[RISCV] `clang` permits invalid operand for `Zclsd` instructions in inline assembly when using any optimization level other than `-O0`

[thebigclub]

The Zclsd extension requires the use of registers in the range of x8-15 for c.ld and c.sd instructions. clang will use an invalid register when the compiler is free to choose the destination. This occurs for both the I and E base ISAs:

$ cat << EOF > test.c
#include <stdint.h>

int fn(void *arg1, int arg2, int arg3, int arg4, int arg5) {
        uint64_t val;

        asm volatile ("c.ld %0, 0(%1)"
                      : "=R" (val)
                      : "r" (arg1)
        );

        return val + arg2 + arg3 + arg4 + arg5;
}
EOF
$ for isa in i e; do for level in 0 1 2 3 s z; do (set -x; clang -march=rv32"$isa"_zclsd -O"$level" -c test.c); echo Return value: $?; done; done
+ clang -march=rv32i_zclsd -O0 -c test.c
Return value: 0
+ clang -march=rv32i_zclsd -O1 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32i_zclsd -O2 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32i_zclsd -O3 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32i_zclsd -Os -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32i_zclsd -Oz -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -O0 -c test.c
Return value: 0
+ clang -march=rv32e_zclsd -O1 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -O2 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -O3 -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -Os -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -Oz -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1
$ clang -v
clang version 21.0.0git (https://github.com/llvm/llvm-project.git 913dcf1aa36f3ea2d67a0d2b05b9d1375987e553)
Target: riscv32-unknown-unknown-elf

[thebigclub]

Forgot to include -Og in my test but it happens for that as well:

$ for isa in i e; do for level in g; do (set -x; clang -march=rv32"$isa"_zclsd -O"$level" -c test.c); echo Return value: $?; done; done
+ clang -march=rv32i_zclsd -Og -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -Og -c test.c
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1

[thebigclub]

And -Ofast:

$ for isa in i e; do for level in fast; do (set -x; clang -march=rv32"$isa"_zclsd -O"$level" -c test.c); echo Return value: $?; done; done
+ clang -march=rv32i_zclsd -Ofast -c test.c
clang: warning: argument '-Ofast' is deprecated; use '-O3 -ffast-math' for the same behavior, or '-O3' to enable only conforming optimizations [-Wdeprecated-ofast]
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld a6, 0(a0)
      |              ^
1 error generated.
Return value: 1
+ clang -march=rv32e_zclsd -Ofast -c test.c
clang: warning: argument '-Ofast' is deprecated; use '-O3 -ffast-math' for the same behavior, or '-O3' to enable only conforming optimizations [-Wdeprecated-ofast]
test.c:6:23: error: invalid operand for instruction
    6 |         asm volatile ("c.ld %0, 0(%1)"
      |                       ^
<inline asm>:1:7: note: instantiated into assembly here
    1 |         c.ld t1, 0(a0)
      |              ^
1 error generated.
Return value: 1

[topperc]

The text of inline assembly is not parsed by the compiler. It's basically treated like a printf format string. The compiler only fills in the %0 parts and hands it off to the assembler. Register allocation constraints must be specified in the constraint string. The correct constraint is cR instead of R. See riscv-non-isa/riscv-c-api-doc#102