[AArch64] Legalize horizontal fmax/fmin reductions on f16 vectors

Expand the horizontal reduction during the instruction selection phase, but only if the target doesn't support the full fp16 instruction set.

Fixes https://bugs.llvm.org/show_bug.cgi?id=49401

Reviewed By: aemerson

Differential Revision: https://reviews.llvm.org/D97840

(cherry picked from commit 8725b24c6d4abaa97425e704652a13dacb35fe3f)

3 files changed, 126 insertions, 5 deletions

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 1be09186dc0a..1451151f4dc5 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -1017,11 +1017,12 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     // Vector reductions
     for (MVT VT : { MVT::v4f16, MVT::v2f32,
                     MVT::v8f16, MVT::v4f32, MVT::v2f64 }) {
-      setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
-      setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
+      if (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16()) {
+        setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
+        setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
 
-      if (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16())
         setOperationAction(ISD::VECREDUCE_FADD, VT, Legal);
+      }
     }
     for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32,
                     MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
diff --git a/llvm/test/CodeGen/AArch64/vecreduce-fmax-legalization.ll b/llvm/test/CodeGen/AArch64/vecreduce-fmax-legalization.ll
index f1ebd8fa85ea..d26db2aefee0 100644
--- a/llvm/test/CodeGen/AArch64/vecreduce-fmax-legalization.ll
+++ b/llvm/test/CodeGen/AArch64/vecreduce-fmax-legalization.ll
@@ -1,11 +1,13 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s --check-prefix=CHECK
+; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-NOFP
+; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon,+fullfp16 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FP
 
 declare half @llvm.vector.reduce.fmax.v1f16(<1 x half> %a)
 declare float @llvm.vector.reduce.fmax.v1f32(<1 x float> %a)
 declare double @llvm.vector.reduce.fmax.v1f64(<1 x double> %a)
 declare fp128 @llvm.vector.reduce.fmax.v1f128(<1 x fp128> %a)
 
+declare half @llvm.vector.reduce.fmax.v4f16(<4 x half> %a)
 declare float @llvm.vector.reduce.fmax.v3f32(<3 x float> %a)
 declare fp128 @llvm.vector.reduce.fmax.v2f128(<2 x fp128> %a)
 declare float @llvm.vector.reduce.fmax.v16f32(<16 x float> %a)
@@ -44,6 +46,64 @@ define fp128 @test_v1f128(<1 x fp128> %a) nounwind {
   ret fp128 %b
 }
 
+define half @test_v4f16(<4 x half> %a) nounwind {
+; CHECK-NOFP-LABEL: test_v4f16:
+; CHECK-NOFP:       // %bb.0:
+; CHECK-NOFP-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-NOFP-NEXT:    mov h3, v0.h[1]
+; CHECK-NOFP-NEXT:    mov h1, v0.h[3]
+; CHECK-NOFP-NEXT:    mov h2, v0.h[2]
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s3, h3
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s3
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s2, h2
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s2
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s1, h1
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s1
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    ret
+;
+; CHECK-FP-LABEL: test_v4f16:
+; CHECK-FP:       // %bb.0:
+; CHECK-FP-NEXT:    fmaxnmv h0, v0.4h
+; CHECK-FP-NEXT:    ret
+  %b = call nnan half @llvm.vector.reduce.fmax.v4f16(<4 x half> %a)
+  ret half %b
+}
+
+define half @test_v4f16_ninf(<4 x half> %a) nounwind {
+; CHECK-NOFP-LABEL: test_v4f16_ninf:
+; CHECK-NOFP:       // %bb.0:
+; CHECK-NOFP-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-NOFP-NEXT:    mov h3, v0.h[1]
+; CHECK-NOFP-NEXT:    mov h1, v0.h[3]
+; CHECK-NOFP-NEXT:    mov h2, v0.h[2]
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s3, h3
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s3
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s2, h2
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s2
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s1, h1
+; CHECK-NOFP-NEXT:    fmaxnm s0, s0, s1
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    ret
+;
+; CHECK-FP-LABEL: test_v4f16_ninf:
+; CHECK-FP:       // %bb.0:
+; CHECK-FP-NEXT:    fmaxnmv h0, v0.4h
+; CHECK-FP-NEXT:    ret
+  %b = call nnan ninf half @llvm.vector.reduce.fmax.v4f16(<4 x half> %a)
+  ret half %b
+}
+
 define float @test_v3f32(<3 x float> %a) nounwind {
 ; CHECK-LABEL: test_v3f32:
 ; CHECK:       // %bb.0:
diff --git a/llvm/test/CodeGen/AArch64/vecreduce-fmin-legalization.ll b/llvm/test/CodeGen/AArch64/vecreduce-fmin-legalization.ll
index 4129fa80b13e..52d6e9773ab2 100644
--- a/llvm/test/CodeGen/AArch64/vecreduce-fmin-legalization.ll
+++ b/llvm/test/CodeGen/AArch64/vecreduce-fmin-legalization.ll
@@ -1,11 +1,13 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s --check-prefix=CHECK
+; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-NOFP
+; RUN: llc < %s -mtriple=aarch64-none-linux-gnu -mattr=+neon,+fullfp16 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FP
 
