Skip to content

AMDGPU: Add a few missing mfma rewrite tests #149026

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

Jump to bottom
Closed
wants to merge 1 commit into from
Closed

AMDGPU: Add a few missing mfma rewrite tests #149026

wants to merge 1 commit into from

Conversation

arsenm
Copy link
Contributor

@arsenm arsenm commented Jul 16, 2025

Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

@arsenm Graphite App
Copy link
Contributor Author

arsenm commented Jul 16, 2025 •
edited
Loading

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

This stack of pull requests is managed by Graphite. Learn more about stacking.

@arsenm arsenm requested review from jrbyrnes and srpande July 16, 2025 07:02
This was referenced Jul 16, 2025
Merged
Closed
@arsenm arsenm marked this pull request as ready for review July 16, 2025 07:02
@llvmbot
Copy link
Member

llvmbot commented Jul 16, 2025

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

Patch is 80.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/149026.diff

2 Files Affected:

  • (modified) llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-agpr-negative-tests.mir (+437)
  • (modified) llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-av-with-load-source.mir (+267-137)
diff --git a/llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-agpr-negative-tests.mir b/llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-agpr-negative-tests.mir
index cd10fbb6020b8..482846e7a9313 100644
--- a/llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-agpr-negative-tests.mir
+++ b/llvm/test/CodeGen/AMDGPU/inflate-reg-class-vgpr-mfma-to-agpr-negative-tests.mir
@@ -20,7 +20,24 @@
     ret void
   }
 
+  define amdgpu_kernel void @inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first() #1 {
+    ret void
+  }
+
+  define amdgpu_kernel void @inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_second() #1 {
+    ret void
+  }
+
+  define amdgpu_kernel void @inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first_physreg() #1 {
+    ret void
+  }
+
+  define amdgpu_kernel void @inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_second_physreg() #1 {
+    ret void
+  }
+
   attributes #0 = { "amdgpu-wave-limiter"="true" "amdgpu-waves-per-eu"="8,8" }
+  attributes #1 = { "amdgpu-wave-limiter"="true" "amdgpu-waves-per-eu"="10,10" }
 ...
 
 # Inflate pattern, except the defining instruction isn't an MFMA.
@@ -393,3 +410,423 @@ body:             |
     S_ENDPGM 0
 
