avr: use compiler-rt

This change adds support for compiler-rt, which supports float64 (unlike
libgcc for AVR). This gets a number of tests to pass that require
float64 support.

We're still using libgcc with this change, but libgcc will probably be
removed eventually once AVR support in compiler-rt is a bit more mature.

I've also pushed a fix for a small regression in our
xtensa_release_14.0.0-patched LLVM branch that has also been merged
upstream. Without it, a floating point comparison against zero always
returns true which is certainly a bug. It is necessary to correctly
print floating point values.

.circleci/config.yml

@@ -22,12 +22,12 @@ commands:
     steps:
       - restore_cache:
           keys:
-            - llvm-source-14-v1
+            - llvm-source-14-v2
       - run:
           name: "Fetch LLVM source"
           command: make llvm-source
       - save_cache:
-          key: llvm-source-14-v1
+          key: llvm-source-14-v2
           paths:
             - llvm-project/clang/lib/Headers
             - llvm-project/clang/include

.github/workflows/linux.yml

@@ -38,7 +38,7 @@ jobs:
         uses: actions/cache@v2
         id: cache-llvm-source
         with:
-          key: llvm-source-14-linux-v1
+          key: llvm-source-14-linux-v2
           path: |
             llvm-project/clang/lib/Headers
             llvm-project/clang/include
@@ -179,7 +179,7 @@ jobs:
         uses: actions/cache@v2
         id: cache-llvm-source
         with:
-          key: llvm-source-14-linux-asserts-v1
+          key: llvm-source-14-linux-asserts-v2
           path: |
             llvm-project/clang/lib/Headers
             llvm-project/clang/include
@@ -276,7 +276,7 @@ jobs:
         uses: actions/cache@v2
         id: cache-llvm-source
         with:
-          key: llvm-source-14-linux-v1
+          key: llvm-source-14-linux-v2
           path: |
             llvm-project/clang/lib/Headers
             llvm-project/clang/include

.github/workflows/windows.yml

@@ -37,7 +37,7 @@ jobs:
         uses: actions/cache@v2
         id: cache-llvm-source
         with:
-          key: llvm-source-14-windows-v1
+          key: llvm-source-14-windows-v2
           path: |
             llvm-project/clang/lib/Headers
             llvm-project/clang/include

builder/builtins.go

@@ -40,7 +40,6 @@ var genericBuiltins = []string{
 	"divdf3.c",
 	"divdi3.c",
 	"divmoddi4.c",
-	"divmodsi4.c",
 	"divsc3.c",
 	"divsf3.c",
 	"divsi3.c",
@@ -127,7 +126,6 @@ var genericBuiltins = []string{
 	"ucmpti2.c",
 	"udivdi3.c",
 	"udivmoddi4.c",
-	"udivmodsi4.c",
 	"udivmodti4.c",
 	"udivsi3.c",
 	"udivti3.c",
@@ -154,6 +152,14 @@ var aeabiBuiltins = []string{
 	"arm/aeabi_memset.S",
 	"arm/aeabi_uidivmod.S",
 	"arm/aeabi_uldivmod.S",
+
+	// These two are not technically EABI builtins but are used by them and only
+	// seem to be used on ARM. LLVM seems to use __divsi3 and __modsi3 on most
+	// other architectures.
+	// Most importantly, they have a different calling convention on AVR so
+	// should not be used on AVR.
+	"divmodsi4.c",
+	"udivmodsi4.c",
 }
 
 // CompilerRT is a library with symbols required by programs compiled with LLVM.

builder/library.go

@@ -148,6 +148,8 @@ func (l *Library) load(config *compileopts.Config, tmpdir string) (job *compileJ
 		// However, ARM has not done this.
 		if strings.HasPrefix(target, "i386") || strings.HasPrefix(target, "x86_64") {
 			args = append(args, "-march="+cpu)
+		} else if strings.HasPrefix(target, "avr") {
+			args = append(args, "-mmcu="+cpu)
 		} else {
 			args = append(args, "-mcpu="+cpu)
 		}
@@ -155,6 +157,13 @@ func (l *Library) load(config *compileopts.Config, tmpdir string) (job *compileJ
 	if strings.HasPrefix(target, "arm") || strings.HasPrefix(target, "thumb") {
 		args = append(args, "-fshort-enums", "-fomit-frame-pointer", "-mfloat-abi=soft", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables")
 	}
+	if strings.HasPrefix(target, "avr") {
+		// AVR defaults to C float and double both being 32-bit. This deviates
+		// from what most code (and certainly compiler-rt) expects. So we need
+		// to force the compiler to use 64-bit floating point numbers for
+		// double.
+		args = append(args, "-mdouble=64")
+	}
 	if strings.HasPrefix(target, "riscv32-") {
 		args = append(args, "-march=rv32imac", "-mabi=ilp32", "-fforce-enable-int128")
 	}

main_test.go

@@ -182,11 +182,6 @@ func runPlatTests(options compileopts.Options, tests []string, t *testing.T) {
 			// Not all tests are currently supported on AVR.
 			// Skip the ones that aren't.
 			switch name {
-			case "atomic.go":
-				// Requires GCC 11.2.0 or above for interface comparison.
-				// https://github.com/gcc-mirror/gcc/commit/f30dd607669212de135dec1f1d8a93b8954c327c
-				continue
-
 			case "reflect.go":
 				// Reflect tests do not work due to type code issues.
 				continue
@@ -203,20 +198,12 @@ func runPlatTests(options compileopts.Options, tests []string, t *testing.T) {
 				// Reflect size calculation crashes.
 				continue
 
-			case "binop.go":
-				// Interface comparison results are inverted.
-				continue
-
 			case "channel.go":
 				// Freezes after recv from closed channel.
 				continue
 
-			case "float.go", "math.go", "print.go":
-				// Stuck in runtime.printfloat64.
-				continue
-
-			case "interface.go":
-				// Several comparison tests fail.
+			case "math.go":
+				// Stuck somewhere, not sure what's happening.
 				continue
 
 			case "cgo/":

targets/avr.json

@@ -6,6 +6,7 @@
 	"gc": "conservative",
 	"linker": "avr-gcc",
 	"scheduler": "none",
+	"rtlib": "compiler-rt",
 	"default-stack-size": 256,
 	"cflags": [
 		"-Werror"