[llvm] [X86][APX] Reuse EFLAGS across multi-predecessor blocks via NF in optimizeCompareInstr (PR #208184)

[Phoebe Wang via llvm-commits]

================
@@ -0,0 +1,353 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -o - %s -mtriple=x86_64-- -run-pass peephole-opt -mattr=+nf | FileCheck %s
+
+# MIR-level coverage for the multi-predecessor EFLAGS reuse in
+# optimizeCompareInstr / findDominatingRedundantFlagInstr. These CFG shapes are
+# awkward to produce from IR (LICM hoists loop-invariant compares, the scheduler
+# repositions EFLAGS clobbers), so they are written directly as MIR.
+
+# A non-NF-convertible EFLAGS clobber in the dominating block that defines the
+# flags shadows the producer, so the redundant compare in bb.3 must be kept.
+---
+name: dom_shadow_clobber
+tracksRegLiveness: true
+body: |
+  ; CHECK-LABEL: name: dom_shadow_clobber
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; CHECK-NEXT:   liveins: $edi, $esi
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[COPY:%[0-9]+]]:gr32 = COPY $edi
+  ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+  ; CHECK-NEXT:   CMP32ri [[COPY]], 2, implicit-def $eflags
+  ; CHECK-NEXT:   CMP32ri [[COPY1]], 7, implicit-def $eflags
+  ; CHECK-NEXT:   JCC_1 %bb.2, 12, implicit $eflags
+  ; CHECK-NEXT:   JMP_1 %bb.1
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[IMUL32rr:%[0-9]+]]:gr32 = nuw nsw IMUL32rr [[COPY]], [[COPY1]], implicit-def dead $eflags
+  ; CHECK-NEXT:   JMP_1 %bb.3
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.3:
+  ; CHECK-NEXT:   CMP32ri [[COPY]], 2, implicit-def $eflags
+  ; CHECK-NEXT:   $al = SETCCr 15, implicit $eflags
+  ; CHECK-NEXT:   RET 0, $al
+  bb.0:
+    liveins: $edi, $esi
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    CMP32ri %0, 2, implicit-def $eflags
+    CMP32ri %1, 7, implicit-def $eflags
+    JCC_1 %bb.2, 12, implicit $eflags
+    JMP_1 %bb.1
+
+  bb.1:
+    %2:gr32 = nuw nsw IMUL32rr %0, %1, implicit-def dead $eflags
+    JMP_1 %bb.3
+
+  bb.2:
+
+  bb.3:
+    CMP32ri %0, 2, implicit-def $eflags
+    $al = SETCCr 15, implicit $eflags
+    RET 0, $al
+...
+
+# Same shape WITHOUT the shadow clobber: the producer in the dominating block is
+# found, the redundant compare in bb.3 is removed and the IMUL on the path is
+# rewritten to its NF variant. Positive control for dom_shadow_clobber.
+---
+name: dom_no_shadow
+tracksRegLiveness: true
+body: |
+  ; CHECK-LABEL: name: dom_no_shadow
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; CHECK-NEXT:   liveins: $edi, $esi
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[COPY:%[0-9]+]]:gr32 = COPY $edi
+  ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+  ; CHECK-NEXT:   CMP32ri [[COPY]], 2, implicit-def $eflags
+  ; CHECK-NEXT:   JCC_1 %bb.2, 12, implicit $eflags
+  ; CHECK-NEXT:   JMP_1 %bb.1
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT:   liveins: $eflags
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[IMUL32rr_NF:%[0-9]+]]:gr32 = nuw nsw IMUL32rr_NF [[COPY]], [[COPY1]]
+  ; CHECK-NEXT:   JMP_1 %bb.3
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT:   liveins: $eflags
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.3:
+  ; CHECK-NEXT:   liveins: $eflags
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   $al = SETCCr 15, implicit $eflags
+  ; CHECK-NEXT:   RET 0, $al
+  bb.0:
+    liveins: $edi, $esi
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    CMP32ri %0, 2, implicit-def $eflags
+    JCC_1 %bb.2, 12, implicit $eflags
+    JMP_1 %bb.1
+
+  bb.1:
+    %2:gr32 = nuw nsw IMUL32rr %0, %1, implicit-def dead $eflags
+    JMP_1 %bb.3
+
+  bb.2:
+
+  bb.3:
+    CMP32ri %0, 2, implicit-def $eflags
+    $al = SETCCr 15, implicit $eflags
+    RET 0, $al
+...
+
+# Profitability: like dom_no_shadow, but bb.3 (the multi-predecessor block)
+# reaches the redundant compare in bb.4 through a single-predecessor chain that
+# itself has an NF-convertible clobber (the IMUL in bb.3). That clobber is on
+# the unconditional path to the compare, so converting it is pure added cost:
----------------
phoebewang wrote:

I don't get the point. The cost is code size. It doesn't matter it's conditional or unconditional path.

https://github.com/llvm/llvm-project/pull/208184