[llvm] [AMDGPU] Visit all PHIs in each call to optimizeLiveType (PR #147522)

[Shilei Tian via llvm-commits]

================
@@ -0,0 +1,94 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes="amdgpu-late-codegenprepare,verify" %s  | FileCheck %s
+
+; This crashed because the PHI with a splat was rejected, but then we marked the PHI
+; as visited and tried to convert one of its user afterwards.
+
+define amdgpu_kernel void @widget(i1 %arg, <4 x i8> %arg1, i64 %arg2) {
+; CHECK-LABEL: define amdgpu_kernel void @widget(
+; CHECK-SAME: i1 [[ARG:%.*]], <4 x i8> [[ARG1:%.*]], i64 [[ARG2:%.*]]) {
+; CHECK-NEXT:  [[BB:.*]]:
+; CHECK-NEXT:    [[WIDGET_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(272) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+; CHECK-NEXT:    [[ARG_KERNARG_OFFSET_ALIGN_DOWN:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[WIDGET_KERNARG_SEGMENT]], i64 36
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr addrspace(4) [[ARG_KERNARG_OFFSET_ALIGN_DOWN]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[TMP0]] to i1
+; CHECK-NEXT:    [[ARG1_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[WIDGET_KERNARG_SEGMENT]], i64 40
+; CHECK-NEXT:    [[ARG1_LOAD:%.*]] = load <4 x i8>, ptr addrspace(4) [[ARG1_KERNARG_OFFSET]], align 8
+; CHECK-NEXT:    [[ARG2_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[WIDGET_KERNARG_SEGMENT]], i64 44
+; CHECK-NEXT:    [[ARG2_LOAD:%.*]] = load i64, ptr addrspace(4) [[ARG2_KERNARG_OFFSET]], align 4
+; CHECK-NEXT:    br label %[[BB_3:.*]]
+; CHECK:       [[BB_3]]:
+; CHECK-NEXT:    [[PHI:%.*]] = phi ptr addrspace(1) [ null, %[[BB]] ], [ [[GETELEMENTPTR:%.*]], %[[BB_14:.*]] ]
+; CHECK-NEXT:    [[PHI4:%.*]] = phi <4 x i8> [ splat (i8 1), %[[BB]] ], [ [[PHI15:%.*]], %[[BB_14]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_6_PREHEADER:.*]], label %[[BB_5:.*]]
+; CHECK:       [[BB_5]]:
+; CHECK-NEXT:    br label %[[BB_14]]
+; CHECK:       [[BB_6_PREHEADER]]:
+; CHECK-NEXT:    br label %[[BB_6:.*]]
+; CHECK:       [[BB_6]]:
+; CHECK-NEXT:    [[PHI7:%.*]] = phi <4 x i8> [ [[PHI13:%.*]], %[[BB_12:.*]] ], [ [[PHI4]], %[[BB_6_PREHEADER]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_8:.*]], label %[[BB_12]]
+; CHECK:       [[BB_8]]:
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_10:.*]], label %[[BB_9:.*]]
+; CHECK:       [[BB_9]]:
+; CHECK-NEXT:    br label %[[BB_10]]
+; CHECK:       [[BB_10]]:
+; CHECK-NEXT:    [[PHI11:%.*]] = phi <4 x i8> [ [[PHI7]], %[[BB_9]] ], [ zeroinitializer, %[[BB_8]] ]
+; CHECK-NEXT:    [[EXTRACTELEMENT:%.*]] = extractelement <4 x i8> [[PHI11]], i64 0
+; CHECK-NEXT:    store i8 [[EXTRACTELEMENT]], ptr addrspace(1) [[PHI]], align 1
+; CHECK-NEXT:    br label %[[BB_12]]
+; CHECK:       [[BB_12]]:
+; CHECK-NEXT:    [[PHI13]] = phi <4 x i8> [ zeroinitializer, %[[BB_10]] ], [ [[PHI7]], %[[BB_6]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_6]], label %[[BB_14]]
+; CHECK:       [[BB_14]]:
+; CHECK-NEXT:    [[PHI15]] = phi <4 x i8> [ [[ARG1_LOAD]], %[[BB_5]] ], [ zeroinitializer, %[[BB_12]] ]
+; CHECK-NEXT:    [[GETELEMENTPTR]] = getelementptr i8, ptr addrspace(1) [[PHI]], i64 [[ARG2_LOAD]]
+; CHECK-NEXT:    br label %[[BB_3]]
+;
+bb:
+  %widget.kernarg.segment = call nonnull align 16 dereferenceable(272) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+  %arg.kernarg.offset.align.down = getelementptr inbounds i8, ptr addrspace(4) %widget.kernarg.segment, i64 36
+  %0 = load i32, ptr addrspace(4) %arg.kernarg.offset.align.down, align 4
+  %1 = trunc i32 %0 to i1
----------------
shiltian wrote:

use named variables

https://github.com/llvm/llvm-project/pull/147522