[RISCV] Mark vector register used for scalar function?

[zzzhhhl]

LLVM will mark vector register as used register with regmask of scalar function.

for example, if we used a fmax (only scalar), llvm will generate a regmask operand. And when we analyze register usage using MRI.isPhysRegUsed(Reg), it shows all the vector register is used. In some scenario, it will cause vector register spill. Does it make senses?

target-flags(riscv-call) @fmax, <regmask $vlenb $x0 $x1 $x8 $x9 $x18 $x19 $x20 $x21 $x22 $x23 $x24 $x25 $x26 $x27 $f8_d $f9_d $f18_d $f19_d $f20_d $f21_d $f22_d $f23_d $f24_d $f25_d $f26_d $f27_d $f8_f $f9_f $f18_f $f19_f $f20_f $f21_f and 24 more...>, implicit-def dead $x1, implicit $f10_d, implicit $f11_d, implicit-def $x2, implicit-def $f10_d

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: None (zzzhhhl)

LLVM will mark vector register as used register with regmask of scalar function.

for example, if we used a fmax (only scalar), llvm will generate a regmask operand. And when we analyze register usage using MRI.isPhysRegUsed(Reg), it shows all the vector register is used. In some scenario, it will cause vector register spill. Does it make senses?

target-flags(riscv-call) @<!-- -->fmax, &lt;regmask $vlenb $x0 $x1 $x8 $x9 $x18 $x19 $x20 $x21 $x22 $x23 $x24 $x25 $x26 $x27 $f8_d $f9_d $f18_d $f19_d $f20_d $f21_d $f22_d $f23_d $f24_d $f25_d $f26_d $f27_d $f8_f $f9_f $f18_f $f19_f $f20_f $f21_f and 24 more...&gt;, implicit-def dead $x1, implicit $f10_d, implicit $f11_d, implicit-def $x2, implicit-def $f10_d

[topperc]

The default ABI says that vector registers are not preserved across calls so spilling across a call would be correct. The compiler doesn't know anything special about fmax and there is no rule that fmax can't use vectors or call some other function that uses vectors.

[zzzhhhl]

The default ABI says that vector registers are not preserved across calls so spilling across a call would be correct. The compiler doesn't know anything special about fmax and there is no rule that fmax can't use vectors or call some other function that uses vectors.

Is there a option to tell compiler that the function can't use vector register to disable this spilling?

[wangpc-pp]

Can you provide a minimal reproducer? @zzzhhhl