@golovin.in> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Dennis Zhou <dennis@kernel.org> Link: https://bugs.llvm.org/show_bug.cgi?id=33587 Link: ClangBuiltLinux#3 Link: ClangBuiltLinux#194 Link: ClangBuiltLinux#781 Link: https://lore.kernel.org/lkml/20180209161833.4605-1-dwmw2@infradead.org/ Link: https://lore.kernel.org/lkml/CAK8P3a1EBaWdbAEzirFDSgHVJMtWjuNt2HGG8z+vpXeNHwETFQ@mail.gmail.com/ Link: https://lkml.kernel.org/r/20200720204925.3654302-12-ndesaulniers@google.com Signed-off-by: Sasha Levin <sashal@kernel.org> Signed-off-by: Seth Forshee <seth.forshee@canonical.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
A new case similar to of #194:
How haven't I hit this earlier? 🤔
The text was updated successfully, but these errors were encountered: