AMDGPU: Add minimumnum/maximumnum to list of canonicalizing opcodes

[arsenm]

This makes no difference in the test, as these always expand now.

[arsenm]

• AMDGPU: Add minimumnum/maximumnum to list of canonicalizing opcodes #139124 👈 (View in Graphite)
• main

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

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

This makes no difference in the test, as these always expand now.

---

Full diff: https://github.com/llvm/llvm-project/pull/139124.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+5-1)
• (modified) llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll (+60)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 76c4546f1207e..4b2ff8fee0a48 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -13145,6 +13145,8 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
   case ISD::FMAXNUM_IEEE:
   case ISD::FMINIMUM:
   case ISD::FMAXIMUM:
+  case ISD::FMINIMUMNUM:
+  case ISD::FMAXIMUMNUM:
   case AMDGPUISD::CLAMP:
   case AMDGPUISD::FMED3:
   case AMDGPUISD::FMAX3:
@@ -13309,7 +13311,9 @@ bool SITargetLowering::isCanonicalized(Register Reg, const MachineFunction &MF,
   case AMDGPU::G_FMINNUM_IEEE:
   case AMDGPU::G_FMAXNUM_IEEE:
   case AMDGPU::G_FMINIMUM:
-  case AMDGPU::G_FMAXIMUM: {
+  case AMDGPU::G_FMAXIMUM:
+  case AMDGPU::G_FMINIMUMNUM:
+  case AMDGPU::G_FMAXIMUMNUM: {
     if (Subtarget->supportsMinMaxDenormModes() ||
         // FIXME: denormalsEnabledForType is broken for dynamic
         denormalsEnabledForType(MRI.getType(Reg), MF))
diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
index c57ddf399b53f..ab476dd96c707 100644
--- a/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
@@ -855,6 +855,66 @@ define float @v_test_canonicalize_amdgcn_exp2(float %a) {
   ret float %canonicalized
 }
 
+; GCN-LABEL: {{^}}v_test_canonicalize_minimum:
+; GCN: s_waitcnt
+; GCN-NEXT: v_min_f32_e32 [[MIN:v[0-9]+]], v0, v1
+; GCN-NEXT: v_mov_b32_e32 [[K:v[0-9]+]], 0x7fc00000
+; GCN-NEXT: v_cmp_o_f32_e32 vcc, v0, v1
+; GCN-NEXT: v_cndmask_b32_e32 v0, [[K]], [[MIN]], vcc
+; VI-FLUSH-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT: s_setpc_b64
+define float @v_test_canonicalize_minimum(float %a, float %b) {
+  %min = call float @llvm.minimum.f32(float %a, float %b)
+  %canonicalized = call float @llvm.canonicalize.f32(float %min)
+  ret float %canonicalized
+}
+
+; GCN-LABEL: {{^}}v_test_canonicalize_maximum:
+; GCN: s_waitcnt
+; GCN-NEXT: v_max_f32_e32 [[MIN:v[0-9]+]], v0, v1
+; GCN-NEXT: v_mov_b32_e32 [[K:v[0-9]+]], 0x7fc00000
+; GCN-NEXT: v_cmp_o_f32_e32 vcc, v0, v1
+; GCN-NEXT: v_cndmask_b32_e32 v0, [[K]], [[MIN]], vcc
+; VI-FLUSH-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT: s_setpc_b64
+define float @v_test_canonicalize_maximum(float %a, float %b) {
+  %min = call float @llvm.maximum.f32(float %a, float %b)
+  %canonicalized = call float @llvm.canonicalize.f32(float %min)
+  ret float %canonicalized
+}
+
+; GCN-LABEL: {{^}}v_test_canonicalize_minimumnum:
+; GCN: s_waitcnt
+; VI-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; VI-NEXT: v_mul_f32_e32 v0, 1.0, v0
+
+; GFX9-NEXT: v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT: v_max_f32_e32 v0, v0, v0
+
+; GCN-NEXT: v_min_f32_e32 v0, v0, v1
+; GCN-NEXT: s_setpc_b64
+define float @v_test_canonicalize_minimumnum(float %a, float %b) {
+  %min = call float @llvm.minimumnum.f32(float %a, float %b)
+  %canonicalized = call float @llvm.canonicalize.f32(float %min)
+  ret float %canonicalized
+}
+
+; GCN-LABEL: {{^}}v_test_canonicalize_maximumnum:
+; GCN: s_waitcnt
+; VI-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; VI-NEXT: v_mul_f32_e32 v0, 1.0, v0
+
+; GFX9-NEXT: v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT: v_max_f32_e32 v0, v0, v0
+
+; GCN-NEXT: v_max_f32_e32 v0, v0, v1
+; GCN-NEXT: s_setpc_b64
+define float @v_test_canonicalize_maximumnum(float %a, float %b) {
+  %min = call float @llvm.maximumnum.f32(float %a, float %b)
+  %canonicalized = call float @llvm.canonicalize.f32(float %min)
+  ret float %canonicalized
+}
+
 ; Avoid failing the test on FreeBSD11.0 which will match the GCN-NOT: 1.0
 ; in the .amd_amdgpu_isa "amdgcn-unknown-freebsd11.0--gfx802" directive
 ; GCN: .amd_amdgpu_isa