[AArch64][GISel] Additional FPTrunc vector lowering

[dmgreen]

Summary

I was attempting to add llvm.reduce.fminimum/fmaximum support for GlobalISel. In the process I noticed that llvm.reduce.fmin/fmax was missing, and could do with being added first. That led on to adding additional vector support for minnum/maxnum, which in turn led to needing to handle fptrunc and fpext for some of the fp16 types. So this patch extends the vector handling for fptrunc, adding support for f16 types which are clamped to 4 elements, and scalarizing the rest.

I went round in circles a little with how smaller than legal vectors should be handled, but this seems simple and seems to work, if not always optimally yet.

[arsenm]

llvm/test/CodeGen/AArch64/fptrunc.ll

3

should use explicit =0/=1 so it doesn't break when the default switches

36

v3 tests often find interesting broken things

[dmgreen]

llvm/test/CodeGen/AArch64/fptrunc.ll

36

Thanks. Will do. I was focusing on power-2 types with power-2 vector sizes, but it makes sense to include some non-power-2 vectors too.

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx,
   }
   case TargetOpcode::G_TRUNC: {
     Observer.changingInstr(MI);
     moreElementsVectorSrc(MI, MoreTy, 1);
     moreElementsVectorDst(MI, MoreTy, 0);
     Observer.changedInstr(MI);
     return Legalized;
   }
+  case TargetOpcode::G_FPTRUNC: {
+    if (TypeIdx != 0)
+      return UnableToLegalize;
+    Observer.changingInstr(MI);
+    LLT SrcTy = LLT::fixed_vector(
+        MoreTy.getNumElements(),
+        MRI.getType(MI.getOperand(1).getReg()).getElementType());
+    moreElementsVectorSrc(MI, SrcTy, 1);
+    moreElementsVectorDst(MI, MoreTy, 0);
+    Observer.changedInstr(MI);
+    return Legalized;
+  }
   default:
     return UnableToLegalize;
   }
 }
 
 LegalizerHelper::LegalizeResult
 LegalizerHelper::equalizeVectorShuffleLengths(MachineInstr &MI) {
   auto [DstReg, DstTy, SrcReg, SrcTy] = MI.getFirst2RegLLTs();

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

getActionDefinitionsBuilder(G_SEXT_INREG)
       .legalFor({s32, s64})
       .legalFor(PackedVectorAllTypeList)
       .lower();
 
   // FP conversions
   getActionDefinitionsBuilder(G_FPTRUNC)
       .legalFor(
           {{s16, s32}, {s16, s64}, {s32, s64}, {v4s16, v4s32}, {v2s32, v2s64}})
-      .clampMaxNumElements(0, s32, 2);
+      .clampNumElements(0, v4s16, v4s16)
+      .clampNumElements(0, v2s32, v2s32)
+      .scalarize(0);
+
   getActionDefinitionsBuilder(G_FPEXT)
       .legalFor(
           {{s32, s16}, {s64, s16}, {s64, s32}, {v4s32, v4s16}, {v2s64, v2s32}})
       .clampMaxNumElements(0, s64, 2);
 
   // Conversions
   getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
       .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})

llvm/test/CodeGen/AArch64/GlobalISel/legalize-fptrunc.mir

 body:             |
   bb.0:
     liveins: $d0
 
     ; CHECK-LABEL: name: fptrunc_v2s16_v2s32
     ; CHECK: liveins: $d0
     ; CHECK-NEXT: {{  $}}
     ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $d0
-    ; CHECK-NEXT: [[FPTRUNC:%[0-9]+]]:_(<2 x s16>) = G_FPTRUNC [[COPY]](<2 x s32>)
-    ; CHECK-NEXT: $s0 = COPY [[FPTRUNC]](<2 x s16>)
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; CHECK-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[UV]](s32), [[UV1]](s32), [[DEF]](s32), [[DEF]](s32)
+    ; CHECK-NEXT: [[FPTRUNC:%[0-9]+]]:_(<4 x s16>) = G_FPTRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CHECK-NEXT: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[FPTRUNC]](<4 x s16>)
+    ; CHECK-NEXT: $s0 = COPY [[UV2]](<2 x s16>)
     ; CHECK-NEXT: RET_ReallyLR implicit $s0
     %0:_(<2 x s32>) = COPY $d0
     %1:_(<2 x s16>) = G_FPTRUNC %0
     $s0 = COPY %1(<2 x s16>)
     RET_ReallyLR implicit $s0
 ...
 ---
 name:            fptrunc_v4s32_v4s64

