[llvm] [CalcSpillWeights] Avoid x87 excess precision influencing weight result (PR #100165)

[Dimitry Andric via llvm-commits]

https://github.com/DimitryAndric created https://github.com/llvm/llvm-project/pull/100165

Fixes #99396

The result of `VirtRegAuxInfo::weightCalcHelper` can be influenced by x87 excess precision, which can result in slightly different register choices when the compiler is hosted on x86_64 or i386. This leads to different object file output when cross-compiling to i386, or native.

Similar to 7af3432e22b0, add a `volatile` qualifier to the local `Weight` variable to force it onto the stack, and avoid the excess precision.


>From c9698176fdee886ebc6a3c2d3230ef759f0520d3 Mon Sep 17 00:00:00 2001
From: Dimitry Andric <dimitry at andric.com>
Date: Tue, 23 Jul 2024 19:02:36 +0200
Subject: [PATCH] [CalcSpillWeights] Avoid x87 excess precision influencing
 weight result

Fixes #99396

The result of `VirtRegAuxInfo::weightCalcHelper` can be influenced by
x87 excess precision, which can result in slightly different register
choices when the compiler is hosted on x86_64 or i386. This leads to
different object file output when cross-compiling to i386, or native.

Similar to 7af3432e22b0, add a `volatile` qualifier to the local
`Weight` variable to force it onto the stack, and avoid the excess
precision.
---
 llvm/lib/CodeGen/CalcSpillWeights.cpp |  4 +-
 llvm/test/CodeGen/X86/pr99396.ll      | 59 +++++++++++++++++++++++++++
 2 files changed, 62 insertions(+), 1 deletion(-)
 create mode 100644 llvm/test/CodeGen/X86/pr99396.ll

diff --git a/llvm/lib/CodeGen/CalcSpillWeights.cpp b/llvm/lib/CodeGen/CalcSpillWeights.cpp
index 1d767a3484bca..60055e6ba0cd3 100644
--- a/llvm/lib/CodeGen/CalcSpillWeights.cpp
+++ b/llvm/lib/CodeGen/CalcSpillWeights.cpp
@@ -257,7 +257,9 @@ float VirtRegAuxInfo::weightCalcHelper(LiveInterval &LI, SlotIndex *Start,
       return -1.0f;
     }
 
-    float Weight = 1.0f;
+    // Force Weight onto the stack so that x86 doesn't add hidden precision,
+    // similar to HWeight below.
+    volatile float Weight = 1.0f;
     if (IsSpillable) {
       // Get loop info for mi.
       if (MI->getParent() != MBB) {
diff --git a/llvm/test/CodeGen/X86/pr99396.ll b/llvm/test/CodeGen/X86/pr99396.ll
new file mode 100644
index 0000000000000..7bc1153b57990
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr99396.ll
@@ -0,0 +1,59 @@
+; RUN: llc < %s -mtriple=i386-unknown-freebsd -enable-misched -relocation-model=pic | FileCheck %s
+
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-i128:128-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i386-unknown-freebsd15.0"
+
+ at c = external local_unnamed_addr global ptr
+
+declare i32 @fn2() local_unnamed_addr
+
+declare i32 @fn3() local_unnamed_addr
+
+define noundef i32 @fn4() #0 {
+entry:
+  %0 = load i32, ptr @fn4, align 4
+; CHECK: movl fn4 at GOT(%ebx), %edi
+; CHECK-NEXT: movl (%edi), %edx
+  %1 = load ptr, ptr @c, align 4
+; CHECK: movl c at GOT(%ebx), %eax
+; CHECK-NEXT: movl (%eax), %esi
+; CHECK-NEXT: testl %esi, %esi
+  %cmp.g = icmp eq ptr %1, null
+  br i1 %cmp.g, label %if.then.g, label %if.end3.g
+
+if.then.g:                                        ; preds = %entry
+  %2 = load i32, ptr inttoptr (i32 1 to ptr), align 4
+  %cmp1.g = icmp slt i32 %2, 0
+  br i1 %cmp1.g, label %if.then2.g, label %if.end3.g
+
+if.then2.g:                                       ; preds = %if.then.g
+  %.g = load volatile i32, ptr null, align 2147483648
+  br label %f.exit
+
+if.end3.g:                                        ; preds = %if.then.g, %entry
+  %h.i.g = icmp eq i32 %0, 0
+  br i1 %h.i.g, label %f.exit, label %while.body.g
+
+while.body.g:                                     ; preds = %if.end3.g, %if.end8.g
+  %buff.addr.019.g = phi ptr [ %incdec.ptr.g, %if.end8.g ], [ @fn4, %if.end3.g ]
+  %g.addr.018.g = phi i32 [ %dec.g, %if.end8.g ], [ %0, %if.end3.g ]
+  %call4.g = tail call i32 @fn3(ptr %1, ptr %buff.addr.019.g, i32 %g.addr.018.g)
+  %cmp5.g = icmp slt i32 %call4.g, 0
+  br i1 %cmp5.g, label %if.then6.g, label %if.end8.g
+
+if.then6.g:                                       ; preds = %while.body.g
+  %call7.g = tail call i32 @fn2(ptr null)
+  br label %f.exit
+
+if.end8.g:                                        ; preds = %while.body.g
+  %dec.g = add i32 %g.addr.018.g, 1
+  %incdec.ptr.g = getelementptr i32, ptr %buff.addr.019.g, i32 1
+  store i64 0, ptr %1, align 4
+  %h.not.g = icmp eq i32 %dec.g, 0
+  br i1 %h.not.g, label %f.exit, label %while.body.g
+
+f.exit:                                           ; preds = %if.end8.g, %if.then6.g, %if.end3.g, %if.then2.g
+  ret i32 0
+}
+
+attributes #0 = { "frame-pointer"="all" "tune-cpu"="generic" }