[llvm] [MIPS][float] Fixed SingleFloat codegen on N32/N64 targets (PR #140575)

[Davide Mor via llvm-commits]

================
@@ -0,0 +1,148 @@
+; RUN: llc -mtriple=mips -relocation-model=static -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM32,O32 %s
+; RUN: llc -mtriple=mipsel -relocation-model=static -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM32,O32 %s
+
+; RUN: llc -mtriple=mips64 -relocation-model=static -target-abi n32 -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM32,N32,NEW,NEWBE %s
+; RUN: llc -mtriple=mips64el -relocation-model=static -target-abi n32 -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM32,N32,NEW,NEWLE %s
+
+; RUN: llc -mtriple=mips64 -relocation-model=static -target-abi n64 -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM64,N64,NEW,NEWBE %s
+; RUN: llc -mtriple=mips64el -relocation-model=static -target-abi n64 -mattr=single-float < %s \
+; RUN:   | FileCheck --check-prefixes=ALL,SYM64,N64,NEW,NEWLE %s
+
+ at floats = global [11 x float] zeroinitializer
+ at doubles = global [11 x double] zeroinitializer
+
+define void @double_args(double %a, ...)
+                         nounwind {
+entry:
+        %0 = getelementptr [11 x double], ptr @doubles, i32 0, i32 1
+        store volatile double %a, ptr %0
+
+        %ap = alloca ptr
+        call void @llvm.va_start(ptr %ap)
+        %b = va_arg ptr %ap, double
+        %1 = getelementptr [11 x double], ptr @doubles, i32 0, i32 2
+        store volatile double %b, ptr %1
+        call void @llvm.va_end(ptr %ap)
+        ret void
----------------
Tazdevil971 wrote:

I tried merging all of the different test cases, honestly it's not that bad, you can see it [here](https://github.com/llvm/llvm-project/blob/8fda8677250b1bb1d3fc476deb8329c55c8b3942/llvm/test/CodeGen/Mips/cconv/arguments-float.ll). You can actually ignore BE/LE variants and end up with just 9 different run combinations.

If you are okay with it I could merge all of the test variants and do some refactoring of the old tests touched by this PR.

https://github.com/llvm/llvm-project/pull/140575