This is an archive of the discontinued LLVM Phabricator instance.

[AArch64][SVE2] Asm: add SQRDMLAH/SQRDMLSH instructions
ClosedPublic

Authored by c-rhodes on May 3 2019, 8:31 AM.

Diff Detail

Event Timeline

c-rhodes created this revision.May 3 2019, 8:31 AM
rovka added a subscriber: rovka.May 6 2019, 3:30 AM
rovka added inline comments.
test/MC/AArch64/SVE2/sqrdmlah-diagnostics.s
8

Why not "error: Invalid restricted vector register, expected z0.h..z7.h" ?

c-rhodes added inline comments.May 7 2019, 8:13 AM
test/MC/AArch64/SVE2/sqrdmlah-diagnostics.s
8

This is being parsed as the vector form of the instruction where everything until the index is invalid, I guess it's hitting that error first. I agree it's a poor diagnostic, I'll see if this can be improved.

c-rhodes marked an inline comment as done.May 13 2019, 6:48 AM
c-rhodes added inline comments.
test/MC/AArch64/SVE2/sqrdmlah-diagnostics.s
8

I understand what's happening here a bit better now, the debug output from asm-matcher explains it pretty well:

Trying to match opcode SQRDMLAH_ZZZ_H
  Matching formal operand class MCK_SVEVectorHReg against actual operand at index 1 (<register 245>): match success using generic matcher
  Matching formal operand class MCK_SVEVectorHReg against actual operand at index 2 (<register 246>): match success using generic matcher
  Matching formal operand class MCK_SVEVectorHReg against actual operand at index 3 (<register 253>): match success using generic matcher
  Matching formal operand class InvalidMatchClass against actual operand at index 4 (<vectorindex 0>): Opcode result: multiple operand mismatches, ignoring this opcode
...
Trying to match opcode SQRDMLAH_ZZZI_H
  Matching formal operand class MCK_SVEVectorHReg against actual operand at index 1 (<register 245>): match success using generic matcher
  Matching formal operand class MCK_SVEVectorHReg against actual operand at index 2 (<register 246>): match success using generic matcher
  Matching formal operand class MCK_SVEVector3bHReg against actual operand at index 3 (<register 253>): Opcode result: multiple operand mismatches, ignoring this opcode

The asm-matcher finds no match for this instruction, and in MatchInstructionImpl the default result is invalid operand:

// Some state to try to produce better error messages.
unsigned RetCode = Match_InvalidOperand;

This isn't new to SVE2, the same diagnostic is given for SVE instructions that have unpredicated and indexed forms, for instance SDOT. The following is from sdot-diagnostics.s:

// ------------------------------------------------------------------------- //
// Invalid restricted register for indexed vector.

sdot  z0.s, z1.b, z8.b[3]
// CHECK: [[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
// CHECK-NEXT: sdot  z0.s, z1.b, z8.b[3]
// CHECK-NOT: [[@LINE-1]]:{{[0-9]+}}:

I'm not sure what can be done from TableGen to improve this

c-rhodes marked an inline comment as not done.May 13 2019, 6:48 AM
rovka accepted this revision.May 14 2019, 1:31 AM

Ok, thanks for looking into it! LGTM

This revision is now accepted and ready to land.May 14 2019, 1:31 AM
This revision was automatically updated to reflect the committed changes.