 ...
+
+# There isn't an assignable AGPR around the first MFMA.
+---
+name:            inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first
+tracksRegLiveness: true
+machineFunctionInfo:
+  isEntryFunction: true
+  stackPtrOffsetReg: '$sgpr32'
+  occupancy:       10
+  sgprForEXECCopy: '$sgpr100_sgpr101'
+body:             |
+  ; CHECK-LABEL: name: inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr0
+  ; CHECK-NEXT:   renamable $sgpr0 = S_MOV_B32 0
+  ; CHECK-NEXT:   renamable $vgpr8 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK-NEXT:   renamable $sgpr1 = COPY renamable $sgpr0
+  ; CHECK-NEXT:   renamable $vgpr18_vgpr19 = COPY killed renamable $sgpr0_sgpr1
+  ; CHECK-NEXT:   renamable $vcc = S_AND_B64 $exec, -1, implicit-def dead $scc
+  ; CHECK-NEXT:   dead renamable $vgpr9 = COPY renamable $vgpr8
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; CHECK-NEXT:   liveins: $vcc, $vgpr18_vgpr19
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   renamable $vgpr16_vgpr17 = GLOBAL_LOAD_DWORDX2 undef renamable $vgpr0_vgpr1, 0, 0, implicit $exec :: (load (s64), addrspace 1)
+  ; CHECK-NEXT:   S_NOP 0, implicit-def renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit-def renamable $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit-def renamable $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39_agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47, implicit-def renamable $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55_agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   early-clobber renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31, 0, 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   S_NOP 0, implicit killed renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit killed renamable $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit killed renamable $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39_agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47, implicit killed renamable $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55_agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   early-clobber renamable $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, 0, 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; CHECK-NEXT:   S_BRANCH %bb.2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   liveins: $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35:0x00000000FFFFFFFF
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+  ; CHECK-NEXT:   renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = COPY killed renamable $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+  ; CHECK-NEXT:   renamable $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr8_agpr9_agpr10_agpr11, undef $sgpr0_sgpr1, 32, 0, implicit $exec :: (store (s128), align 32, addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr12_agpr13_agpr14_agpr15, undef $sgpr0_sgpr1, 48, 0, implicit $exec :: (store (s128), addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr0_agpr1_agpr2_agpr3, undef $sgpr0_sgpr1, 0, 0, implicit $exec :: (store (s128), align 128, addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR killed renamable $vgpr0, killed renamable $agpr4_agpr5_agpr6_agpr7, killed undef $sgpr0_sgpr1, 16, 0, implicit $exec :: (store (s128), addrspace 1)
+  ; CHECK-NEXT:   S_ENDPGM 0
+  bb.0:
+    S_NOP 0, implicit-def $agpr0
+    renamable $sgpr0 = S_MOV_B32 0
+    undef %0.sub8:vreg_512_align2 = V_MOV_B32_e32 0, implicit $exec
+    renamable $sgpr1 = COPY renamable $sgpr0
+    %1:vreg_64_align2 = COPY killed renamable $sgpr0_sgpr1
+    renamable $vcc = S_AND_B64 $exec, -1, implicit-def dead $scc
+    %0.sub9:vreg_512_align2 = COPY %0.sub8
+
+  bb.1:
+    liveins: $vcc
+
+    undef %0.sub0_sub1:vreg_512_align2 = GLOBAL_LOAD_DWORDX2 undef %2:vreg_64_align2, 0, 0, implicit $exec :: (load (s64), addrspace 1)
+    S_NOP 0, implicit-def %6:areg_512_align2, implicit-def %7:areg_512_align2, implicit-def %8:areg_512_align2, implicit-def %9:areg_512_align2
+    %3:vreg_512_align2 = V_MFMA_F32_32X32X8F16_vgprcd_e64 %1, %1, %0, 0, 0, 0, implicit $mode, implicit $exec
+    S_NOP 0, implicit %6, implicit %7, implicit %8, implicit %9
+    %4:vreg_512_align2 = V_MFMA_F32_32X32X8F16_vgprcd_e64 %1, %1, %3, 0, 0, 0, implicit $mode, implicit $exec
+    S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+    S_BRANCH %bb.2
+
+  bb.2:
+    ; No VGPRs available for %0 or %4
+    S_NOP 0, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+    S_NOP 0, implicit-def $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+    S_NOP 0, implicit-def $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23
+    S_NOP 0, implicit-def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+    S_NOP 0, implicit-def $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39
+    S_NOP 0, implicit-def $vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47
+    S_NOP 0, implicit-def $vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55
+    S_NOP 0, implicit-def $vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+    %5:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub8_sub9_sub10_sub11, undef $sgpr0_sgpr1, 32, 0, implicit $exec :: (store (s128), align 32, addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub12_sub13_sub14_sub15, undef $sgpr0_sgpr1, 48, 0, implicit $exec :: (store (s128), addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub0_sub1_sub2_sub3, undef $sgpr0_sgpr1, 0, 0, implicit $exec :: (store (s128), align 128, addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub4_sub5_sub6_sub7, killed undef $sgpr0_sgpr1, 16, 0, implicit $exec :: (store (s128), addrspace 1)
+    S_ENDPGM 0
+
+...
+
+# There isn't an assignable AGPR around the second MFMA.
+---
+name:            inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_second
+tracksRegLiveness: true
+machineFunctionInfo:
+  isEntryFunction: true
+  stackPtrOffsetReg: '$sgpr32'
+  occupancy:       10
+  sgprForEXECCopy: '$sgpr100_sgpr101'
+body:             |
+  ; CHECK-LABEL: name: inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_second
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr0
+  ; CHECK-NEXT:   renamable $sgpr0 = S_MOV_B32 0
+  ; CHECK-NEXT:   renamable $vgpr8 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK-NEXT:   renamable $sgpr1 = COPY renamable $sgpr0
+  ; CHECK-NEXT:   renamable $vgpr18_vgpr19 = COPY killed renamable $sgpr0_sgpr1
+  ; CHECK-NEXT:   renamable $vcc = S_AND_B64 $exec, -1, implicit-def dead $scc
+  ; CHECK-NEXT:   dead renamable $vgpr9 = COPY renamable $vgpr8
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; CHECK-NEXT:   liveins: $vcc, $vgpr18_vgpr19
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   renamable $vgpr16_vgpr17 = GLOBAL_LOAD_DWORDX2 undef renamable $vgpr0_vgpr1, 0, 0, implicit $exec :: (load (s64), addrspace 1)
