Skip to contentSkip to conversation

x86/idle: Work around LLVM assembler bug with MONITOR and MWAIT insn

PATCH Superseded Uros Bizjak 2 participants 6 messages Apr 3, 2025 x86

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/idle: Work around LLVM assembler bug with MONITOR and MWAIT insn

LLVM assembler is not able to assemble correct forms of MONITOR and MWAIT instructions with explicit operands:

error: invalid operand for instruction

          monitor %rax,%ecx,%edx
                       ^~~~

Use instruction mnemonics with implicit operands when LLVM assembler is detected to work around this issue.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Fixes cd3b85b27542 ("x86/idle: Use MONITOR and MWAIT mnemonics in <asm/mwait.h>")
---
 arch/x86/include/asm/mwait.h | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
arch/x86/include/asm/mwait.h +13 -3
index 5141d2acc0ef..d956d1be4873 100644
@@ -25,9 +25,18 @@
 #define TPAUSE_C01_STATE		1
 #define TPAUSE_C02_STATE		0
 
+#ifdef CONFIG_LD_IS_LLD
+# define ASM_MONITOR	"monitor"
+# define ASM_MWAIT	"mwait"
+#else
+# define ASM_MONITOR	"monitor %[eax], %[ecx], %[edx]"
+# define ASM_MWAIT	"mwait %[eax], %[ecx]"
+#endif
+
 static __always_inline void __monitor(const void *eax, u32 ecx, u32 edx)
 {
-	asm volatile("monitor %0, %1, %2" :: "a" (eax), "c" (ecx), "d" (edx));
+	asm volatile(ASM_MONITOR
+		     :: [eax] "a" (eax), [ecx] "c" (ecx), [edx] "d" (edx));
 }
 
 static __always_inline void __monitorx(const void *eax, u32 ecx, u32 edx)
@@ -41,7 +50,7 @@ static __always_inline void __mwait(u32 eax, u32 ecx)
 {
 	mds_idle_clear_cpu_buffers();
 
-	asm volatile("mwait %0, %1" :: "a" (eax), "c" (ecx));
+	asm volatile(ASM_MWAIT :: [eax] "a" (eax), [ecx] "c" (ecx));
 }
 
 /*
@@ -92,7 +101,8 @@ static __always_inline void __sti_mwait(u32 eax, u32 ecx)
 {
 	mds_idle_clear_cpu_buffers();
 
-	asm volatile("sti; mwait %0, %1" :: "a" (eax), "c" (ecx));
+	asm volatile("sti\n\t"
+		     ASM_MWAIT :: [eax] "a" (eax), [ecx] "c" (ecx));
 }
 
 /*
Signature · 1 line
2.49.0
maintainer
#1 lore raw
Quoted 10 lines · Uros Bizjak

LLVM assembler is not able to assemble correct forms of MONITOR and MWAIT instructions with explicit operands:

error: invalid operand for instruction

          monitor %rax,%ecx,%edx
                       ^~~~

Use instruction mnemonics with implicit operands when LLVM assembler is detected to work around this issue.

Quoted 7 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Fixes cd3b85b27542 ("x86/idle: Use MONITOR and MWAIT mnemonics in <asm/mwait.h>")

No, you should whack that one - the toolchains are clearly not ready yet...

arch/x86/include/asm/mwait.h line 34

Quoted 7 lines
arch/x86/include/asm/mwait.h +7 -0
+#ifdef CONFIG_LD_IS_LLD
+# define ASM_MONITOR	"monitor"
+# define ASM_MWAIT	"mwait"
+#else
+# define ASM_MONITOR	"monitor %[eax], %[ecx], %[edx]"
+# define ASM_MWAIT	"mwait %[eax], %[ecx]"
+#endif

... instead of piling more ifdeffery ontop and making the code uglier than before.

Go fix the LLVM assembler to deal with explicit operands or GCC and LLVM should agree on common syntax they both assemble or whatever and the kernel should do *one* thing exactly and not carry toolchain-specific ifdeffery at every spot.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#2 lore raw
Quoted 20 lines · Borislav Petkov
Quoted 10 lines · Uros Bizjak

LLVM assembler is not able to assemble correct forms of MONITOR and MWAIT instructions with explicit operands:

error: invalid operand for instruction

          monitor %rax,%ecx,%edx
                       ^~~~

Use instruction mnemonics with implicit operands when LLVM assembler is detected to work around this issue.

Quoted 7 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Fixes cd3b85b27542 ("x86/idle: Use MONITOR and MWAIT mnemonics in <asm/mwait.h>")
Quoted 1 line

No, you should whack that one - the toolchains are clearly not ready yet...

The least common denominator is an insn with implicit operands. I'll post V2 that fixes it that way.

Uros.

maintainer
#3 lore raw
Quoted 2 lines · Uros Bizjak

The least common denominator is an insn with implicit operands. I'll post V2 that fixes it that way.

I guess.

You don't need to post a patch ontop - your previous one should be removed and the whole effort can start all over again.

Maybe Ingo would let the test bots chew the patches a bit first before applying them so quickly so that we can avoid this churn too.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#4 lore raw
Quoted 9 lines · Borislav Petkov
Quoted 2 lines · Uros Bizjak

The least common denominator is an insn with implicit operands. I'll post V2 that fixes it that way.

I guess.

You don't need to post a patch ontop - your previous one should be removed and the whole effort can start all over again.

IMO, in this case the fixup also documents the LLVM bug, so perhaps the fixup on top is desirable for documentation purposes.

BTW: The test with CC=clang make argument didn't expose this issue.

Uros.

maintainer
#5 lore raw
Quoted 2 lines · Uros Bizjak

IMO, in this case the fixup also documents the LLVM bug, so perhaps the fixup on top is desirable for documentation purposes.

No.

1. You put a comment over the function so that someone sees it and *actually* fixes it in the compilers

2. Explain in the commit message of the correct patches *why* it cannot be done like you were trying initially and leave a lore link into them so that people can find the old discussion.

Lemme zap those patches so that they can be done properly.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette