Improve loc information of assembly listings

Previous loc information on assembly listing was being printed for
every line, while the intent was to print it only when it changes.

Also update the tests to expose and test that case, and remaster
other tests to match.

internal/driver/driver_test.go

@@ -520,7 +520,7 @@ func cpuProfile() *profile.Profile {
 			Address: 0x3002,
 			Line: []profile.Line{
 				{Function: cpuF[5], Line: 5},
-				{Function: cpuF[3], Line: 7},
+				{Function: cpuF[3], Line: 9},
 			},
 		},
 	}

internal/driver/testdata/pprof.cpu.callgrind

@@ -72,7 +72,7 @@ calls=0 -8193 1
 ob=(1)
 fl=(3)
 fn=(5)
-+1 7 0
++1 9 0
 cfl=(3)
 cfn=(3)
 calls=0 +1 5

internal/driver/testdata/pprof.cpu.flat.addresses.disasm

@@ -1,14 +1,14 @@
 Total: 1.12s
 ROUTINE ======================== line1000
      1.10s      1.10s (flat, cum) 98.21% of Total
-     1.10s      1.10s       1000: instruction one                         ; line1000 file1000.src:1
+     1.10s      1.10s       1000: instruction one                         ;line1000 file1000.src:1
          .          .       1001: instruction two
          .          .       1002: instruction three
          .          .       1003: instruction four
 ROUTINE ======================== line3000
       10ms      1.12s (flat, cum)   100% of Total
-      10ms      1.01s       3000: instruction one                         ; line3000 file3000.src:6
-         .      100ms       3001: instruction two                         ; line3000 file3000.src:9
-         .       10ms       3002: instruction three                       ; line3000 file3000.src:7
+      10ms      1.01s       3000: instruction one                         ;line3000 file3000.src:6
+         .      100ms       3001: instruction two                         ;line3000 file3000.src:9
+         .       10ms       3002: instruction three
          .          .       3003: instruction four
          .          .       3004: instruction five

internal/driver/testdata/pprof.cpu.flat.addresses.weblist

@@ -90,13 +90,13 @@ Duration: 10s, Total samples = 1.12s (11.20%)<br>Total: 1.12s</div><h1>line1000<
 <span class=line>      5</span> <span class=nop>           .          . line5 </span>
 <span class=line>      6</span> <span class=deadsrc>        10ms      1.01s line6 </span><span class=asm>                10ms      1.01s     3000: instruction one                                  <span class=disasmloc>file3000.src:6</span>
 </span>
-<span class=line>      7</span> <span class=deadsrc>           .       10ms line7 </span><span class=asm>                   .       10ms     3002: instruction three                                <span class=disasmloc>file3000.src:7</span>
+<span class=line>      7</span> <span class=nop>           .          . line7 </span>
+<span class=line>      8</span> <span class=nop>           .          . line8 </span>
+<span class=line>      9</span> <span class=deadsrc>           .      110ms line9 </span><span class=asm>                   .      100ms     3001: instruction two                                  <span class=disasmloc>file3000.src:9</span>
+                   .       10ms     3002: instruction three                                <span class=disasmloc>file3000.src:9</span>
                    .          .     3003: instruction four                                 <span class=disasmloc></span>
                    .          .     3004: instruction five                                 <span class=disasmloc></span>
 </span>
-<span class=line>      8</span> <span class=nop>           .          . line8 </span>
-<span class=line>      9</span> <span class=deadsrc>           .      100ms line9 </span><span class=asm>                   .      100ms     3001: instruction two                                  <span class=disasmloc>file3000.src:9</span>
-</span>
 <span class=line>     10</span> <span class=nop>           .          . line0 </span>
 <span class=line>     11</span> <span class=nop>           .          . line1 </span>
 <span class=line>     12</span> <span class=nop>           .          . line2 </span>

internal/report/report.go

@@ -368,35 +368,45 @@ func printAssembly(w io.Writer, rpt *Report, obj plugin.ObjTool) error {
 
 		function, file, line := "", "", 0
 		for _, n := range ns {
-			flat := valueOrDot(n.flatValue(), rpt)
-			cum := valueOrDot(n.cumValue(), rpt)
-			if n.function == function && n.file == file && n.line == line {
+			locStr := ""
+			// Skip loc information if it hasn't changed from previous instruction.
+			if n.function != function || n.file != file || n.line != line {
+				function, file, line = n.function, n.file, n.line
+				if n.function != "" {
+					locStr = n.function + " "
+				}
+				if n.file != "" {
+					locStr += n.file
+					if n.line != 0 {
+						locStr += fmt.Sprintf(":%d", n.line)
+					}
+				}
+			}
+			switch {
+			case locStr == "":
+				// No location info, just print the instruction.
 				fmt.Fprintf(w, "%10s %10s %10x: %s\n",
-					flat, cum,
+					valueOrDot(n.flatValue(), rpt),
+					valueOrDot(n.cumValue(), rpt),
 					n.address, n.instruction,
 				)
-				continue
-			}
-			line := ""
-			if n.line > 0 {
-				line = fmt.Sprintf(":%d", n.line)
-			}
-			if len(n.instruction) <= 40 {
-				fmt.Fprintf(w, "%10s %10s %10x: %-40s; %s %s%s\n",
-					flat, cum,
+			case len(n.instruction) < 40:
+				// Short instruction, print loc on the same line.
+				fmt.Fprintf(w, "%10s %10s %10x: %-40s;%s\n",
+					valueOrDot(n.flatValue(), rpt),
+					valueOrDot(n.cumValue(), rpt),
+					n.address, n.instruction,
+					locStr,
+				)
+			default:
+				// Long instruction, print loc on a separate line.
+				fmt.Fprintf(w, "%74s;%s\n", "", locStr)
+				fmt.Fprintf(w, "%10s %10s %10x: %s\n",
+					valueOrDot(n.flatValue(), rpt),
+					valueOrDot(n.cumValue(), rpt),
 					n.address, n.instruction,
-					n.function, n.file, line,
 				)
-				continue
 			}
-			fmt.Fprintf(w, "%75s; %s %s%s\n",
-				"",
-				n.function, n.file, line,
-			)
-			fmt.Fprintf(w, "%10s %10s %10x: %s\n",
-				flat, cum,
-				n.address, n.instruction,
-			)
 		}
 	}
 	return nil