+  ; CHECK-NEXT:   early-clobber renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31, 0, 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   S_NOP 0, implicit-def renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit-def renamable $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit-def renamable $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39_agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47, implicit-def renamable $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55_agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   early-clobber renamable $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, 0, 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   S_NOP 0, implicit killed renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit killed renamable $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit killed renamable $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39_agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47, implicit killed renamable $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55_agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; CHECK-NEXT:   S_BRANCH %bb.2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   liveins: $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35:0x00000000FFFFFFFF
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+  ; CHECK-NEXT:   renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = COPY killed renamable $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+  ; CHECK-NEXT:   renamable $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr8_agpr9_agpr10_agpr11, undef $sgpr0_sgpr1, 32, 0, implicit $exec :: (store (s128), align 32, addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr12_agpr13_agpr14_agpr15, undef $sgpr0_sgpr1, 48, 0, implicit $exec :: (store (s128), addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR renamable $vgpr0, renamable $agpr0_agpr1_agpr2_agpr3, undef $sgpr0_sgpr1, 0, 0, implicit $exec :: (store (s128), align 128, addrspace 1)
+  ; CHECK-NEXT:   GLOBAL_STORE_DWORDX4_SADDR killed renamable $vgpr0, killed renamable $agpr4_agpr5_agpr6_agpr7, killed undef $sgpr0_sgpr1, 16, 0, implicit $exec :: (store (s128), addrspace 1)
+  ; CHECK-NEXT:   S_ENDPGM 0
+  bb.0:
+    S_NOP 0, implicit-def $agpr0
+    renamable $sgpr0 = S_MOV_B32 0
+    undef %0.sub8:vreg_512_align2 = V_MOV_B32_e32 0, implicit $exec
+    renamable $sgpr1 = COPY renamable $sgpr0
+    %1:vreg_64_align2 = COPY killed renamable $sgpr0_sgpr1
+    renamable $vcc = S_AND_B64 $exec, -1, implicit-def dead $scc
+    %0.sub9:vreg_512_align2 = COPY %0.sub8
+
+  bb.1:
+    liveins: $vcc
+
+    undef %0.sub0_sub1:vreg_512_align2 = GLOBAL_LOAD_DWORDX2 undef %2:vreg_64_align2, 0, 0, implicit $exec :: (load (s64), addrspace 1)
+    %3:vreg_512_align2 = V_MFMA_F32_32X32X8F16_vgprcd_e64 %1, %1, %0, 0, 0, 0, implicit $mode, implicit $exec
+    S_NOP 0, implicit-def %6:areg_512_align2, implicit-def %7:areg_512_align2, implicit-def %8:areg_512_align2, implicit-def %9:areg_512_align2
+    %4:vreg_512_align2 = V_MFMA_F32_32X32X8F16_vgprcd_e64 %1, %1, %3, 0, 0, 0, implicit $mode, implicit $exec
+    S_NOP 0, implicit %6, implicit %7, implicit %8, implicit %9
+    S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+    S_BRANCH %bb.2
+
+  bb.2:
+    ; No VGPRs available for %0 or %4
+    S_NOP 0, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+    S_NOP 0, implicit-def $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+    S_NOP 0, implicit-def $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23
+    S_NOP 0, implicit-def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+    S_NOP 0, implicit-def $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39
+    S_NOP 0, implicit-def $vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47
+    S_NOP 0, implicit-def $vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55
+    S_NOP 0, implicit-def $vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+    %5:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub8_sub9_sub10_sub11, undef $sgpr0_sgpr1, 32, 0, implicit $exec :: (store (s128), align 32, addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub12_sub13_sub14_sub15, undef $sgpr0_sgpr1, 48, 0, implicit $exec :: (store (s128), addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub0_sub1_sub2_sub3, undef $sgpr0_sgpr1, 0, 0, implicit $exec :: (store (s128), align 128, addrspace 1)
+    GLOBAL_STORE_DWORDX4_SADDR %5, %4.sub4_sub5_sub6_sub7, killed undef $sgpr0_sgpr1, 16, 0, implicit $exec :: (store (s128), addrspace 1)
+    S_ENDPGM 0
+
+...
+
+# There isn't an assignable AGPR around the first MFMA, with physreg interference
+---
+name:            inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first_physreg
+tracksRegLiveness: true
+machineFunctionInfo:
+  isEntryFunction: true
+  stackPtrOffsetReg: '$sgpr32'
+  occupancy:       10
+  sgprForEXECCopy: '$sgpr100_sgpr101'
+body:             |
+  ; CHECK-LABEL: name: inflate_result_to_agpr__V_MFMA_F32_32X32X8F16_vgprcd_e64_chain_no_agprs_first_physreg
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr0
+  ; CHECK-NEXT:   renamable $sgpr0 = S_MOV_B32 0
+  ; CHECK-NEXT:   renamable $vgpr8 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK-NEXT:   renamable $sgpr1 = COPY renamable $sgpr0
+  ; CHECK-NEXT:   renamable $vgpr18_vgpr19 = COPY killed renamable $sgpr0_sgpr1
+  ; CHECK-NEXT:   renamable $vcc = S_AND_B64 $exec, -1, implicit-def dead $scc
+  ; CHECK-NEXT:   dead renamable $vgpr9 = COPY renamable $vgpr8
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; CHECK-NEXT:   liveins: $vcc, $vgpr18_vgpr19
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   renamable $vgpr16_vgpr17 = GLOBAL_LOAD_DWORDX2 undef renamable $vgpr0_vgpr1, 0, 0, implicit $exec :: (load (s64), addrspace 1)
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55
+  ; CHECK-NEXT:   S_NOP 0, implicit-def $agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   early-clobber renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31, 0, 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr32_agpr33_agpr34_agpr35_agpr36_agpr37_agpr38_agpr39
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr40_agpr41_agpr42_agpr43_agpr44_agpr45_agpr46_agpr47
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr48_agpr49_agpr50_agpr51_agpr52_agpr53_agpr54_agpr55
+  ; CHECK-NEXT:   S_NOP 0, implicit $agpr56_agpr57_agpr58_agpr59_agpr60_agpr61_agpr62_agpr63
+  ; CHECK-NEXT:   early-clobber renamable $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35 = V_MFMA_F32_32X32X8F16_vgprcd_e64 $vgpr18_vgpr19, $vgpr18_vgpr19, killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, 0, ...
[truncated]