 declare half @llvm.vector.reduce.fmin.v1f16(<1 x half> %a)
 declare float @llvm.vector.reduce.fmin.v1f32(<1 x float> %a)
 declare double @llvm.vector.reduce.fmin.v1f64(<1 x double> %a)
 declare fp128 @llvm.vector.reduce.fmin.v1f128(<1 x fp128> %a)
 
+declare half @llvm.vector.reduce.fmin.v4f16(<4 x half> %a)
 declare float @llvm.vector.reduce.fmin.v3f32(<3 x float> %a)
 declare fp128 @llvm.vector.reduce.fmin.v2f128(<2 x fp128> %a)
 declare float @llvm.vector.reduce.fmin.v16f32(<16 x float> %a)
@@ -44,6 +46,64 @@ define fp128 @test_v1f128(<1 x fp128> %a) nounwind {
   ret fp128 %b
 }
 
+define half @test_v4f16(<4 x half> %a) nounwind {
+; CHECK-NOFP-LABEL: test_v4f16:
+; CHECK-NOFP:       // %bb.0:
+; CHECK-NOFP-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-NOFP-NEXT:    mov h3, v0.h[1]
+; CHECK-NOFP-NEXT:    mov h1, v0.h[3]
+; CHECK-NOFP-NEXT:    mov h2, v0.h[2]
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s3, h3
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s3
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s2, h2
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s2
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s1, h1
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s1
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    ret
+;
+; CHECK-FP-LABEL: test_v4f16:
+; CHECK-FP:       // %bb.0:
+; CHECK-FP-NEXT:    fminnmv h0, v0.4h
+; CHECK-FP-NEXT:    ret
+  %b = call nnan half @llvm.vector.reduce.fmin.v4f16(<4 x half> %a)
+  ret half %b
+}
+
+define half @test_v4f16_ninf(<4 x half> %a) nounwind {
+; CHECK-NOFP-LABEL: test_v4f16_ninf:
+; CHECK-NOFP:       // %bb.0:
+; CHECK-NOFP-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-NOFP-NEXT:    mov h3, v0.h[1]
+; CHECK-NOFP-NEXT:    mov h1, v0.h[3]
+; CHECK-NOFP-NEXT:    mov h2, v0.h[2]
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s3, h3
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s3
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s2, h2
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s2
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    fcvt s0, h0
+; CHECK-NOFP-NEXT:    fcvt s1, h1
+; CHECK-NOFP-NEXT:    fminnm s0, s0, s1
+; CHECK-NOFP-NEXT:    fcvt h0, s0
+; CHECK-NOFP-NEXT:    ret
+;
+; CHECK-FP-LABEL: test_v4f16_ninf:
+; CHECK-FP:       // %bb.0:
+; CHECK-FP-NEXT:    fminnmv h0, v0.4h
+; CHECK-FP-NEXT:    ret
+  %b = call nnan ninf half @llvm.vector.reduce.fmin.v4f16(<4 x half> %a)
+  ret half %b
+}
+
 define float @test_v3f32(<3 x float> %a) nounwind {
 ; CHECK-LABEL: test_v3f32:
 ; CHECK:       // %bb.0: