[llvm] [X86] Prevent APX NDD compression when it creates a partial write (PR #132051)

[Phoebe Wang via llvm-commits]

================
@@ -0,0 +1,89 @@
+# NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+# RUN: llc %s -mtriple=x86_64-unknown -mattr=+ndd,+egpr -start-before=break-false-deps -verify-machineinstrs -show-mc-encoding -o - | FileCheck --check-prefixes=ASM,RCDEFAULT %s
+# RUN: llc %s -mtriple=x86_64-unknown -mattr=+ndd,+egpr -start-before=break-false-deps -verify-machineinstrs -partial-reg-update-clearance=1 -show-mc-encoding -o - | FileCheck --check-prefixes=ASM,RC1 %s
+#
+# Check that BreakFalseDeps detects cases where an ND instruction would cause a partial register write
+# if compressed to a legacy op. MIR has been modified to force different register assignments.
+#
+# For partial_write, the ADD16rr_ND is compressible, but will become a partial write after compression.
+# Compression is inhibited if the eax definition is within the partial-reg-update-clearance threshold.
+#
+# For no_partial_write, the ADD16rr_ND is incompressible hence it cannot become a partial write.
+# This case checks that an implicit-def of eax is not added by breakPartialRegDependency.
+#
+--- |
+  define signext i16 @partial_write(ptr %p, i32 %a, i32 %b, i16 signext %x, i16 signext %y) #0 {
+  ; RCDEFAULT-LABEL: partial_write:
+  ; RCDEFAULT:       # %bb.0: # %entry
+  ; RCDEFAULT-NEXT:    addl %esi, %edx, %eax # encoding: [0x62,0xf4,0x7c,0x18,0x01,0xf2]
+  ; RCDEFAULT-NEXT:    movl %eax, (%rdi) # encoding: [0x89,0x07]
+  ; RCDEFAULT-NEXT:    addw %cx, %ax, %ax # encoding: [0x62,0xf4,0x7d,0x18,0x01,0xc8]
+  ; RCDEFAULT-NEXT:    retq # encoding: [0xc3]
+  ;
+  ; RC1-LABEL: partial_write:
+  ; RC1:       # %bb.0: # %entry
+  ; RC1-NEXT:    addl %esi, %edx, %eax # encoding: [0x62,0xf4,0x7c,0x18,0x01,0xf2]
+  ; RC1-NEXT:    movl %eax, (%rdi) # encoding: [0x89,0x07]
+  ; RC1-NEXT:    addw %cx, %ax # EVEX TO LEGACY Compression encoding: [0x66,0x01,0xc8]
+  ; RC1-NEXT:    retq # encoding: [0xc3]
+  entry:
+    %add = add nsw i32 %b, %a
+    store i32 %add, ptr %p, align 4, !tbaa !1
+    %add1 = trunc i32 %add to i16
+    %add2 = add i16 %add1, %x
+    ret i16 %add2
+  }
+
+  define signext i16 @no_partial_write(ptr %p, i32 %a, i32 %b, i16 signext %x, i16 signext %y) #0 {
+  ; ASM-LABEL: no_partial_write:
+  ; ASM:       # %bb.0: # %entry
+  ; ASM-NEXT:    addl %esi, %edx # EVEX TO LEGACY Compression encoding: [0x01,0xf2]
+  ; ASM-NEXT:    movl %edx, (%rdi) # encoding: [0x89,0x17]
+  ; ASM-NEXT:    addw %cx, %dx, %ax # encoding: [0x62,0xf4,0x7d,0x18,0x01,0xca]
+  ; ASM-NEXT:    retq # encoding: [0xc3]
+  entry:
+    %add = add nsw i32 %b, %a
+    store i32 %add, ptr %p, align 4, !tbaa !1
+    %add1 = trunc i32 %add to i16
+    %add2 = add i16 %add1, %x
+    ret i16 %add2
+  }
+  attributes #0 = { mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write) uwtable "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx16,+cx8,+egpr,+fxsr,+mmx,+ndd,+sse,+sse2,+x87" "tune-cpu"="generic" }
+
+  !llvm.module.flags = !{!0}
+
+  !0 = !{i32 1, !"wchar_size", i32 4}
+  !1 = !{!2, !2, i64 0}
+  !2 = !{!"int", !3, i64 0}
+  !3 = !{!"omnipotent char", !4, i64 0}
+  !4 = !{!"Simple C/C++ TBAA"}
----------------
phoebewang wrote:

Nit: These are not needed.

https://github.com/llvm/llvm-project/pull/132051