[AMDGPU] Fix predicates for BUFFER_ATOMIC_FMIN/FMAX patterns #89066
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Use OtherPredicates to avoid interfering with other uses of SubtargetPredicate for GFX12.
|
@llvm/pr-subscribers-backend-amdgpu Author: Jay Foad (jayfoad) ChangesUse OtherPredicates to avoid interfering with other uses of Full diff: https://github.com/llvm/llvm-project/pull/89066.diff 2 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/BUFInstructions.td b/llvm/lib/Target/AMDGPU/BUFInstructions.td
index 273f92abf35465..8053d89aeb0a89 100644
--- a/llvm/lib/Target/AMDGPU/BUFInstructions.td
+++ b/llvm/lib/Target/AMDGPU/BUFInstructions.td
@@ -1726,7 +1726,7 @@ let SubtargetPredicate = isGFX12Plus in {
defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["noret"]>;
}
-let SubtargetPredicate = isGFX6GFX7GFX10Plus in {
+let OtherPredicates = [isGFX6GFX7GFX10Plus] in {
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmin", f32, "BUFFER_ATOMIC_FMIN">;
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmax", f32, "BUFFER_ATOMIC_FMAX">;
}
diff --git a/llvm/test/CodeGen/AMDGPU/fp-min-max-buffer-atomics.ll b/llvm/test/CodeGen/AMDGPU/fp-min-max-buffer-atomics.ll
index 0c62b52eb92afc..587340c7aa342c 100644
--- a/llvm/test/CodeGen/AMDGPU/fp-min-max-buffer-atomics.ll
+++ b/llvm/test/CodeGen/AMDGPU/fp-min-max-buffer-atomics.ll
@@ -4,12 +4,14 @@
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck %s -check-prefix=GFX10
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s -check-prefix=GFX1030
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck %s -check-prefix=GFX1100
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs | FileCheck %s -check-prefix=GFX12
; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s -check-prefix=G_SI
; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=hawaii -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX7
; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX10
; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX1030
; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX1100
+; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs | FileCheck %s -check-prefix=GFX12
declare float @llvm.amdgcn.raw.buffer.atomic.fmin.f32(float, <4 x i32>, i32, i32, i32 immarg)
declare float @llvm.amdgcn.raw.buffer.atomic.fmax.f32(float, <4 x i32>, i32, i32, i32 immarg)
@@ -70,6 +72,18 @@ define amdgpu_kernel void @raw_buffer_atomic_min_noret_f32(<4 x i32> inreg %rsrc
; GFX1100-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_min_noret_f32:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: s_clause 0x1
+; GFX12-NEXT: s_load_b64 s[4:5], s[0:1], 0x34
+; GFX12-NEXT: s_load_b128 s[0:3], s[0:1], 0x24
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX12-NEXT: buffer_atomic_min_num_f32 v0, v1, s[0:3], null offen
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_min_noret_f32:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd
@@ -170,6 +184,15 @@ define amdgpu_ps void @raw_buffer_atomic_min_rtn_f32(<4 x i32> inreg %rsrc, floa
; GFX1100-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_min_rtn_f32:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: buffer_atomic_min_num_f32 v0, v1, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_store_b32 v[0:1], v0, off
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_min_rtn_f32:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
@@ -292,6 +315,20 @@ define amdgpu_kernel void @raw_buffer_atomic_min_rtn_f32_off4_slc(<4 x i32> inre
; GFX1100-NEXT: ds_store_b32 v1, v0
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_min_rtn_f32_off4_slc:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: s_clause 0x1
+; GFX12-NEXT: s_load_b96 s[4:6], s[0:1], 0x34
+; GFX12-NEXT: s_load_b128 s[0:3], s[0:1], 0x24
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX12-NEXT: s_mov_b32 s4, 4
+; GFX12-NEXT: buffer_atomic_min_num_f32 v0, v1, s[0:3], s4 offen th:TH_ATOMIC_NT_RETURN
+; GFX12-NEXT: v_mov_b32_e32 v1, s6
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: ds_store_b32 v1, v0
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_min_rtn_f32_off4_slc:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xd
@@ -427,6 +464,18 @@ define amdgpu_kernel void @raw_buffer_atomic_max_noret_f32(<4 x i32> inreg %rsrc
; GFX1100-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_max_noret_f32:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: s_clause 0x1
+; GFX12-NEXT: s_load_b64 s[4:5], s[0:1], 0x34
+; GFX12-NEXT: s_load_b128 s[0:3], s[0:1], 0x24
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX12-NEXT: buffer_atomic_max_num_f32 v0, v1, s[0:3], null offen
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_max_noret_f32:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xd
@@ -527,6 +576,15 @@ define amdgpu_ps void @raw_buffer_atomic_max_rtn_f32(<4 x i32> inreg %rsrc, floa
; GFX1100-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_max_rtn_f32:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: buffer_atomic_max_num_f32 v0, v1, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_store_b32 v[0:1], v0, off
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_max_rtn_f32:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
@@ -641,6 +699,20 @@ define amdgpu_kernel void @raw_buffer_atomic_max_rtn_f32_off4_slc(<4 x i32> inre
; GFX1100-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX1100-NEXT: s_endpgm
;
+; GFX12-LABEL: raw_buffer_atomic_max_rtn_f32_off4_slc:
+; GFX12: ; %bb.0: ; %main_body
+; GFX12-NEXT: s_load_b256 s[0:7], s[0:1], 0x24
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX12-NEXT: s_mov_b32 s4, 4
+; GFX12-NEXT: buffer_atomic_max_num_f32 v0, v1, s[0:3], s4 offen th:TH_ATOMIC_NT_RETURN
+; GFX12-NEXT: v_mov_b32_e32 v1, 0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_store_b32 v1, v0, s[6:7]
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
; G_SI-LABEL: raw_buffer_atomic_max_rtn_f32_off4_slc:
; G_SI: ; %bb.0: ; %main_body
; G_SI-NEXT: s_load_dwordx8 s[0:7], s[0:1], 0x9
|
jayfoad
commented
Apr 17, 2024
| @@ -1726,7 +1726,7 @@ let SubtargetPredicate = isGFX12Plus in { | |||
| defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["noret"]>; | |||
| } | |||
|
|
|||
| let SubtargetPredicate = isGFX6GFX7GFX10Plus in { | |||
| let OtherPredicates = [isGFX6GFX7GFX10Plus] in { | |||
There was a problem hiding this comment.
It is annoying not to be able to use SubtargetPredicate for something that is clearly a subtarget predicate, but it clashes with the use of SubtargetPredicate = HasUnrestrictedSOffset inside SIBufferAtomicPat.
There was a problem hiding this comment.
Yeah, I came across this in the past too.
There was a problem hiding this comment.
I would expect these to be swapped, with HasUnrestrictedSOffset in OtherPredicates
There was a problem hiding this comment.
Right, but that just causes the converse problem, as there are already a bunch of cases that override OtherPredicates in SIBufferAtomicPat.
|
#78701 fixed a different instance of the same problem. |
piotrAMD
approved these changes
Apr 17, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Use OtherPredicates to avoid interfering with other uses of
SubtargetPredicate for GFX12.