Skip to contentSkip to conversation

Bug#1003490: u-boot: FTBFS on arch:all with qemu-ppce500 target

DISCUSSION Vagrant Cascadian 2 participants 3 messages Jan 26, 2022 4 messages not in the archive

lore
Share & tools
lore
#4 lore raw
In reply to a message not in the archive

A bug in Debian is causing a build failure of the qemu-ppce500 target:

https://bugs.debian.org/1003490

I've CCed u-boot in case they're not aware of the issue.

Some background follows...

Quoted 27 lines · Aurelien Jarno
Quoted 26 lines · Vagrant Cascadian
Quoted 25 lines · Lennart Sorensen
Quoted 24 lines · Vagrant Cascadian

Something in the toolchain recently changed which causes u-boot arch:all build to FTBFS... I suspect binutils, as building in "bookworm" still works fine where binutils hasn't yet migrated.

On arch:all builds the qemu-ppce500 target is cross-compiled.

Full log:

https://buildd.debian.org/status/fetch.php?pkg=u-boot&arch=all&ver=2022.01%2Bdfsg-1&stamp=1641860624&raw=0

The hopefully relevent lines from the build log:

powerpc-linux-gnu-gcc -Wp,-MD,arch/powerpc/cpu/mpc85xx/.tlb.o.d -nostdinc -isystem /usr/lib/gcc-cross/powerpc-linux-gnu/11/include -Iinclude -I/<<PKGBUILDDIR>>/include -I/<<PKGBUILDDIR>>/arch/powerpc/include -include /<<PKGBUILDDIR>>/include/linux/kconfig.h -I/<<PKGBUILDDIR>>/arch/powerpc/cpu/mpc85xx -Iarch/powerpc/cpu/mpc85xx -D__KERNEL__ -D__UBOOT__ -Wall -Wstrict-prototypes -Wno-format-security -fno-builtin -ffreestanding -std=gnu11 -fshort-wchar -fno-strict-aliasing -fno-PIE -Os -fno-stack-protector -fno-delete-null-pointer-checks -Wno-pointer-sign -Wno-stringop-truncation -Wno-zero-length-bounds -Wno-array-bounds -Wno-stringop-overflow -Wno-maybe-uninitialized -fmacro-prefix-map=/<<PKGBUILDDIR>>/= -g -fstack-usage -Wno-format-nonliteral -Wno-address-of-packed-member -Wno-unused-but-set-variable -Werror=date-time -Wno-packed-not-aligned -D__powerpc__ -ffixed-r2 -m32 -fno-ira-hoist-pressure -Wa,-me500 -msoft-float -mno-string -fpic -mrelocatable -ffunction-sections -fdata-sections -mcall-linux -msingle-pic-base -fno-jump-tables -pipe -DKBUILD_BASENAME='"tlb"' -DKBUILD_MODNAME='"tlb"' -c -o arch/powerpc/cpu/mpc85xx/tlb.o /<<PKGBUILDDIR>>/arch/powerpc/cpu/mpc85xx/tlb.c
...
{standard input}: Assembler messages:
{standard input}:127: Error: unrecognized opcode: `tlbre'
{standard input}:418: Error: unrecognized opcode: `tlbre'
{standard input}:821: Error: unrecognized opcode: `msync'
{standard input}:821: Error: unrecognized opcode: `tlbwe'
{standard input}:884: Error: unrecognized opcode: `tlbsx'
make[4]: *** [/<<PKGBUILDDIR>>/scripts/Makefile.build:253: arch/powerpc/cpu/mpc85xx/tlb.o] Error 1 make[3]: *** [/<<PKGBUILDDIR>>/Makefile:1810: arch/powerpc/cpu/mpc85xx] Error 2 make[3]: *** Waiting for unfinished jobs....
powerpc-linux-gnu-gcc -Wp,-MD,arch/powerpc/lib/.traps.o.d -nost

...

