[RISC-V] memcpy vector expansion should use 8B elements

[antonblanchard]

clang is using the element size in memcpy vector expansions:

#include <string.h>

void foo_64(unsigned long *dst, unsigned long *src)
{
	memcpy(dst, src, 16);
}

void foo_8(unsigned char *dst, unsigned char *src)
{
	memcpy(dst, src, 16);
}

foo_64:
	vsetivli	zero, 2, e64, m1, ta, ma
	vle64.v	v8, (a1)
	vse64.v	v8, (a0)
	ret

foo_8:
	vsetivli	zero, 16, e8, m1, ta, ma
	vle8.v	v8, (a1)
	vse8.v	v8, (a0)
	ret

This prevents expansion of the memcpy with 64bit elements on a CPU with vector support but no unaligned vector (eg Spacemit K1). If we always used 8B elements for the copy we wouldn't have this issue. I notice gcc does this.

[topperc]

I'm not sure what issue you're raising here. An unsigned long * must be 8 byte aligned by ABI rules. Do you have a different example?

[llvmbot]

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

Author: Anton Blanchard (antonblanchard)

clang is using the element size in memcpy vector expansions:

#include <string.h>

void foo_64(unsigned long *dst, unsigned long *src)
{
	memcpy(dst, src, 16);
}

void foo_8(unsigned char *dst, unsigned char *src)
{
	memcpy(dst, src, 16);
}

foo_64:
	vsetivli	zero, 2, e64, m1, ta, ma
	vle64.v	v8, (a1)
	vse64.v	v8, (a0)
	ret

foo_8:
	vsetivli	zero, 16, e8, m1, ta, ma
	vle8.v	v8, (a1)
	vse8.v	v8, (a0)
	ret

This prevents expansion of the memcpy with 64bit elements on a CPU with vector support but no unaligned vector (eg Spacemit K1). If we always used 8B elements for the copy we wouldn't have this issue. I notice gcc does this.

[antonblanchard]

Sorry, the test case is probably junk. I'm seeing the issue in a c++ global static initializer. It's trying to copy something that is 2B aligned but we are using a 64B element size for the memcpy:

  addi    a2,a2,-1128 # 2139d4a
  vsetivli        zero,2,e64,m1,ta,ma
  vle64.v v8,(a2)
  vse64.v v8,(a0)

I'll see if I can come up with a relevant test case.