llvm/test/CodeGen/AArch64/fptrunc.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=aarch64-none-eabi -global-isel=0 -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64-none-eabi -global-isel=1 -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+
+define float @fptrunc_f64_f32(double %a) {
+; CHECK-LABEL: fptrunc_f64_f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvt s0, d0
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc double %a to float
+  ret float %c
+}
+
+define half @fptrunc_f64_f16(double %a) {
+; CHECK-LABEL: fptrunc_f64_f16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvt h0, d0
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc double %a to half
+  ret half %c
+}
+
+define half @fptrunc_f32_f16(float %a) {
+; CHECK-LABEL: fptrunc_f32_f16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvt h0, s0
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc float %a to half
+  ret half %c
+}
+
+define <2 x float> @fptrunc_v2f64_v2f32(<2 x double> %a) {
+; CHECK-LABEL: fptrunc_v2f64_v2f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvtn v0.2s, v0.2d
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc <2 x double> %a to <2 x float>
+  ret <2 x float> %c
+}
+
+define <3 x float> @fptrunc_v3f64_v3f32(<3 x double> %a) {
+; CHECK-SD-LABEL: fptrunc_v3f64_v3f32:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-SD-NEXT:    // kill: def $d1 killed $d1 def $q1
+; CHECK-SD-NEXT:    // kill: def $d2 killed $d2 def $q2
+; CHECK-SD-NEXT:    mov v0.d[1], v1.d[0]
+; CHECK-SD-NEXT:    fcvtn v0.2s, v0.2d
+; CHECK-SD-NEXT:    fcvtn2 v0.4s, v2.2d
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v3f64_v3f32:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-GI-NEXT:    // kill: def $d1 killed $d1 def $q1
+; CHECK-GI-NEXT:    mov v0.d[1], v1.d[0]
+; CHECK-GI-NEXT:    fcvt s1, d2
+; CHECK-GI-NEXT:    fcvtn v0.2s, v0.2d
+; CHECK-GI-NEXT:    mov s2, v0.s[1]
+; CHECK-GI-NEXT:    mov v0.s[1], v2.s[0]
+; CHECK-GI-NEXT:    mov v0.s[2], v1.s[0]
+; CHECK-GI-NEXT:    mov v0.s[3], v0.s[0]
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <3 x double> %a to <3 x float>
+  ret <3 x float> %c
+}
+
+define <4 x float> @fptrunc_v4f64_v4f32(<4 x double> %a) {
+; CHECK-LABEL: fptrunc_v4f64_v4f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvtn v0.2s, v0.2d
+; CHECK-NEXT:    fcvtn2 v0.4s, v1.2d
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc <4 x double> %a to <4 x float>
+  ret <4 x float> %c
+}
+
+define <2 x half> @fptrunc_v2f64_v2f16(<2 x double> %a) {
+; CHECK-SD-LABEL: fptrunc_v2f64_v2f16:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    mov d1, v0.d[1]
+; CHECK-SD-NEXT:    fcvt h0, d0
+; CHECK-SD-NEXT:    fcvt h1, d1
+; CHECK-SD-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v2f64_v2f16:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov d1, v0.d[1]
+; CHECK-GI-NEXT:    fcvt h0, d0
+; CHECK-GI-NEXT:    fcvt h1, d1
+; CHECK-GI-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-GI-NEXT:    mov v0.h[2], v0.h[0]
+; CHECK-GI-NEXT:    mov v0.h[3], v0.h[0]
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <2 x double> %a to <2 x half>
+  ret <2 x half> %c
+}
+
+define <3 x half> @fptrunc_v3f64_v3f16(<3 x double> %a) {
+; CHECK-SD-LABEL: fptrunc_v3f64_v3f16:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fcvt h1, d1
+; CHECK-SD-NEXT:    fcvt h0, d0
+; CHECK-SD-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-SD-NEXT:    fcvt h1, d2
+; CHECK-SD-NEXT:    mov v0.h[2], v1.h[0]
+; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v3f64_v3f16:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    fcvt h0, d0
+; CHECK-GI-NEXT:    fcvt h1, d1
+; CHECK-GI-NEXT:    fcvt h2, d2
+; CHECK-GI-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-GI-NEXT:    mov v0.h[2], v2.h[0]
+; CHECK-GI-NEXT:    mov v0.h[3], v0.h[0]
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <3 x double> %a to <3 x half>
+  ret <3 x half> %c
+}
+
+define <4 x half> @fptrunc_v4f64_v4f16(<4 x double> %a) {
+; CHECK-SD-LABEL: fptrunc_v4f64_v4f16:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    mov d2, v0.d[1]
+; CHECK-SD-NEXT:    fcvt h0, d0
+; CHECK-SD-NEXT:    fcvt h2, d2
+; CHECK-SD-NEXT:    mov v0.h[1], v2.h[0]
+; CHECK-SD-NEXT:    fcvt h2, d1
+; CHECK-SD-NEXT:    mov d1, v1.d[1]
+; CHECK-SD-NEXT:    mov v0.h[2], v2.h[0]
+; CHECK-SD-NEXT:    fcvt h1, d1
+; CHECK-SD-NEXT:    mov v0.h[3], v1.h[0]
+; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v4f64_v4f16:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov d2, v0.d[1]
+; CHECK-GI-NEXT:    fcvt h0, d0
+; CHECK-GI-NEXT:    mov d3, v1.d[1]
+; CHECK-GI-NEXT:    fcvt h1, d1
+; CHECK-GI-NEXT:    fcvt h2, d2
+; CHECK-GI-NEXT:    mov v0.h[1], v2.h[0]
+; CHECK-GI-NEXT:    fcvt h2, d3
+; CHECK-GI-NEXT:    mov v0.h[2], v1.h[0]
+; CHECK-GI-NEXT:    mov v0.h[3], v2.h[0]
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <4 x double> %a to <4 x half>
+  ret <4 x half> %c
+}
+
+define <2 x half> @fptrunc_v2f32_v2f16(<2 x float> %a) {
+; CHECK-SD-LABEL: fptrunc_v2f32_v2f16:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-SD-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v2f32_v2f16:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 def $q0
+; CHECK-GI-NEXT:    mov s1, v0.s[1]
+; CHECK-GI-NEXT:    mov v0.s[1], v1.s[0]
+; CHECK-GI-NEXT:    mov v0.s[2], v0.s[0]
+; CHECK-GI-NEXT:    mov v0.s[3], v0.s[0]
+; CHECK-GI-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-GI-NEXT:    mov h1, v0.h[1]
+; CHECK-GI-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-GI-NEXT:    mov v0.h[2], v0.h[0]
+; CHECK-GI-NEXT:    mov v0.h[3], v0.h[0]
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <2 x float> %a to <2 x half>
+  ret <2 x half> %c
+}
+
+define <3 x half> @fptrunc_v3f32_v3f16(<3 x float> %a) {
+; CHECK-SD-LABEL: fptrunc_v3f32_v3f16:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fptrunc_v3f32_v3f16:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov s1, v0.s[1]
+; CHECK-GI-NEXT:    mov s2, v0.s[2]
+; CHECK-GI-NEXT:    mov v0.s[1], v1.s[0]
+; CHECK-GI-NEXT:    mov v0.s[2], v2.s[0]
+; CHECK-GI-NEXT:    mov v0.s[3], v0.s[0]
+; CHECK-GI-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-GI-NEXT:    mov h1, v0.h[1]
+; CHECK-GI-NEXT:    mov h2, v0.h[2]
+; CHECK-GI-NEXT:    mov v0.h[1], v1.h[0]
+; CHECK-GI-NEXT:    mov v0.h[2], v2.h[0]
+; CHECK-GI-NEXT:    mov v0.h[3], v0.h[0]
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = fptrunc <3 x float> %a to <3 x half>
+  ret <3 x half> %c
+}
+
+define <4 x half> @fptrunc_v4f32_v4f16(<4 x float> %a) {
+; CHECK-LABEL: fptrunc_v4f32_v4f16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc <4 x float> %a to <4 x half>
+  ret <4 x half> %c
+}
+
+define <8 x half> @fptrunc_v8f32_v8f16(<8 x float> %a) {
+; CHECK-LABEL: fptrunc_v8f32_v8f16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-NEXT:    fcvtn2 v0.8h, v1.4s
+; CHECK-NEXT:    ret
+entry:
+  %c = fptrunc <8 x float> %a to <8 x half>
+  ret <8 x half> %c
+}