Quoted 13 lines
Quoted 5 lines

The binutils versions appear to be:

succeeding, bookworm 2.37-10.1
failing, sid 2.37.50.20220106-2

Yep, this is due to commit b25f942e18d6ecd7ec3e2d2e9930eb4f996c258a on the binutils side [1], which changes the behavior of .machine directives to override, rather than augment, the base CPU. GCC is called with -Wa,-me500 to enable PowerPC e500 instructions on the assembler side, but as the default GCC machine is ppc, a .set machine ppc is emitted at the beginning of the assembly code.

Quoted 3 lines

One option would be to force the CPU to e500 on the GCC side, however support for it has been removed. The options is therefore to force the machine in the assembly code. This is what the attached patch does.

Somehow I missed that you had attached a patch! I will try to get this tested and uploaded to Debian soon...

live well,
vagrant

Quoted 51 lines
arch/powerpc/cpu/mpc85xx/tlb.c +17 -4
@@ -50,7 +50,10 @@ void read_tlbcam_entry(int idx, u32 *val
 	u32 _mas1;
 
 	mtspr(MAS0, FSL_BOOKE_MAS0(1, idx, 0));
-	asm volatile("tlbre;isync");
+	asm volatile(".machine push;\n"
+		     ".machine e500;\n"
+		     "tlbre;isync;\n"
+		     ".machine pop;\n");
 	_mas1 = mfspr(MAS1);
 
 	*valid = (_mas1 & MAS1_VALID);
@@ -109,7 +112,10 @@ void init_used_tlb_cams(void)
 	/* walk all the entries */
 	for (i = 0; i < num_cam; i++) {
 		mtspr(MAS0, FSL_BOOKE_MAS0(1, i, 0));
-		asm volatile("tlbre;isync");
+		asm volatile(".machine push;\n"
+			     ".machine e500;\n"
+			     "tlbre;isync;\n"
+			     ".machine pop;");
 		if (mfspr(MAS1) & MAS1_VALID)
 			use_tlb_cam(i);
 	}
@@ -183,7 +189,10 @@ void disable_tlb(u8 esel)
 #ifdef CONFIG_ENABLE_36BIT_PHYS
 	mtspr(MAS7, 0);
 #endif
-	asm volatile("isync;msync;tlbwe;isync");
+	asm volatile(".machine push;\n"
+		     ".machine e500;\n"
+		     "isync;msync;tlbwe;isync;\n"
+		     ".machine pop;\n");
 
 #ifdef CONFIG_ADDR_MAP
 	if (gd->flags & GD_FLG_RELOC)
@@ -193,7 +202,11 @@ void disable_tlb(u8 esel)
 
 static void tlbsx (const volatile unsigned *addr)
 {
-	__asm__ __volatile__ ("tlbsx 0,%0" : : "r" (addr), "m" (*addr));
+	__asm__ __volatile__ (".machine push;\n"
+			      ".machine e500;\n"
+			      "tlbsx 0,%0;\n"
+			      ".machine pop;\n"
+			      : : "r" (addr), "m" (*addr));
 }
 
 /* return -1 if we didn't find anything */
#5 lore raw
Quoted 15 lines · Vagrant Cascadian
Quoted 14 lines · Aurelien Jarno
Quoted 13 lines · Vagrant Cascadian
Quoted 12 lines · Lennart Sorensen
Quoted 11 lines · Vagrant Cascadian

Something in the toolchain recently changed which causes u-boot arch:all build to FTBFS... I suspect binutils, as building in "bookworm" still works fine where binutils hasn't yet migrated.

On arch:all builds the qemu-ppce500 target is cross-compiled.

Full log:

https://buildd.debian.org/status/fetch.php?pkg=u-boot&arch=all&ver=2022.01%2Bdfsg-1&stamp=1641860624&raw=0

The hopefully relevent lines from the build log:

...

Quoted 24 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

{standard input}: Assembler messages:
{standard input}:127: Error: unrecognized opcode: `tlbre'
{standard input}:418: Error: unrecognized opcode: `tlbre'
{standard input}:821: Error: unrecognized opcode: `msync'
{standard input}:821: Error: unrecognized opcode: `tlbwe'
{standard input}:884: Error: unrecognized opcode: `tlbsx'
make[4]: *** [/<<PKGBUILDDIR>>/scripts/Makefile.build:253: arch/powerpc/cpu/mpc85xx/tlb.o] Error 1 make[3]: *** [/<<PKGBUILDDIR>>/Makefile:1810: arch/powerpc/cpu/mpc85xx] Error 2 make[3]: *** Waiting for unfinished jobs....
powerpc-linux-gnu-gcc -Wp,-MD,arch/powerpc/lib/.traps.o.d -nost

...

Quoted 13 lines
Quoted 5 lines

The binutils versions appear to be:

succeeding, bookworm 2.37-10.1
failing, sid 2.37.50.20220106-2

Yep, this is due to commit b25f942e18d6ecd7ec3e2d2e9930eb4f996c258a on the binutils side [1], which changes the behavior of .machine directives to override, rather than augment, the base CPU. GCC is called with -Wa,-me500 to enable PowerPC e500 instructions on the assembler side, but as the default GCC machine is ppc, a .set machine ppc is emitted at the beginning of the assembly code.

Quoted 6 lines
Quoted 3 lines

One option would be to force the CPU to e500 on the GCC side, however support for it has been removed. The options is therefore to force the machine in the assembly code. This is what the attached patch does.

Somehow I missed that you had attached a patch! I will try to get this tested and uploaded to Debian soon...

Your patch fixed building qemu-ppce500, but now I think we have a potentially similar problem with qemu-riscv64 and qemu-riscv64_smode:

/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c: Assembler messages:
/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c:94: Error: unrecognized opcode
`csrs sstatus,a5'
/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c:95: Error: unrecognized opcode
`csrw 0x003,0'
make[4]: *** [/<<PKGBUILDDIR>>/scripts/Makefile.build:254:
arch/riscv/cpu/cpu.o] Error 1
make[3]: *** [/<<PKGBUILDDIR>>/Makefile:1810: arch/riscv/cpu] Error 2 make[3]: Leaving directory
'/<<PKGBUILDDIR>>/debian/build/qemu-riscv64_smode'

live well,
vagrant

Quoted 51 lines
Quoted 51 lines
arch/powerpc/cpu/mpc85xx/tlb.c +17 -4
@@ -50,7 +50,10 @@ void read_tlbcam_entry(int idx, u32 *val
 	u32 _mas1;
 
 	mtspr(MAS0, FSL_BOOKE_MAS0(1, idx, 0));
-	asm volatile("tlbre;isync");
+	asm volatile(".machine push;\n"
+		     ".machine e500;\n"
+		     "tlbre;isync;\n"
+		     ".machine pop;\n");
 	_mas1 = mfspr(MAS1);
 
 	*valid = (_mas1 & MAS1_VALID);
@@ -109,7 +112,10 @@ void init_used_tlb_cams(void)
 	/* walk all the entries */
 	for (i = 0; i < num_cam; i++) {
 		mtspr(MAS0, FSL_BOOKE_MAS0(1, i, 0));
-		asm volatile("tlbre;isync");
+		asm volatile(".machine push;\n"
+			     ".machine e500;\n"
+			     "tlbre;isync;\n"
+			     ".machine pop;");
 		if (mfspr(MAS1) & MAS1_VALID)
 			use_tlb_cam(i);
 	}
@@ -183,7 +189,10 @@ void disable_tlb(u8 esel)
 #ifdef CONFIG_ENABLE_36BIT_PHYS
 	mtspr(MAS7, 0);
 #endif
-	asm volatile("isync;msync;tlbwe;isync");
+	asm volatile(".machine push;\n"
+		     ".machine e500;\n"
+		     "isync;msync;tlbwe;isync;\n"
+		     ".machine pop;\n");
 
