Skip to content

[AMDGPU][True16][CodeGen] select vgpr16 for asm inline 16bit vreg #140946

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
May 21, 2025
Merged

[AMDGPU][True16][CodeGen] select vgpr16 for asm inline 16bit vreg #140946

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c1b6d35
16bit for asm inline reg
broxigarchen May 21, 2025
67d4d9f
auto generate check line
broxigarchen May 21, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion llvm/lib/Target/AMDGPU/SIISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16062,7 +16062,8 @@ SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI_,
case 'v':
switch (BitWidth) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

switch seems overkill here. Could just handle it with:

  if (BitWidth == 16 && !Subtarget->useRealTrue16Insts())
    BitWidth = 32;

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Or even move the handling for the BitWidth == 16 case inside getVGPRClassForBitWidth?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This reminds me we need constraints for the aligned and unaligned versions of register classes

case 16:
RC = &AMDGPU::VGPR_32RegClass;
RC = Subtarget->useRealTrue16Insts() ? &AMDGPU::VGPR_16RegClass
: &AMDGPU::VGPR_32RegClass;
break;
default:
RC = TRI->getVGPRClassForBitWidth(BitWidth);
Expand Down
79 changes: 79 additions & 0 deletions llvm/test/CodeGen/AMDGPU/inlineasm-16-fake16.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,79 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can you auto generate check lines?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure


define amdgpu_kernel void @s_input_output_i16() #0 {
; GFX11-LABEL: s_input_output_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: s_mov_b32 s0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_and_b32 s0, s0, 0xffff
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use s0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call i16 asm sideeffect "s_mov_b32 $0, -1", "=s"()
tail call void asm sideeffect "; use $0", "s"(i16 %v) #0
ret void
}

define amdgpu_kernel void @s_input_output_f16() #0 {
; GFX11-LABEL: s_input_output_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: s_mov_b32 s0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use s0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call half asm sideeffect "s_mov_b32 $0, -1", "=s"() #0
tail call void asm sideeffect "; use $0", "s"(half %v)
ret void
}

define amdgpu_kernel void @v_input_output_f16() #0 {
; GFX11-LABEL: v_input_output_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: v_mov_b32 v0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call half asm sideeffect "v_mov_b32 $0, -1", "=v"() #0
tail call void asm sideeffect "; use $0", "v"(half %v)
ret void
}

define amdgpu_kernel void @v_input_output_i16() #0 {
; GFX11-LABEL: v_input_output_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: v_mov_b32 v0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call i16 asm sideeffect "v_mov_b32 $0, -1", "=v"() #0
tail call void asm sideeffect "; use $0", "v"(i16 %v)
ret void
}

define amdgpu_kernel void @i16_imm_input_phys_vgpr() {
; GFX11-LABEL: i16_imm_input_phys_vgpr:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: v_mov_b32_e32 v0, 0xffff
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
entry:
call void asm sideeffect "; use $0 ", "{v0}"(i16 65535)
ret void
}

attributes #0 = { nounwind }
91 changes: 91 additions & 0 deletions llvm/test/CodeGen/AMDGPU/inlineasm-16-true16.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,91 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s

define amdgpu_kernel void @s_input_output_i16() #0 {
; GFX11-LABEL: s_input_output_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: s_mov_b32 s0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_and_b32 s0, s0, 0xffff
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use s0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call i16 asm sideeffect "s_mov_b32 $0, -1", "=s"()
tail call void asm sideeffect "; use $0", "s"(i16 %v) #0
ret void
}

define amdgpu_kernel void @s_input_output_f16() #0 {
; GFX11-LABEL: s_input_output_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: s_mov_b32 s0, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use s0
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call half asm sideeffect "s_mov_b32 $0, -1", "=s"() #0
tail call void asm sideeffect "; use $0", "s"(half %v)
ret void
}

define amdgpu_kernel void @v_input_output_f16() #0 {
; GFX11-LABEL: v_input_output_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: v_mov_b16 v0.l, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0.l
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call half asm sideeffect "v_mov_b16 $0, -1", "=v"() #0
tail call void asm sideeffect "; use $0", "v"(half %v)
ret void
}

define amdgpu_kernel void @v_input_output_i16() #0 {
; GFX11-LABEL: v_input_output_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: v_mov_b16 v0.l, -1
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0.l
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
%v = tail call i16 asm sideeffect "v_mov_b16 $0, -1", "=v"() #0
tail call void asm sideeffect "; use $0", "v"(i16 %v)
ret void
}

define amdgpu_kernel void @i16_imm_input_phys_vgpr_lo() {
; GFX11-LABEL: i16_imm_input_phys_vgpr_lo:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: v_mov_b16_e32 v0.l, -1
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0.l
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
entry:
call void asm sideeffect "; use $0 ", "{v0.l}"(i16 65535)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you please add one more test with v0.h?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

And do we need a constraint to specify specifically an l or h for a virtual register?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added a .h case

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

And do we need a constraint to specify specifically an l or h for a virtual register?

I do not think it is practically needed. At least it is not needed for correctness at this point.

ret void
}

define amdgpu_kernel void @i16_imm_input_phys_vgpr_hi() {
; GFX11-LABEL: i16_imm_input_phys_vgpr_hi:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: v_mov_b16_e32 v0.h, -1
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v0.h
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_endpgm
entry:
call void asm sideeffect "; use $0 ", "{v0.h}"(i16 65535)
ret void
}

attributes #0 = { nounwind }
Loading