From 744c0057e7dc0d1d046a4867cece2f31fee9bb23 Mon Sep 17 00:00:00 2001
From: Nashe Mncube <nashe.mncube@arm.com>
Date: Thu, 22 Feb 2024 19:15:52 +0000
Subject: [AArch64][CodeGen] Fix crash when fptrunc returns fp16 with +nofp
 attr (#81724)

When performing lowering of the fptrunc opcode returning fp16 with the
+nofp flag enabled we could trigger a compiler crash. This is because we
had no custom lowering implemented. This patch
the case in which we need to promote an fp16 return type
for fptrunc when the +nofp attr is enabled.
---
 .../AArch64/16bit-float-promotion-with-nofp.ll     |  31 ++++++
 .../CodeGen/AArch64/strictfp_f16_abi_promote.ll    | 115 ++++++++++++++++++---
 2 files changed, 129 insertions(+), 17 deletions(-)
 create mode 100644 llvm/test/CodeGen/AArch64/16bit-float-promotion-with-nofp.ll

(limited to 'llvm/test')

diff --git a/llvm/test/CodeGen/AArch64/16bit-float-promotion-with-nofp.ll b/llvm/test/CodeGen/AArch64/16bit-float-promotion-with-nofp.ll
new file mode 100644
index 0000000..bfe9ab8
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/16bit-float-promotion-with-nofp.ll
@@ -0,0 +1,31 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=aarch64 -mattr=-fp-armv8 -o - %s | FileCheck %s
+
+define half @f2h(float %a) {
+; CHECK-LABEL: f2h:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    bl __gnu_f2h_ieee
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %0 = fptrunc float %a to half
+  ret half %0
+}
+
+define bfloat @f2bfloat(float %a) {
+; CHECK-LABEL: f2bfloat:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    bl __truncsfbf2
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %0 = fptrunc float %a to bfloat
+  ret bfloat %0
+}
+
diff --git a/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll b/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
index a34f7ab..9fa5208 100644
--- a/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
+++ b/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
@@ -131,26 +131,107 @@ define void @v4f16_arg(<4 x half> %arg, ptr %ptr) #0 {
   ret void
 }
 
-; FIXME:
-; define half @f16_return(float %arg) #0 {
-;   %fptrunc = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
-;   ret half %fptrunc
-; }
+ define half @f16_return(float %arg) #0 {
+; NOFP16-LABEL: f16_return:
+; NOFP16:       // %bb.0:
+; NOFP16-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; NOFP16-NEXT:    .cfi_def_cfa_offset 16
+; NOFP16-NEXT:    .cfi_offset w30, -16
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; NOFP16-NEXT:    ret
+   %fptrunc = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
+   ret half %fptrunc
+ }
 
-; define <2 x half> @v2f16_return(<2 x float> %arg) #0 {
-;   %fptrunc = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
-;   ret <2 x half> %fptrunc
-; }
+ define <2 x half> @v2f16_return(<2 x float> %arg) #0 {
+; NOFP16-LABEL: v2f16_return:
+; NOFP16:       // %bb.0:
+; NOFP16-NEXT:    str x30, [sp, #-32]! // 8-byte Folded Spill
+; NOFP16-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
+; NOFP16-NEXT:    .cfi_def_cfa_offset 32
+; NOFP16-NEXT:    .cfi_offset w19, -8
+; NOFP16-NEXT:    .cfi_offset w20, -16
+; NOFP16-NEXT:    .cfi_offset w30, -32
+; NOFP16-NEXT:    mov w19, w0
+; NOFP16-NEXT:    mov w0, w1
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w20, w0
+; NOFP16-NEXT:    mov w0, w19
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w1, w20
+; NOFP16-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
+; NOFP16-NEXT:    ldr x30, [sp], #32 // 8-byte Folded Reload
+; NOFP16-NEXT:    ret
+   %fptrunc = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
+   ret <2 x half> %fptrunc
+ }
 
-; define <3 x half> @v3f16_return(<3 x float> %arg) #0 {
-;   %fptrunc = call <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
-;   ret <3 x half> %fptrunc
-; }
+ define <3 x half> @v3f16_return(<3 x float> %arg) #0 {
+; NOFP16-LABEL: v3f16_return:
+; NOFP16:       // %bb.0:
+; NOFP16-NEXT:    stp x30, x21, [sp, #-32]! // 16-byte Folded Spill
+; NOFP16-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
+; NOFP16-NEXT:    .cfi_def_cfa_offset 32
+; NOFP16-NEXT:    .cfi_offset w19, -8
+; NOFP16-NEXT:    .cfi_offset w20, -16
+; NOFP16-NEXT:    .cfi_offset w21, -24
+; NOFP16-NEXT:    .cfi_offset w30, -32
+; NOFP16-NEXT:    mov w20, w0
+; NOFP16-NEXT:    mov w0, w2
+; NOFP16-NEXT:    mov w19, w1
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w21, w0
+; NOFP16-NEXT:    mov w0, w19
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w19, w0
+; NOFP16-NEXT:    mov w0, w20
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w1, w19
+; NOFP16-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
+; NOFP16-NEXT:    mov w2, w21
+; NOFP16-NEXT:    ldp x30, x21, [sp], #32 // 16-byte Folded Reload
+; NOFP16-NEXT:    ret
+   %fptrunc = call <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
+   ret <3 x half> %fptrunc
+ }
 
-; define <4 x half> @v4f16_return(<4 x float> %arg) #0 {
-;   %fptrunc = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
-;   ret <4 x half> %fptrunc
-; }
+ define <4 x half> @v4f16_return(<4 x float> %arg) #0 {
+; NOFP16-LABEL: v4f16_return:
+; NOFP16:       // %bb.0:
+; NOFP16-NEXT:    str x30, [sp, #-48]! // 8-byte Folded Spill
+; NOFP16-NEXT:    stp x22, x21, [sp, #16] // 16-byte Folded Spill
+; NOFP16-NEXT:    stp x20, x19, [sp, #32] // 16-byte Folded Spill
+; NOFP16-NEXT:    .cfi_def_cfa_offset 48
+; NOFP16-NEXT:    .cfi_offset w19, -8
+; NOFP16-NEXT:    .cfi_offset w20, -16
+; NOFP16-NEXT:    .cfi_offset w21, -24
+; NOFP16-NEXT:    .cfi_offset w22, -32
+; NOFP16-NEXT:    .cfi_offset w30, -48
+; NOFP16-NEXT:    mov w21, w0
+; NOFP16-NEXT:    mov w0, w3
+; NOFP16-NEXT:    mov w19, w2
+; NOFP16-NEXT:    mov w20, w1
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w22, w0
+; NOFP16-NEXT:    mov w0, w19
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w19, w0
+; NOFP16-NEXT:    mov w0, w20
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w20, w0
+; NOFP16-NEXT:    mov w0, w21
+; NOFP16-NEXT:    bl __gnu_f2h_ieee
+; NOFP16-NEXT:    mov w1, w20
+; NOFP16-NEXT:    mov w2, w19
+; NOFP16-NEXT:    mov w3, w22
+; NOFP16-NEXT:    ldp x20, x19, [sp, #32] // 16-byte Folded Reload
+; NOFP16-NEXT:    ldp x22, x21, [sp, #16] // 16-byte Folded Reload
+; NOFP16-NEXT:    ldr x30, [sp], #48 // 8-byte Folded Reload
+; NOFP16-NEXT:    ret
+   %fptrunc = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")
+   ret <4 x half> %fptrunc
+ }
 
 ; FIXME:
 ; define void @outgoing_f16_arg(ptr %ptr) #0 {
-- 
cgit v1.1