 #ifdef CONFIG_ADDR_MAP
 	if (gd->flags & GD_FLG_RELOC)
@@ -193,7 +202,11 @@ void disable_tlb(u8 esel)
 
 static void tlbsx (const volatile unsigned *addr)
 {
-	__asm__ __volatile__ ("tlbsx 0,%0" : : "r" (addr), "m" (*addr));
+	__asm__ __volatile__ (".machine push;\n"
+			      ".machine e500;\n"
+			      "tlbsx 0,%0;\n"
+			      ".machine pop;\n"
+			      : : "r" (addr), "m" (*addr));
 }
 
 /* return -1 if we didn't find anything */
#6 lore raw
Quoted 51 lines · Vagrant Cascadian
Quoted 15 lines · Vagrant Cascadian
Quoted 14 lines · Aurelien Jarno
Quoted 13 lines · Vagrant Cascadian
Quoted 12 lines · Lennart Sorensen
Quoted 11 lines · Vagrant Cascadian

Something in the toolchain recently changed which causes u-boot arch:all build to FTBFS... I suspect binutils, as building in "bookworm" still works fine where binutils hasn't yet migrated.

On arch:all builds the qemu-ppce500 target is cross-compiled.

Full log:

https://buildd.debian.org/status/fetch.php?pkg=u-boot&arch=all&ver=2022.01%2Bdfsg-1&stamp=1641860624&raw=0

The hopefully relevent lines from the build log:

...

Quoted 24 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

{standard input}: Assembler messages:
{standard input}:127: Error: unrecognized opcode: `tlbre'
{standard input}:418: Error: unrecognized opcode: `tlbre'
{standard input}:821: Error: unrecognized opcode: `msync'
{standard input}:821: Error: unrecognized opcode: `tlbwe'
{standard input}:884: Error: unrecognized opcode: `tlbsx'
make[4]: *** [/<<PKGBUILDDIR>>/scripts/Makefile.build:253: arch/powerpc/cpu/mpc85xx/tlb.o] Error 1 make[3]: *** [/<<PKGBUILDDIR>>/Makefile:1810: arch/powerpc/cpu/mpc85xx] Error 2 make[3]: *** Waiting for unfinished jobs....
powerpc-linux-gnu-gcc -Wp,-MD,arch/powerpc/lib/.traps.o.d -nost

...

Quoted 13 lines
Quoted 5 lines

The binutils versions appear to be:

succeeding, bookworm 2.37-10.1
failing, sid 2.37.50.20220106-2

Yep, this is due to commit b25f942e18d6ecd7ec3e2d2e9930eb4f996c258a on the binutils side [1], which changes the behavior of .machine directives to override, rather than augment, the base CPU. GCC is called with -Wa,-me500 to enable PowerPC e500 instructions on the assembler side, but as the default GCC machine is ppc, a .set machine ppc is emitted at the beginning of the assembly code.

Quoted 6 lines
Quoted 3 lines

One option would be to force the CPU to e500 on the GCC side, however support for it has been removed. The options is therefore to force the machine in the assembly code. This is what the attached patch does.

Somehow I missed that you had attached a patch! I will try to get this tested and uploaded to Debian soon...

Your patch fixed building qemu-ppce500, but now I think we have a potentially similar problem with qemu-riscv64 and qemu-riscv64_smode:

Quoted 10 lines

/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c: Assembler messages:
/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c:94: Error: unrecognized opcode
`csrs sstatus,a5'
/<<PKGBUILDDIR>>/arch/riscv/cpu/cpu.c:95: Error: unrecognized opcode
`csrw 0x003,0'
make[4]: *** [/<<PKGBUILDDIR>>/scripts/Makefile.build:254:
arch/riscv/cpu/cpu.o] Error 1
make[3]: *** [/<<PKGBUILDDIR>>/Makefile:1810: arch/riscv/cpu] Error 2 make[3]: Leaving directory
'/<<PKGBUILDDIR>>/debian/build/qemu-riscv64_smode'

I guess this is due to:
http://lists.infradead.org/pipermail/linux-riscv/2022-January/011728.html

Unfortunately as the change has been done in a not yet released binutils version, there is no way for the fix to have been done upstream yet.

Note that this also breaks at least linux and opensbi.

Signature · 2 lines
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
[email protected]                 http://www.aurel32.net