[AMDGPU] performMinMaxCombine should not optimize patterns of vectors to min3/max3.

Summary: There are no packed instructions for min3 or max3. So, performMinMaxCombine should not optimize vectors of f16 to min3/max3.

Author: FarhanaAleen

Reviewed By: arsenm

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D45219

llvm-svn: 329131

3 files changed, 59 insertions, 1 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index f0d5e926e3c..2a7549ed77f 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6446,7 +6446,7 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
 
 
   if (Opc != AMDGPUISD::FMIN_LEGACY && Opc != AMDGPUISD::FMAX_LEGACY &&
-      VT != MVT::f64 &&
+      !VT.isVector() && VT != MVT::f64 &&
       ((VT != MVT::f16 && VT != MVT::i16) || Subtarget->hasMin3Max3_16())) {
     // max(max(a, b), c) -> max3(a, b, c)
     // min(min(a, b), c) -> min3(a, b, c)
diff --git a/llvm/test/CodeGen/AMDGPU/fmax3.ll b/llvm/test/CodeGen/AMDGPU/fmax3.ll
index 2e6d3f3c1e8..937bd74a0fe 100644
--- a/llvm/test/CodeGen/AMDGPU/fmax3.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmax3.ll
@@ -84,9 +84,38 @@ define amdgpu_kernel void @test_fmax3_olt_1_f16(half addrspace(1)* %out, half ad
   ret void
 }
 
+; Checks whether the test passes; performMinMaxCombine() should not optimize vector patterns of max3
+; since there are no pack instructions for fmax3.
+; GCN-LABEL: {{^}}no_fmax3_v2f16:
+
+; SI: v_cvt_f16_f32_e32
+; SI: v_max_f32_e32
+; SI-NEXT: v_max_f32_e32
+; SI-NEXT: v_max3_f32
+; SI-NEXT: v_max3_f32
+
+; VI: v_max_f16_e32
+; VI-NEXT: v_max_f16_e32
+; VI-NEXT: v_max_f16_e32
+; VI-NEXT: v_max_f16_e32
+; VI-NEXT: v_max_f16_e32
+; VI-NEXT: v_max_f16_e32
+
+; GFX9: v_pk_max_f16
+; GFX9-NEXT: v_pk_max_f16
+; GFX9-NEXT: v_pk_max_f16
+define <2 x half> @no_fmax3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) {
+entry:
+  %max = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)
+  %max1 = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %c, <2 x half> %max)
+  %res = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %max1, <2 x half> %d)
+  ret <2 x half> %res
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.maxnum.f32(float, float) #1
 declare half @llvm.maxnum.f16(half, half) #1
+declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>)
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone speculatable }
diff --git a/llvm/test/CodeGen/AMDGPU/fmin3.ll b/llvm/test/CodeGen/AMDGPU/fmin3.ll
index 5fc5895c3ec..d22333384dc 100644
--- a/llvm/test/CodeGen/AMDGPU/fmin3.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmin3.ll
@@ -82,9 +82,38 @@ define amdgpu_kernel void @test_fmin3_olt_1_f16(half addrspace(1)* %out, half ad
   ret void
 }
 
+; Checks whether the test passes; performMinMaxCombine() should not optimize vector patterns of min3
+; since there are no pack instructions for fmin3.
+; GCN-LABEL: {{^}}no_fmin3_v2f16:
+
+; SI: v_cvt_f16_f32_e32
+; SI: v_min_f32_e32
+; SI-NEXT: v_min_f32_e32
+; SI-NEXT: v_min3_f32
+; SI-NEXT: v_min3_f32
+
+; VI: v_min_f16_e32
+; VI-NEXT: v_min_f16_e32
+; VI-NEXT: v_min_f16_e32
+; VI-NEXT: v_min_f16_e32
+; VI-NEXT: v_min_f16_e32
+; VI-NEXT: v_min_f16_e32
+
+; GFX9: v_pk_min_f16
+; GFX9: v_pk_min_f16
+; GFX9: v_pk_min_f16
+define <2 x half> @no_fmin3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) {
+entry:
+  %min = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %a, <2 x half> %b)
+  %min1 = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %c, <2 x half> %min)
+  %res = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %min1, <2 x half> %d)
+  ret <2 x half> %res
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.minnum.f32(float, float) #1
 declare half @llvm.minnum.f16(half, half) #1
+declare <2 x half> @llvm.minnum.v2f16(<2 x half>, <2 x half>)
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone speculatable }