[llvm-branch-commits] [llvm] [AMDGPU] Use a bf16 operand profile for v_fma_mix_bf16_t16 (PR #224910)

[via llvm-branch-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Dmitry Sidorov (MrSidims)

<details>
<summary>Changes</summary>

Assisted-by: Claude Code Opus 5

---

<sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>

---
Full diff: https://github.com/llvm/llvm-project/pull/224910.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/VOP3PInstructions.td (+1-1) 
- (added) llvm/test/CodeGen/AMDGPU/mad-mix-bf16-t16-inline-imm.mir (+68) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
index 714f7f7ece7b0..e9b6731b1623b 100644
--- a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -549,7 +549,7 @@ defm V_FMA_MIXHI_BF16 : VOP3_VOP3PInst<"v_fma_mixhi_bf16", VOP3P_Mix_Profile<VOP
 }
 
 // Pseudo true16 inst for v_fma_mixlo/hi_bf16
-defm V_FMA_MIX_BF16_t16 : VOP3_VOP3PInst_t16<"v_fma_mix_bf16_t16", VOP3P_Mix_Profile_t16<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
+defm V_FMA_MIX_BF16_t16 : VOP3_VOP3PInst_t16<"v_fma_mix_bf16_t16", VOP3P_Mix_Profile_t16<VOP_BF16_BF16_BF16_BF16, VOP3_OPSEL>>;
 } // End FPDPRounding = 1
 } // End isCommutable = 1
 
diff --git a/llvm/test/CodeGen/AMDGPU/mad-mix-bf16-t16-inline-imm.mir b/llvm/test/CodeGen/AMDGPU/mad-mix-bf16-t16-inline-imm.mir
new file mode 100644
index 0000000000000..a6fd68df66b9b
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/mad-mix-bf16-t16-inline-imm.mir
@@ -0,0 +1,68 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
+# RUN: llc -mtriple=amdgpu12.50 -mattr=+real-true16 -run-pass=si-fold-operands -o - %s | FileCheck %s
+
+# Check that the bf16 encoding of 1.0 is recognized as an inline constant in
+# each source of v_fma_mix_bf16_t16.
+
+---
+name:            fold_bf16_one_into_src0
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+
+    ; CHECK-LABEL: name: fold_bf16_one_into_src0
+    ; CHECK: liveins: $vgpr0, $vgpr1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    ; CHECK-NEXT: [[V_FMA_MIX_BF16_t16_:%[0-9]+]]:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 4, 16256, 0, [[COPY]], 0, [[COPY1]], 0, 0, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0, implicit [[V_FMA_MIX_BF16_t16_]]
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vgpr_32 = COPY $vgpr1
+    %2:sreg_32 = S_MOV_B32 16256
+    %3:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 0, %2, 0, %0, 0, %1, 0, 0, 0, implicit $mode, implicit $exec
+    S_ENDPGM 0, implicit %3
+...
+
+---
+name:            fold_bf16_one_into_src1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+
+    ; CHECK-LABEL: name: fold_bf16_one_into_src1
+    ; CHECK: liveins: $vgpr0, $vgpr1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    ; CHECK-NEXT: [[V_FMA_MIX_BF16_t16_:%[0-9]+]]:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 0, [[COPY]], 0, 16256, 0, [[COPY1]], 0, 0, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0, implicit [[V_FMA_MIX_BF16_t16_]]
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vgpr_32 = COPY $vgpr1
+    %2:sreg_32 = S_MOV_B32 16256
+    %3:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 0, %0, 0, %2, 0, %1, 0, 0, 0, implicit $mode, implicit $exec
+    S_ENDPGM 0, implicit %3
+...
+
+---
+name:            fold_bf16_one_into_src2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+
+    ; CHECK-LABEL: name: fold_bf16_one_into_src2
+    ; CHECK: liveins: $vgpr0, $vgpr1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    ; CHECK-NEXT: [[V_FMA_MIX_BF16_t16_:%[0-9]+]]:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 0, [[COPY]], 0, [[COPY1]], 0, 16256, 0, 0, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0, implicit [[V_FMA_MIX_BF16_t16_]]
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vgpr_32 = COPY $vgpr1
+    %2:sreg_32 = S_MOV_B32 16256
+    %3:vgpr_16 = nofpexcept V_FMA_MIX_BF16_t16 0, %0, 0, %1, 0, %2, 0, 0, 0, implicit $mode, implicit $exec
+    S_ENDPGM 0, implicit %3
+...

``````````

</details>


https://github.com/llvm/llvm-project/pull/224910