@arsenm
Copy link
Contributor Author

arsenm commented Jul 24, 2025

ping

@arsenm arsenm force-pushed the users/arsenm/amdgpu/add-mfma-rewrite-other-split-pattern-tests branch from cb47016 to 15d9c6a Compare July 25, 2025 01:14
@arsenm arsenm force-pushed the users/arsenm/amdgpu/handle-multiple-agpr-mfma-rewrites branch from 3ec2a31 to 8e03847 Compare July 25, 2025 01:14
@arsenm arsenm mentioned this pull request Jul 25, 2025
Closed
@arsenm
AMDGPU: Add a few missing mfma rewrite tests
4c13a1e 
Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.
@arsenm arsenm force-pushed the users/arsenm/amdgpu/add-mfma-rewrite-other-split-pattern-tests branch from 15d9c6a to 4c13a1e Compare July 26, 2025 23:58
@arsenm arsenm force-pushed the users/arsenm/amdgpu/handle-multiple-agpr-mfma-rewrites branch from 8e03847 to 33aad42 Compare July 26, 2025 23:58
arsenm added a commit that referenced this pull request Aug 7, 2025
@arsenm
AMDGPU: Add a few missing mfma rewrite tests
f2bacff 
Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

I'm accumulating too many failure cases, and MIR tests are very prone
to painful merge conflicts, so I've added a few more tests and extracted
new tests from #147975.

Closes #149026
@arsenm arsenm mentioned this pull request Aug 7, 2025
Merged
arsenm added a commit that referenced this pull request Aug 7, 2025
@arsenm
AMDGPU: Add a few missing mfma rewrite tests
cba7e52 
Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

I'm accumulating too many failure cases, and MIR tests are very prone
to painful merge conflicts, so I've added a few more tests and extracted
new tests from #147975.

Closes #149026
arsenm added a commit that referenced this pull request Aug 7, 2025
@arsenm
AMDGPU: Add a few missing mfma rewrite tests
224e38d 
Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

I'm accumulating too many failure cases, and MIR tests are very prone
to painful merge conflicts, so I've added a few more tests and extracted
new tests from #147975.

Closes #149026
arsenm added a commit that referenced this pull request Aug 7, 2025
@arsenm
AMDGPU: Add a few missing mfma rewrite tests (#152434)
f44d8d5 
Test other splitting situations that appear in greedy.
This includes ensuring we have a case that hits a local split
and instruction split (most of the tests hit the region split path).

Also test a few cases where the final result isn't fully used, resulting
in partial copy bundles instead of a simple full copy. Test physreg
and virtreg agpr interference with a reassignment candidate.

I'm accumulating too many failure cases, and MIR tests are very prone
to painful merge conflicts, so I've added a few more tests and extracted
new tests from #147975.

Closes #149026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@srpande srpande srpande approved these changes

@Pierre-vh Pierre-vh Pierre-vh approved these changes

@jrbyrnes jrbyrnes Awaiting requested review from jrbyrnes

Assignees
No one assigned
Labels
backend:AMDGPU
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@arsenm @llvmbot @srpande @Pierre-vh